51 of 51 fields researched, in 16 groups. Open a group, pick a field, and see what its companies already hand to AI, with the source.
Each line is one company's adoption, from any supplier, with the page we read. Free to reuse under CC-BY-4.0. Cite as: massrelay.ai, What AI is for — the field guide.
Have estate maps and a drone pilot or contracted drone service; start with one block and compare imagery readings with agronomist checks; expect ongoing flight, data-processing and agronomy staffing.
Machine plucking expanded to offset a manual labour shortage, drones applying fertiliser and pesticide and supervising farm practice, and AI being explored for efficiency and quality in Ceylon tea
Where
LK, South Asia
Segment
enterprise (listed tea and coffee estates)
Problem
manual labour shortage and rising production costs
Technology
machine plucking, agricultural drones, AI (exploratory)
none stated (no tea productivity or cost figures given; the article cites only an 18 % electricity cut and a 500 ha coffee estate)
Effort
high
Risks
Machine plucking can lower leaf quality; labour displacement; no productivity figures published; drone rules.
How to adopt
Pick the most labour-bound task (plucking or spraying), pilot machines or drones on one division, train existing staff to operate them, then extend.
How a room runs it
the ROOM for a tea estate division; the estate manager, the field officers, the drone operator and a plantation teammate (an agent definition with the tea skills); the agent reads the plucking and drone-flight logs the estate already keeps and GENERATES the week's plucking-round, spray and fertiliser jobs for the field officers (block, product, window) as plan steps with owners and due dates, and raises a decision for the manager when a block's input spend passes its envelope (approve more / hold) in the Inbox as a proposal with a deadline; the humans keep the plucking, the flying, the sign-off and the spend above the envelope; the receipt records each finished job with block, product and date; budgets degrade before they block
The engagement on top
a custom engagement: the estate's field-record sheet or ERP and the drone operator's flight log wired to the room through connectors loaded on this pull, an agronomist or plantation accountant as the paid professional on the match, and the sector pack's task templates for the recurring rounds; the client receives the reports and gives direction in the room
Mechanised tea plucking: a growers' agency subsidiary designed battery-powered plucking machines for smallholders with credit purchase; a listed estate invested in mechanised plucking and cut 267 jobs in 2023
Tea Machinery Company (Temec, subsidiary of Kenya Tea Development Agency); Greenland Fedha; Sasini Plc (Kenya)
Outcome
Sasini cut 267 jobs in 2023 and staff costs from Kes381.9 million to Kes359.9 million (Sasini statements as reported)
Effort
medium
Risks
Leaf quality below buyer standard; worker resistance and job loss; cost of machines to small growers.
How to adopt
Start with a shared machine for a group of growers on a credit or hire plan, check leaf quality against the buyer standard, and plan for the labour effects.
How a room runs it
the ROOM for a smallholder tea group; the group's chair, the growers, the plucking-machine operator and a tea teammate; the agent reads the group's plot list and the factory delivery records and GENERATES the plucking rounds and machine-booking steps for the operator (plot, day) with due dates as plan steps with owners and due dates, and raises a decision for each grower (book the machine and its fee, with a deadline) in the Inbox as a proposal with a deadline; the humans keep the plucking, the leaf-quality call and the payment; the receipt records each plot plucked, the leaf delivered and the machine fee, split across growers; budgets degrade before they block
The engagement on top
a custom engagement: the group's member register and the factory's leaf-delivery statements wired to the room through connectors loaded on this pull, a cooperative accountant or agronomist as the paid professional on the match, and the sector pack's task templates for the recurring rounds; the client receives the reports and gives direction in the room
Drone mapping of oil palm estates: an orthomosaic map, an ML tree count and GIS identification of empty planting spaces for replanting, delivered on a web platform
Where
CI, Africa
Segment
enterprise (private palm estates)
Problem
no accurate estate map or tree inventory for planning replanting
United Oil Company (Ivory Coast, about 16,300 ha across three plantations); Ghana Flying Labs; SKT Live
Outcome
none stated (the article says the mapping ensured accurate data but gives no counts or accuracy)
Effort
medium
Risks
Vendor-only evidence; count accuracy unreported; reliance on an outside survey provider; data ownership.
How to adopt
Commission a drone survey of one plantation, check the tree count against a ground sample, and use the gap map to plan replanting.
How a room runs it
the ROOM for a palm estate's replanting programme; the estate manager, the replanting supervisor, the survey contractor and a palm teammate; the agent reads the tree-count and gap map the contractor delivers and GENERATES the replanting rounds for the supervisor (block, number of palms, planting window) with due dates as plan steps with owners and due dates, and raises a decision for the manager (approve the seedling order, with a deadline) in the Inbox as a proposal with a deadline; the humans keep the survey flight, the planting and the spend above the envelope; the receipt records each map delivered, each block replanted and each seedling order; budgets degrade before they block
The engagement on top
a custom engagement: the estate's block register and the survey contractor's map files wired to the room through connectors loaded on this pull, a survey contractor or agronomist as the paid professional on the match, and the sector pack's task templates for the recurring rounds; the client receives the reports and gives direction in the room
Source: vendor · 2024 · vendor only; to corroborate
Precision agronomy: drones with multispectral imaging, soil and weather sensors (moisture, temperature, light, pH), satellite data; ML recommends fertilisation and irrigation
plantation managers and operators on palm oil and tea estates; 'several large plantation companies' (unnamed)
Outcome
AI image processing 10,000 ha in 4 h vs 14+ days for 20 people (figure from William Tao, COO of Insight Robotics, a drone-analytics supplier to palm growers; reported by Bloomberg 2019); no yield figures
Effort
high
Risks
Recommendation accuracy unproven without yield data; dependence on drone or analytics vendors; skills gaps on estates.
How to adopt
Have estate maps and a drone pilot or contracted drone service; start with one block and compare imagery readings with agronomist checks; expect ongoing flight, data-processing and agronomy staffing.
How a room runs it
the ROOM for an estate block: the estate manager, the field supervisor and an agronomy teammate (an agent definition with the plantation skills); the agent reads the readings and imagery the estate already has and GENERATES the week's tasks for the supervisor (which plots to irrigate, dose, inspect) as plan steps with owners and due dates, and the manager's advisory as a decision block with a deadline; what the humans keep: the field work, the sign-off, the spend above the envelope; the receipt records every finished task
The engagement on top
a custom engagement: the estate's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (sensor feeds, the imagery folder, the ERP), the agronomist joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Early detection of abnormal conditions (pests, disease) from imagery and sensors so managers act before damage spreads
Where
MY, Southeast Asia
Segment
enterprise estates
Problem
pests and disease
Technology
drones, satellite anomaly detection, ML pattern detection
Who uses it
estate managers
Outcome
none stated
Effort
medium
Risks
No detection rate given, so false alarms and misses are unknown; drone flight regulation; connectivity at remote estates.
How to adopt
Subscribe to satellite hot-spot alerts for estate boundaries; use drones to verify hard-to-reach sites; assign a manager to act on each alert.
How a room runs it
the same room: an alert from imagery or sensors becomes a TASK for the field team (inspect block 7 by Thursday) generated by the teammate, and a DECISION for the manager (treat / wait) with the evidence attached; nobody reads a dashboard — the work arrives as tasks and a decision in the Inbox
The engagement on top
inspection and treatment tasks handed to field staff over WhatsApp (the channel is built); the agronomist's verdict as a receipt
Barriers: skilled labour limited, connectivity limits on plantations, climate uncertainty — technology 'to support rather than replace farmers'
Where
MY, Southeast Asia
Segment
all
Problem
adoption barriers
Technology
edge computing for offline analysis
Who uses it
all actors
Outcome
Mata Aerotech's drone services covered 200,000 ha of oil palm estates; about 1,000 commercial drones in Malaysia (industry sources, unattributed); no quantified labour saving
Effort
medium
Risks
Robotics immature; workforce resistance or displacement; connectivity limits on plantations.
How to adopt
Treat labour shortage as the business case: pick one labour-heavy task (spraying, monitoring), pilot drones or machines on it, train existing staff to operate them.
How a room runs it
the room meets the barriers as it is: field staff work over WhatsApp with an agent that tolerates offline entry; the advice arrives in Bahasa; the agent supports the supervisor and never replaces the sign-off
Sensor-based smart irrigation and fertigation for oil palm: soil-moisture and weather sensors send data over IoT to a dashboard that advises irrigation and fertigation, on two hectares of a research farm with two local growers; the study also compared irrigation systems
Where
CO, Latin America
Segment
enterprise and growers (pilot with research body)
Problem
irrigation of oil palm in a dry region wastes water; growers have not adopted efficient irrigation
Technology
soil-moisture and weather sensors, IoT, an advisory dashboard, drip, sprinkler and fertigation systems, water harvesting
Who uses it
Delphy (Netherlands) with Solidaridad Network, FutureWater and Cenipalma (the research arm of Fedepalma, Colombia); the Dutch Embassy in Colombia; two growers
Outcome
Efficient irrigation in the Sevilla basin could save up to 67 % of water with sprinklers and up to 72 % with drip (project study; sprinkler theft is the main barrier)
Effort
high
Risks
Estimates are for one basin; sprinkler theft; upfront cost of drip; sensor upkeep.
How to adopt
Install sensors on a small demonstration block, compare irrigation options on cost and water use, and choose the system the site can protect from theft.
How a room runs it
the ROOM for a palm grower and the research farm; the grower, the farm's irrigation technician and an irrigation teammate; the agent reads the soil-moisture and weather feed the sensor dashboard already produces and GENERATES the irrigation and fertigation rounds for the technician (block, volume, window) with due dates as plan steps with owners and due dates, and raises a decision for the grower when water or input spend passes its envelope (approve more / hold) in the Inbox as a proposal with a deadline; the humans keep the valve and pump work, the sign-off and the spend above the envelope; the receipt records each irrigation round with block and volume; budgets degrade before they block
The engagement on top
a custom engagement: the sensor dashboard export and the water-use sheet wired to the room through connectors loaded on this pull, an irrigation engineer as the paid professional on the match, and the sector pack's task templates for the recurring rounds; the client receives the reports and gives direction in the room
Targeted irrigation and dosing: automated irrigation that activates only when needed; lower water and fertiliser use
Where
ID, Southeast Asia
Segment
SMB and enterprise
Problem
resource waste
Technology
automated irrigation, edge computing on site, cloud for storage and long-term analysis
Who uses it
operators
Outcome
Up to 20 % cost savings and up to 15 % yield improvement, reported by unnamed plantation companies (vendor page; no measurer named); a Heriot-Watt paper (Sustainability 2023, Malaysia) reports positive return on investment above 1.5 ha but gives no figures the company’s own figure
Install soil-moisture sensors and controllable valves on a pilot block; set thresholds with an agronomist; budget for sensor maintenance and connectivity.
How a room runs it
the teammate schedules irrigation and dosing as tasks against the plots' baselines and raises a decision when a plot drifts over its input budget — the budget envelope is the plot's, the wall degrades before it blocks
Source: vendor · 2024 · one vendor source; to corroborate
Harvest-timing prediction feeding logistics planning, 'faster and more transparent' distribution; centralised cloud records for 'long-term analysis and compliance'
Where
ID, Southeast Asia
Segment
enterprise
Problem
logistics and traceability; compliance
Technology
ML harvest prediction, cloud records
Who uses it
estates, buyers, logistics
Outcome
Up to 20 % cost savings and up to 15 % yield improvement, reported by unnamed plantation companies (vendor page; no measurer named); no figure for harvest prediction the company’s own figure
Effort
medium
Risks
Forecast quality depends on clean historical data; vendor-only evidence; cloud data hosting and compliance questions.
How to adopt
Digitise harvest and delivery records first; then add forecasting on that history; connect to logistics and buyer planning.
How a room runs it
each harvest lot is a deliverable the teammate opens, routes to logistics as tasks, and closes with a receipt a buyer can verify; the compliance file is the record, not a separate system
The engagement on top
export deal rooms later (D261 — another monorepo on the public doors)
Source: vendor · 2024 · one vendor source; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 9 adoptions here, 0 were low effort, 5 medium and 4 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Trial spraying drones on a small share of fields with a distributor; train operators; check licensing and spray rules.
Camera-and-AI sprayer that sprays only where weeds are detected, sold per acre or on an annual licence
Where
US, North America
Segment
SMB and enterprise (row-crop farms)
Problem
herbicide cost and over-spraying
Technology
See & Spray (John Deere, ExactApply nozzles)
Who uses it
John Deere customers, United States
Outcome
More than 5 million acres sprayed in 2025, an average cut of nearly 50% in non-residual herbicide use and about 31 million gallons of herbicide mix saved (Deere, Nov 2025)
Effort
high
Risks
Missed weeds in heavy residue; company-reported savings; machine cost and connectivity.
Next: an Unlimited Annual License for high-use operations is offered for 2026 source
How to adopt
Start on one farm with one crop, compare against a broadcast strip, then widen acres.
How a room runs it
the field room: the grower, the agronomist, the sprayer operator and the dealer; the agent turns the season's coverage maps and weed pressure into a spray plan and generates the job tasks per field with dates, and any chemical choice becomes a proposal for the agronomist; the receipt records the product, the rate and the saved volume
The engagement on top
connectors to the machine data portal and the farm records, a spray-plan template, a label-compliance rule
Autonomous tractors, rice transplanters and combines in a staged path toward fully unmanned farm work
Where
JP, East Asia
Segment
enterprise (machinery maker) and farms
Problem
farm labour shortage and an ageing workforce
Technology
Agri Robo series (Kubota) with Nvidia, NTT and the University of Tokyo
Who uses it
Kubota, Japan
Outcome
About 700 Agri Robo machines had shipped by 2024; steps one (assisted steering) and two (unmanned under supervision) complete (Government of Japan, Jul 2025)
Effort
high
Risks
Safety around unattended machines; terrain limits; capital cost for small farms.
Next: Kubota's step three, fully automated operation, is in progress, with robots coordinating each other source
How to adopt
Begin with assisted steering, move to supervised unmanned passes on flat fields, then widen.
How a room runs it
the farm-operations room: the farm manager, the machine operator and the dealer; the agent plans the day's field passes and generates the machine tasks per field with dates, and any unattended-operation start becomes a proposal the supervisor approves; the receipt records the plan, the machine and the approver
The engagement on top
connectors to the machine telematics and the field map, a field-pass template, a supervised-autonomy rule
Satellite imagery and machine learning underwrite small input loans to maize farmers, bundled with seed, fertiliser, advice and insurance sold through local agents
Where
KE, Africa
Segment
SMB (smallholder maize farmers)
Problem
smallholders lack credit history and collateral, so banks do not lend to them
Apollo Agriculture (Kenya; also Zambia); investors SoftBank Vision Fund, Triodos Investment Management
Outcome
More than 100,000 farmers served; 7 in 10 farmers say productivity 'very much increased' (reported by lender Triodos; a perception, not measured yield); no repayment figure
Effort
high
Risks
Productivity is self-reported; no default data on the page; weather and price shocks; over-lending to the poor.
How to adopt
Pilot with one crop and one region, score from satellite and agent-collected data, set small loan sizes tied to the season, and compare repayment with a control group before scaling.
How a room runs it
the ROOM for an agri-lender's field operation: the credit manager, the agent-network lead, an agronomist and a farm-credit teammate (an agent definition with the lender's scoring policy); the agent teammate reads the farm records and repayment data the lender already holds and GENERATES the agent-network lead's tasks (visit these 40 farms before planting by the 10th) as plan steps with owners and due dates, and each lending limit change for the credit manager (raise / hold, with the evidence attached) as a decision block in the Inbox with a deadline; what the humans keep: the loan decision, the field visit and the insurance claim; the receipt records every finished task and every decision, and the client receives the reports and gives direction there Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the loan book, the agent app data, the satellite and weather feeds), an agricultural-finance specialist joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Drone spraying as a paid service: a farmer's roughly 8 ha sprayed by drone in about an hour versus a week with eight people; DJI and XAG hold about 80 % of the market
Where
CN, East Asia
Segment
SMB (farmers and drone-service operators)
Problem
rural labour shortages and slow manual spraying
Technology
agricultural spray drones (DJI, XAG)
Who uses it
XAG and DJI (China); farmer He Guozhu; South China Agricultural University (Prof. Lan Yubin)
Outcome
About 8 ha sprayed in about an hour versus a week with eight people (farmer He Guozhu, reported by CGTN)
Effort
medium
Risks
Pesticide drift and safety rules; displaced field labour; figures are company or farmer statements.
How to adopt
Hire a drone spraying service for one round, compare time and cost with manual work, then decide whether to buy or keep hiring.
How a room runs it
the ROOM for a farm group using a drone service; the farm group's chair, the farmers, the drone operator and a crop-protection teammate; the agent reads the plot list and the operator's flight log and GENERATES the spraying round for the operator (plot, product, day) with due dates as plan steps with owners and due dates, and raises a decision for each farmer (accept the service and its fee, with a deadline) in the Inbox as a proposal with a deadline; the humans keep the flying, the consent and the payment; the receipt records each plot sprayed with product and date, the fee split across the farmers served; budgets degrade before they block
The engagement on top
a custom engagement: the group's plot register and the operator's flight log wired to the room through connectors loaded on this pull, a drone service provider or extension officer as the paid professional on the match, and the sector pack's task templates for the recurring rounds; the client receives the reports and gives direction in the room
Traceability app for independent oil palm smallholders: farmers' plots are mapped and fresh-fruit-bunch origin is traced so mills can accept the bunches; data collected from 3,132 farmers in five locations over about three years
Where
ID, Southeast Asia
Segment
public / NGO with smallholders and mills
Problem
independent smallholders risk exclusion from supply chains that demand traceability; mills must prove origin
Technology
Hamurni mapping and traceability app
Who uses it
WWF-Indonesia; independent oil palm smallholders in Pelalawan and Kuantan Singingi (Riau), Katingan (Central Kalimantan), Kapuas Hulu and Sintang (West Kalimantan); palm oil mills
Outcome
Data collected from 3,132 farmers in five locations over about three years (WWF-Indonesia's own count)
Effort
high
Risks
Data accuracy; smallholder privacy; costs falling on the NGO or mill.
How to adopt
Map smallholder plots and register farmers; trace fruit-bunch origin to mills; fund field enumerators.
How a room runs it
the ROOM for a smallholder group selling to a mill: the group's chair, the farmers, the mill's buyer and a traceability teammate; the agent reads the plot records the farmers already hold and GENERATES the pre-harvest steps (confirm plot boundary, record harvest date, attach the delivery) for each farmer with due dates, and a decision for the buyer (accept the lot, with a deadline); the humans keep the harvest, the weighing and the acceptance; the receipt is the lot's record from plot to delivery, a file a buyer can verify
The engagement on top
a custom engagement: the group's plot register or the mill's intake system wired to the room through connectors loaded on this pull, a certification or sustainability professional as the paid professional on the match, and the sector pack's task templates for lot preparation and compliance files
Digital Village for oil palm smallholders: an app with fertiliser dosage per farm and harvest and maintenance records, an always-on Q&A service, an online input shop, and prices the cooperative unit no longer has to relay by hand
Where
ID, Southeast Asia
Segment
SMB (smallholders) and village cooperative
Problem
smallholders lag in technology adoption; prices and inputs reach them through intermediaries
Technology
Petani SawitPRO app, chat Q&A service ('Dokter Sawitina'), online input store
Who uses it
Karya Harapan Mukti village, Bungo regency, Jambi; the village cooperative unit (KUD Karya Mukti); SawitPRO (Indonesia)
Outcome
none stated
Effort
medium
Risks
Outcomes unmeasured; connectivity; data ownership by the cooperative versus the vendor.
How to adopt
Digitise farm and harvest records with a village cooperative; add price transparency and an input shop; train a local operator.
How a room runs it
the ROOM for a village cooperative: the cooperative's chair, the smallholders, the village head and a palm agronomy teammate; the agent reads the harvest and maintenance records the farmers already keep and GENERATES each farmer's fertiliser and upkeep tasks with due dates, and the cooperative's weekly collection plan as steps for the chair; a decision lands in the chair's Inbox when an input order passes the cooperative's envelope (approve / split the order, with a deadline); the humans keep the weighing, the price and the payment; the receipt records each lot and each input order, and the shared bill splits a group purchase among the members
The engagement on top
a custom engagement: the cooperative's member register and weighbridge sheet wired to the room through connectors loaded on this pull, a cooperative accountant or agronomist as the paid professional on the match, and the sector pack's task templates for smallholder upkeep rounds; farmers reached over WhatsApp
Camera-based targeted spraying: a sprayer detects weeds in real time and sprays only them; a university extension demonstration on five soybean fields (415 acres) saved herbicide in every field
Where
US, North America
Segment
SMB and mid (row-crop farms)
Problem
blanket herbicide spraying wastes product and money
Technology
John Deere See & Spray Ultimate (camera and machine-learning spot spraying)
Who uses it
Iowa State University Extension and Outreach (US): Doug Houser, Matt Darr, Ryan Huffman; five farm fields in Boone and Story counties
Outcome
Average herbicide saving 76 % over five soybean fields (range 43.9 % to 90.6 %), nearly $6,500 or $15.70 per acre (Iowa State Extension)
Effort
high
Risks
Savings fall with heavy weed pressure; machine cost; only five fields and one season.
How to adopt
Check the sprayer fits your fleet, run one season on a few fields, compare product use with your usual practice, and keep weed-pressure records.
How a room runs it
the ROOM for a row-crop farm; the farm owner, the sprayer operator, the agronomist and a crop-protection teammate; the agent reads the spray-job logs the sprayer already records and GENERATES the spraying rounds for the operator (field, product, window) with due dates as plan steps with owners and due dates, and raises a decision for the owner when a field's herbicide spend passes its envelope (approve / hold) in the Inbox as a proposal with a deadline; the humans keep the spraying, the product choice and the spend above the envelope; the receipt records each field sprayed with product and volume; budgets degrade before they block
The engagement on top
a custom engagement: the sprayer's job-log export and the input invoices wired to the room through connectors loaded on this pull, an agronomist as the paid professional on the match, and the sector pack's task templates for the recurring rounds; the client receives the reports and gives direction in the room
Fazendas Reunidas (Rio Verde, Goias, Brazil); CEO Alexandre Baumgart; Solinftec
Outcome
none stated (the readable text gives no results figure)
Effort
high
Risks
Capital cost; connectivity; technicians still needed; results unpublished.
How to adopt
Pilot a robot fleet on one block of a large farm, train technicians, and measure product use against a conventional block.
How a room runs it
the ROOM for a large grain farm; the farm manager, the field technician and a crop-protection teammate; the agent reads the robots' field reports and GENERATES the weekly inspect, mix-preparation and refill steps for the technician (field, product, window) with due dates as plan steps with owners and due dates, and raises a decision for the manager when a field's herbicide use passes its envelope (approve / hold) in the Inbox as a proposal with a deadline; the humans keep the set-up, the mix preparation and the spend above the envelope; the receipt records each field treated with product and date; budgets degrade before they block
The engagement on top
a custom engagement: the robot platform's field reports and the stock sheet wired to the room through connectors loaded on this pull, an agronomist as the paid professional on the match, and the sector pack's task templates for the recurring rounds; the client receives the reports and gives direction in the room
Generative AI advisory chatbot used by extension agents and farmers in local languages, linked to national farm-service infrastructure
Where
IN;KE, South Asia
Segment
public and SMB (smallholder farmers)
Problem
too few extension agents per farmer
Technology
Farmer.Chat (Digital Green with OpenAI)
Who uses it
Digital Green (non-profit), India and Kenya, later Ethiopia and Nigeria
Outcome
More than 4,500 extension agents were using the platform in India and Kenya by May 2024; early assessments reported a large drop in the cost of extension, unquantified in the article
Effort
medium
Risks
Wrong advice reaching farmers; language errors; low connectivity.
How to adopt
Give it to extension agents first, ground it in vetted sources, evaluate with a trial.
How a room runs it
the village-advice room: the extension agent, the farmer group and an agronomy teammate; the agent answers crop questions in the farmer's language from vetted sources and generates follow-up tasks (a field visit, a pest check) with dates, and anything that recommends a chemical becomes a proposal for the agent's supervisor; the receipt records the question, the source and the reply
The engagement on top
connectors to WhatsApp and the state advisory source, a local-language pack, a supervisor-approval rule
Computer-vision laser weeder that identifies and burns weeds without herbicide, across more than 100 crops
Where
US, North America
Segment
SMB and mid (vegetable and row-crop growers)
Problem
herbicide resistance and hand-weeding labour
Technology
LaserWeeder (Carbon Robotics)
Who uses it
Carbon Robotics customers, farms across the United States
Outcome
The machine eliminates over 200,000 weeds an hour and the company cites up to 80% savings in weed-control cost (Fox Business, Jun 2023; company figures)
Effort
high
Risks
Capital cost; performance in heavy residue or wet fields; company-reported savings.
How to adopt
Trial on one high-value crop, compare cost per acre with hand crews, then add fields.
How a room runs it
the weeding room: the grower, the operator and the agronomist; the agent reads each pass's weed counts and proposes the next fields and speed, generating the crew tasks with dates, and an organic-certification claim becomes a proposal the compliance lead reviews; the receipt records the field, the pass and the figures
The engagement on top
connectors to the machine data export and the farm records, a field-priority template, a certification-evidence rule
Domain-specific generative AI assistant for field staff and agronomists, trained on trial data and product labels
Where
US, North America
Segment
enterprise (crop-science company)
Problem
slow expert answers to farmer product and agronomy questions
Technology
E.L.Y. (Bayer Crop Science with Microsoft and EY, on Azure AI)
Who uses it
Bayer Crop Science, United States (global rollout planned)
Outcome
Prototype built in 90 days and shared with 1,000 employees in 2024, expert answers generated within seconds, planned for several thousand employees in 2025 (Microsoft customer story, Feb 2025) the company’s own figure
Effort
high
Risks
Wrong dosage or label advice; liability for recommendations; stale trial data.
Next: expansion to several thousand Bayer Crop Science employees in 2025 and a scalable foundation for supporting farmers globally source
How to adopt
Pilot with agronomists, restrict to approved sources, keep label compliance reviewed.
How a room runs it
the advice room: the farmer, the agronomist and the retailer; the agent drafts an answer from the label, trial data and local conditions, and each product recommendation becomes a proposal the agronomist approves; the receipt records the sources and the label version
The engagement on top
connectors to the trial database and the label library, a geography and crop filter, an agronomist-review rule
Photo-based crop diagnosis in a farmer app: the farmer photographs a pest or disease problem, an image model returns a diagnosis and crop-protection advice, geotagged uploads trigger area early warnings; features work offline; Indonesia in progressive rollout
Where
ID;IN;PK;TH;BD, Southeast Asia;South Asia
Segment
SMB (smallholder farmers)
Problem
pests and diseases misdiagnosed; advice reaches smallholders late
Syngenta Asia Pacific with Plantix; rollout in India, Pakistan, Indonesia, Thailand and Bangladesh
Outcome
93 % accuracy and 500,000 farmers targeted (Syngenta's own figures) the company’s own figure
Effort
low
Risks
Misdiagnosis harms crops; vendor figures unverified; input-sales conflict of interest.
How to adopt
Offer farmers a photo-diagnosis app that works offline; train extension agents; link advice to locally available inputs.
How a room runs it
the ROOM for a farmer group with an extension officer: the farmers, the group's chair and an agronomy teammate; a farmer posts a photo of the problem in the room over WhatsApp and the agent reads the diagnosis the farmer's app already gives, then GENERATES a task for the farmer (inspect the neighbouring plot by Thursday, spray only if confirmed) and a decision for the extension officer or the group's agronomist (confirm the treatment, with a deadline); the humans keep the diagnosis sign-off and the spend on chemicals; the receipt records the photo, the verdict and the treatment done; the agent never diagnoses on its own model
The engagement on top
a custom engagement: the farmer group's member and plot list wired to the room through a connector, the input retailer's price list, an agronomist or extension officer as the paid professional on the match, and the sector pack's task templates for pest and disease rounds; Bahasa messages over WhatsApp
Source: vendor · 2022 · vendor only; to corroborate
Community-service pilot: assembled agro-drones sprayed pesticide on three farmers' rice fields in Majannang village, Gowa, South Sulawesi (about 1 L of mixture over about 180 m2 in about 9 minutes); terraced fields made spraying less effective
Where
ID, Southeast Asia
Segment
SMB (smallholder farmers)
Problem
smallholders cannot afford drones or the skill to run them; terraced plots
Technology
agricultural drone (assembled by the team), pesticide spraying
Who uses it
three farmers in Majannang village, Parigi district, Gowa regency (Indonesia); an eight-author university-style team (affiliation not stated on the page)
Outcome
About 1 L of mixture over about 180 m2 in about 9 minutes on three farms; terraced fields made the result not optimal (the authors, SNP2M 2024)
Effort
medium
Risks
Terraced fields cut effectiveness; very small sample; pesticide handling safety.
How to adopt
Pilot a drone-spraying service on a few plots; compare time and coverage with manual spraying; account for terrain.
How a room runs it
the ROOM for a smallholder group: the farmer group's chair, the three to thirty farmers, the drone service provider and an agronomy teammate; the agent reads the farmers' plot list (plot, crop, date of last spray) and GENERATES the spraying round as plan steps for the provider (which plots, which day, which product) and a decision for each farmer (accept the service and its fee, with a deadline) in their Inbox over WhatsApp; the humans keep the flying, the consent and the payment; the receipt records the plot, the product and the date for each farmer; the shared bill splits the provider's fee across the farmers served
The engagement on top
a custom engagement: the farmer group's plot register (a sheet or the cooperative's member list) wired to the room through a connector loaded on this pull, a drone service provider or extension officer as the paid professional on the match, and the sector pack's task templates for group service rounds
Drone trial for spraying, fertiliser and seed spreading on rice paddies: 15 ha of about 5,000 ha trialled since early 2025; tasks that needed at least ten people done by one operator in a few hours (company quote)
Where
ID, Southeast Asia
Segment
public / state-owned enterprise
Problem
labour-heavy field operations on large paddy estates; chemical use
Technology
autonomous agricultural drones (XAG P100 Pro), local distributor support
Who uses it
Sang Hyang Seri (state-owned rice company, Indonesia); PT Blessed Bentara Agri Indonesia (distributor)
Outcome
Tasks needing 10+ people done by one operator in a few hours (company quote); up to 30 % chemical cut cited from ministry studies, unverified the company’s own figure
Effort
high
Risks
Vendor-only evidence; drift and chemical-safety rules; drone licensing; displaced field labour.
How to adopt
Trial spraying drones on a small share of fields with a distributor; train operators; check licensing and spray rules.
How a room runs it
the ROOM for a paddy estate: the estate manager, the field-operations lead, the drone operator and an agronomy teammate (an agent definition with the rice-farming skills); the agent reads the estate's plot list and the drone operator's flight log and GENERATES the week's spray and spread jobs for the operator as plan steps with owners and due dates (plot, product, window), and raises a decision in the manager's Inbox when a plot's chemical use passes its envelope (approve more / hold, with a deadline); the humans keep the flying, the sign-off and the spend above the envelope; the receipt records each finished job with plot, product and date; budgets degrade before they block
The engagement on top
a custom engagement: the estate's own systems wired to the room through connectors loaded on this pull (the plot register or ERP, the drone operator's flight log, the stock sheet), an agronomist as the paid professional on the match, and the sector pack's task templates for spraying and spreading rounds; operators reached over WhatsApp
Source: vendor · 2025 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 14 adoptions here, 1 were low effort, 4 medium and 9 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Unify candidate data in one platform, connect regional systems, pilot an AI agent on pre-screening with recruiters reviewing every match.
Indonesian employers using AI in recruitment: about 20% of companies, mostly for candidate screening and writing job ads
Where
ID, Southeast Asia
Segment
SMB, mid and enterprise (survey of Indonesian employers)
Problem
manual CV screening and ad writing consume recruiter time
Technology
AI CV screening, AI job-ad drafting, assessment and interview tools
Who uses it
Indonesian companies of all sizes (1,273 recruitment staff surveyed by Jobstreet)
Outcome
about 20 % of surveyed companies use AI in recruitment; among them screening 76 %, job ads 74 % (Jobstreet survey, n=1,273)
Effort
low
Risks
Screening bias; candidate privacy; over-reliance without human review.
How to adopt
Start with job-ad drafting and CV screening using a SaaS recruiting tool; keep recruiters deciding; log who was screened out.
How a room runs it
what the 80% still not using AI can start with is the smallest room: the agency owner, one recruiter and a teammate that does the two jobs the survey names: screen the CVs that arrive and draft the ad; the screen becomes tasks for the recruiter and a shortlist decision for the owner; nothing is rejected without a person's name on it; the receipt records every screen and its reason
The engagement on top
WhatsApp as the candidate door, the agency's spreadsheet or job-board export read through a connector, an HR professional on the match; ad and screen task templates
Generative AI co-pilot for the talent-acquisition team: drafts job descriptions, builds sourcing strings, creates tailored screening questions, summarises profiles and schedules screened candidates for a project
Where
IN, South Asia
Segment
enterprise (IT services)
Problem
recruiters spend days on job descriptions, skill mapping and profile screening while projects wait for staff
Technology
LLM co-pilot inside the recruitment platform (several models compared by the team)
Who uses it
HCLTech (India)
Outcome
25 % faster turnaround of the sourcing stages and an estimated 5 million USD saving a year; job descriptions 2-4 days to 30 minutes-1 day (HCLTech's figures in a Brandon Hall award entry)
Effort
medium
Risks
Estimated savings are self-reported; biased summaries; confidential candidate data sent to a model.
How to adopt
Start with the recruiters' repeatable drafting tasks (job descriptions, screening questions, profile summaries); compare models on the team's own samples; keep a recruiter reviewing every output; measure turnaround per stage before and after.
How a room runs it
the ROOM for a recruiting engagement: the talent-acquisition head, the recruiters, the delivery managers and a sourcing teammate (an agent definition with the company's job-family rules); the agent drafts job descriptions and screening questions and GENERATES the sourcing plan as steps with owners and due dates (shortlist for project X by Thursday, confirm skill adjacencies with the delivery manager), and puts each shortlist into the delivery manager's Inbox as a proposal with a deadline; the humans keep the interview and the hiring decision; the receipt records which profiles the agent summarised and who accepted each shortlist
The engagement on top
connectors to the recruitment platform and the skills catalogue, loaded on this pull; a paid talent-acquisition practitioner on the match
Source: analyst · 2024 · vendor only; to corroborate
Agentic AI across the recruiting lifecycle: one-to-one applicant engagement, pre-screening with match scores, automated job creation and interview preparation on a unified candidate data layer
Where
CH, Europe
Segment
enterprise
Problem
slow, manual high-volume pre-screening and campaign preparation; 60+ regional systems
Technology
Salesforce Data 360, Agentforce AI agents, match scoring
Who uses it
The Adecco Group (Switzerland, global staffing)
Outcome
Candidates screened are 2x as likely to be hired; hiring 28 % faster; campaign prep -93 % (supplier's customer story, measurer not stated) the company’s own figure
Effort
high
Risks
Bias and fairness in screening; candidate consent and data protection; vendor figures unaudited.
How to adopt
Unify candidate data in one platform, connect regional systems, pilot an AI agent on pre-screening with recruiters reviewing every match.
How a room runs it
the ROOM for a staffing desk: the account manager, two recruiters, the client's hiring lead and a recruiting teammate (an agent definition with the staffing skills); the agent reads the candidate pool and the open requisitions the agency already has and GENERATES the day's tasks as plan steps with owners and due dates (call these six candidates by Wednesday, prepare the interview brief for the client), and puts the shortlist into the client lead's Inbox as a decision with a deadline; the humans keep the interview, the offer and the placement fee; the receipt records who was screened, who was submitted and who signed off
The engagement on top
a custom engagement: the agency's own ATS and job boards wired to the room through connectors loaded on this pull, candidates reached over WhatsApp (the channel is built), a paid recruitment professional on the match, the staffing pack's task templates (requisition intake, screen, submit, interview prep)
Source: vendor · 2025 · vendor only; to corroborate
AI-matched talent platform: Randstad runs one global talent platform and connects Workday's Recruiting Agent to its talent network so recruiters surface qualified past applicants and network candidates inside existing workflows
Where
NL, Europe
Segment
enterprise
Problem
time-to-fill and screening effort; talent scarcity
500,000 shifts filled in Q1 2025 (company release); 54 % capacity claim is Workday's the company’s own figure
Effort
high
Risks
Bias in matching; regional availability limits; dependence on two vendors.
How to adopt
Consolidate talent data on one platform; connect an AI recruiting agent into the existing HR system; run in one market first.
How a room runs it
the same staffing room, with the client in it: the agent proposes the candidate leads as tasks for the recruiter (reach out, verify availability) and a submit-or-hold decision for the account manager; the matching stays a suggestion the person accepts; budgets for outreach degrade (fewer channels first) before they block; the receipt records which leads were contacted and which were submitted
The engagement on top
connectors to the agency's ATS or HRIS and its candidate database; the recruiter's own judgement as the paid professional; a shift-fill task template for high-volume clients
Source: vendor · 2025 · vendor only; to corroborate
Conversational AI hiring assistant that chats with applicants, answers their questions, collects basic details, schedules interviews with the hiring manager and sends offers to the candidates the manager selects; rolled out to 3,500+ restaurants in North America and Europe
Where
US, North America
Segment
enterprise (restaurants, 3,500+ sites)
Problem
restaurant general managers lose hours to administrative hiring tasks while hourly roles stay open
Technology
Paradox conversational AI hiring platform (assistant named Ava Cado), in English, Spanish, French and German
Who uses it
Chipotle Mexican Grill (US); Paradox
Outcome
Time to hire for in-restaurant roles expected to fall by up to 75 % (Chipotle's expectation at launch, not a measured result); no measured figure stated on the page the company’s own figure
Effort
medium
Risks
Bias in screening; candidates who cannot reach a human; data protection for applicant data; vendor claims not independently measured.
How to adopt
Choose a hiring assistant that connects to the applicant tracking system and calendars; write the screening questions and the human hand-off points first; pilot in a few sites for hourly roles, then widen; audit candidate outcomes for bias.
How a room runs it
the ROOM for a multi-site hiring engagement: the HR lead, the area managers and a recruiting teammate (an agent definition with the company's hiring rules); the agent reads open roles and applicants the room is allowed to see and GENERATES the hiring plan as steps with owners and due dates (confirm the interview slots for site 12 by Friday, send the shortlist to the site manager), and puts each offer and each exception into the manager's Inbox as a proposal with a deadline; the humans keep who is hired and the offer terms; the receipt records what was asked of each candidate and who approved each offer
The engagement on top
connectors to the applicant tracking system and the calendar, loaded on this pull; a paid HR practitioner on the match to set the screening questions and the fairness checks
Your own team builds it. Across the 5 adoptions here, 1 were low effort, 2 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Put employee, attendance and leave data into an HRIS; run payroll in parallel one cycle before cutover; train HR staff.
HR virtual agent that answers employees' questions (payslips, leave policy) and automates about 80 HR tasks such as letters, vacation requests and manager workflows; complex cases go to human advisors; connected to the HR, ERP and expense systems
Where
US, North America
Segment
enterprise (global)
Problem
a global workforce sends millions of routine HR questions to a service centre
Technology
AskHR virtual agent with watsonx Orchestrate agentic layer, connected to Workday, SAP and Concur
Who uses it
IBM (US)
Outcome
94 % containment of common questions, 75 % fewer support tickets since 2016, 40 % lower HR operating cost over four years, 11.5 million interactions in 2024 (IBM's own figures) the company’s own figure
Effort
high
Risks
Wrong policy answers; access to personal records; self-reported containment definition; employee trust.
How to adopt
Index the HR policies and connect the agent to the HR system with read access first; define which requests the agent may complete and which go to an advisor; add task automation one request type at a time; track containment and escalation quality.
How a room runs it
the ROOM for an HR-service engagement: the HR director, the HR advisors, the managers and an HR-desk teammate (an agent definition with the company's policies); the agent answers routine questions from the policies and the employee's own record and GENERATES the follow-ups as plan steps with owners and due dates (update the leave policy answer by Monday, route the compensation change to payroll), and puts each exception and each compensation change into the manager's or advisor's Inbox as a proposal with a deadline; the humans keep exceptions and anything touching pay; the receipt records every answer's policy source and who approved each change
The engagement on top
connectors to the HR information system and the policy library, loaded on this pull; a paid HR-operations practitioner on the match
Source: vendor · 2025 · vendor only; to corroborate
HR chatbot integrated with SAP SuccessFactors on Workplace: leave, travel and overtime requests and policy questions in English and Vietnamese
Where
VN, Southeast Asia
Segment
enterprise (Southeast Asia)
Problem
approvals consume executive time; policies hard to find
Technology
Leena AI chatbot, SAP SuccessFactors, attendance systems, Workplace
Who uses it
Coca-Cola Vietnam
Outcome
50 % of travel and 60 % of leave requests via the bot, approval effort -60 % (company and supplier figures)
Effort
medium
Risks
Misrouted approvals; language quality; employees bypassing the bot.
How to adopt
Integrate a chatbot with the HRIS and the workplace chat tool; start with leave, travel and overtime requests; add local-language support.
How a room runs it
the staff asks in chat and the request becomes a task with an owner: leave and overtime go to the manager as a proposal with a deadline, the HR teammate checks the policy first; the manager's approvals take minutes instead of a queue; the humans keep the approval; the receipt records each request and its approver
The engagement on top
connectors to SAP SuccessFactors or the company's HRIS and attendance system, WhatsApp as the door, an HR professional on the match
HR chatbot connected to the HR information system that handles leave, payroll and policy questions, automates leave applications and approvals and raises a ticket for a business partner when a query is personal or complex
several HR systems and portals run apart, and routine requests such as leave balance or payslip fall on HR business partners
Technology
Leena AI chatbot (Electra) integrated with SAP SuccessFactors
Who uses it
KEC International (India); Leena AI
Outcome
Nearly 90 % of queries resolved by the bot, query resolution time down 80 %, 5,000+ HR hours saved in a quarter, 9x ROI in the quarter (advertorial; no named measurer)
Effort
medium
Risks
Unattributed figures in an advertorial; wrong answers on personal data; low adoption if the bot cannot act.
How to adopt
Connect the chatbot to the HR information system for leave and payroll; collect the real wording of employee questions; route personal or complex queries to a ticket; measure resolution time before and after.
How a room runs it
the ROOM for an HR-desk engagement: the HR head, the HR business partners, the employees' managers and an HR-desk teammate (an agent definition with the company's leave and payroll rules); the agent answers from the HR record and GENERATES the follow-ups as plan steps with owners and due dates (fix the unclear travel-policy answer by Friday, approve the leave backlog), and puts leave approvals and policy exceptions into the business partner's Inbox as proposals with a deadline; the humans keep exceptions and personal cases; the receipt records each answer's source and each approval
The engagement on top
connectors to the HR information system, loaded on this pull; a paid HR-operations practitioner on the match to map the real wording of employee questions
Source: press · 2019 · vendor only; to corroborate
HR chatbot tied to the HR information system that answers each employee's questions with their own records (leave balance, claim status) and policy
Where
SG, Southeast Asia
Segment
enterprise (Southeast Asia)
Problem
HR queries swamp the HR team
Technology
AI chatbot integrated with HRIS, voice and text
Who uses it
OCBC Bank (Singapore)
Outcome
none stated
Effort
medium
Risks
Wrong answers on personal records; access control to HR data; maintenance of policy content.
How to adopt
Connect a chatbot to the HR information system so answers use the employee's own records; start with leave and claims questions.
How a room runs it
the HR room with the employees reachable: the HR partner, the line managers, the head of HR and an HR teammate that answers from the policies and records it is allowed to read; a question that needs a person (an exception to the leave policy) becomes a task for the HR partner and a decision for the manager, with a deadline; masked fields stay masked; the receipt records what was asked, answered and escalated
The engagement on top
connectors to the HRIS and the policy library, WhatsApp as the staff door, the HR professional for exceptions; leave, claim and onboarding task templates
HRIS with automated payroll: employee database, digital attendance, shift and leave management, payroll with gross-to-net simulation replacing Excel and email
Where
ID, Southeast Asia
Segment
mid (Indonesia)
Problem
manual HR in Excel and email; payroll slow and error-prone after the TER tax rules
Technology
Mekari Talenta HRIS and payroll
Who uses it
Olympus Medical Indonesia (subsidiary of Olympus Japan)
Outcome
Payroll '80 % faster' (supplier's figure) the company’s own figure
Effort
medium
Risks
Payroll errors during migration; employee data privacy; vendor lock-in.
How to adopt
Put employee, attendance and leave data into an HRIS; run payroll in parallel one cycle before cutover; train HR staff.
How a room runs it
the ROOM for a small HR and payroll desk: the HR manager, the finance lead, the country head and an HR teammate (an agent definition with payroll and attendance skills); each month the agent reads the attendance and leave the company already holds and GENERATES the payroll close as plan steps with owners and due dates (confirm overtime, approve adjustments, sign the run), and each salary adjustment as a decision in the head's Inbox with a deadline; the humans keep the approval and the money release; the receipt records every adjustment and who approved it
The engagement on top
connectors to the company's HRIS and attendance export (Talenta or its peer), the payroll and tax professional on the match, the HR pack's monthly close templates
Source: vendor · 2024 · vendor only; to corroborate
AI interviews added to the recruitment process of an Indonesian insurer while keeping hiring quality
Where
ID, Southeast Asia
Segment
enterprise (Indonesia)
Problem
hiring volume and interview scheduling
Technology
AI interview tool
Who uses it
AIA Indonesia
Outcome
none stated
Effort
medium
Risks
Evidence is one vendor sentence; bias; candidate consent.
How to adopt
Add an AI interview step to one hiring stream; define the quality metric first and compare with human interviews.
How a room runs it
the in-house HR room runs the interview round: the recruiter, the hiring manager and a teammate that schedules, runs the first structured interview questions and writes the summary; the hiring manager gets the candidate decision in the Inbox with a deadline; the humans keep the offer and the final interview; the receipt records the questions, the summary and who decided
The engagement on top
connectors to the company's ATS, WhatsApp for scheduling, an HR professional on the match
Source: vendor · 2024 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 6 adoptions here, 0 were low effort, 5 medium and 1 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Index the firm's own knowledge base and build an assistant that cites sources and names experts; restrict to firm-approved content.
Internal generative AI platform used for drafting, research, code and project plans with task prioritisation, rolled out to 75,000 staff
Where
GB;AE, Europe;Middle East
Segment
enterprise
Problem
routine drafting and research slow billable work
Technology
PairD (Deloitte AI Institute, in-house)
Who uses it
Deloitte (Europe and Middle East; also shared with the charity Scope)
Outcome
Rolled out to 75,000 employees across Europe and the Middle East, 100,000 expected within six months (Deloitte, to CIO Dive); no productivity figure
Effort
high
Risks
Data confidentiality; quality of output; staff over-reliance.
How to adopt
Provide a secured internal generative-AI platform with training and usage policy; roll out in waves.
How a room runs it
the small-firm version of the same room: the principal, two consultants and a teammate that drafts the proposal, the research brief and the project plan; the plan lands as tasks with owners and dates, the proposal as a decision for the principal with a deadline; the consultant reviews every draft; the receipt records what the agent drafted and who edited it
The engagement on top
connectors to the firm's proposal library and CRM, a senior consultant on the match as reviewer, proposal and brief task templates
Firm-wide enterprise chat assistant for 100,000+ staff, with custom GPTs for tax review, proposals and developer help, and resale to clients
Where
US;GB, global
Segment
enterprise
Problem
slow proposal writing, review and research across a very large firm
Technology
ChatGPT Enterprise (GPT-4o), custom GPTs
Who uses it
PwC (US and UK) with OpenAI
Outcome
100,000+ employees in the US and UK given access; 3,000+ internal generative AI use cases identified, nearly 40% addressed by deployed internal tools (PwC figures via CIO Dive, May 2024)
Effort
high
Risks
Client confidentiality; hallucinated findings in advice; professional liability.
Next: PwC is building custom GPTs for tax-return review, proposal responses and developer assistance, and PwC becomes the first reseller so clients can buy access through the firm source
How to adopt
Start with enterprise licences and a use-case inventory; build custom assistants per workflow; keep partner sign-off.
How a room runs it
the engagement room: the partner, the manager, the associates and a drafting teammate; the agent drafts the proposal response and the review checklist and GENERATES the review tasks per associate with due dates, and each client-facing claim becomes a proposal in the partner's Inbox with a deadline; the humans keep the judgement and the sign-off; the receipt records the draft, the reviewer and the sources
The engagement on top
connectors to the firm's document store and knowledge base, a senior reviewer on the match, proposal and tax-review task templates
Firm-wide generative AI assistant that searches the firm's knowledge base, summarises the five to seven most relevant items with links and names the experts
Where
US, North America
Segment
enterprise
Problem
consultants cannot find prior work and experts across 100,000+ documents
Technology
Lilli (in-house generative AI platform over 40+ knowledge sources)
Who uses it
McKinsey & Company (global)
Outcome
Around 7,000 employees had access to the platform in Aug 2023; it aggregates 40+ knowledge sources and 100,000+ documents (McKinsey, to CIO Dive); no time-saved figure
Effort
high
Risks
Confidential client data leakage; hallucinated answers; source quality.
How to adopt
Index the firm's own knowledge base and build an assistant that cites sources and names experts; restrict to firm-approved content.
How a room runs it
the ROOM for an engagement: the partner, the consultants, the client lead and a research teammate (an agent definition with the firm's method); the agent reads the firm's prior decks and the client's documents it is allowed to see and GENERATES the engagement plan as steps with owners and due dates (pull three comparable cases by Tuesday, draft the interview guide), and puts the recommendation memo into the client lead's Inbox as a decision with a deadline; the humans keep the judgement, the client conversation and the sign-off; the receipt records sources used and who accepted each deliverable
The engagement on top
connectors to the firm's document store and the client's data room, a paid practitioner on the match, the consulting pack's engagement templates (scoping, interviews, synthesis, memo)
Vendor-only evidence; outdated content; confidentiality across offices.
How to adopt
Build a searchable index of past proposals and RFPs; recommend related past work to bid teams.
How a room runs it
the bid room across offices: the regional lead, the bid manager and a teammate that reads the RFP, finds the firm's closest past proposals and turns them into section tasks for named authors with due dates; the go or no-go is a decision for the regional lead; the receipt records which past work was reused
The engagement on top
connectors to the firm's proposal store and file shares; templates for the reuse-and-adapt steps
Source: vendor · 2026 · vendor only; to corroborate
AI proposal-writing system built on OpenAI models on AWS for an IT consulting firm
Where
multi, global
Segment
mid (IT consulting)
Problem
proposals written from scratch under time pressure
Technology
OpenAI models, AWS serverless, built by Toptal
Who uses it
an unnamed IT consulting firm (Toptal client)
Outcome
Proposals 20 % faster (supplier's figure) the company’s own figure
Effort
medium
Risks
Vendor-only evidence; client data in prompts; generic text.
How to adopt
Feed past proposals into a drafting tool on a cloud model; keep a human editor on every proposal.
How a room runs it
the sales room of a consulting firm: the partner, the bid manager, a delivery lead and a proposal teammate that reads the RFP and the firm's past proposals and GENERATES the response plan (sections, owners, due dates) and the draft; the pricing and the go or no-go land in the partner's Inbox as decisions with the bid deadline; the humans keep price and promise; the receipt records the version submitted
The engagement on top
connectors to the firm's proposal folder and CRM; the bid professional on the match; RFP response templates
Source: vendor · 2026 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 5 adoptions here, 0 were low effort, 2 medium and 3 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Build an assistant on existing app and messaging channels with an escalation path to human operators.
A bank grew ChatGPT from 3,300 accounts to 11,000 and in December 2025 announced a rollout of ChatGPT Enterprise to all 120,000 staff, with staff building custom assistants
Where
ES, Europe
Segment
enterprise
Problem
Routine writing and analysis work eats staff hours across 25 markets
Technology
ChatGPT Enterprise, custom GPTs, Google Workspace with Gemini for about 100,000 more staff
Who uses it
BBVA, Spain (25 markets)
Outcome
Staff reportedly saved nearly three hours a week and more than 80 % used it daily (as reported; measurer not stated)
Effort
medium
Risks
Self-reported savings; data governance across 25 markets; dependence on one supplier.
How to adopt
Pilot with a few thousand seats, let staff build and share their own assistants, measure use, then extend to the whole workforce with policy and training.
How a room runs it
the ROOM for a bank-wide rollout: the head of the programme, each country lead and an adoption teammate; the teammate GENERATES the wave plan as steps with owners and due dates (which country, which team, which training) and the policy exceptions as decisions in the risk officer's Inbox with a deadline; the humans keep the risk sign-off and the customer-facing use; the receipt records every wave, every exception and who approved it
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the bank's identity, document and training systems), and a change-management professional joins as the paid professional on the match
A bank gave its staff a portal to several external language models for drafting, summarising and spreadsheet work, and then extended it from one division to tens of thousands of employees
Where
US, North America
Segment
enterprise
Problem
Staff write, summarise and research by hand across a very large organisation, and no single model supplier should be a lock-in
Technology
LLM Suite portal over external large language models (launched with OpenAI), internal marketing and meeting-summary uses
Who uses it
JPMorgan Chase, United States
Outcome
More than 60,000 employees had access (CNBC report, August 2024); no productivity figure stated
Effort
medium
Risks
Data leaving the firm through prompts; lock-in to one supplier; uneven use by staff.
How to adopt
Choose a portal that can call more than one model supplier; start with one division, define which documents may enter, then widen access in stages with training.
How a room runs it
the ROOM for a company-wide assistant: the COO, each department head and a staff-assistant teammate; the teammate GENERATES the department rollout plan as steps with owners and due dates (who is trained, which documents may be summarised) and files the policy questions as decisions in the compliance lead's Inbox with a deadline; the humans keep the sign-off on what the assistant may touch; the receipt records who asked what, with which model and which documents
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the document stores, the identity provider), and an AI-adoption lead joins as the paid professional on the match
A telco runs a virtual assistant on its app, web and messaging channels that handles top-ups, bills, packages, branch appointments and complaints and hands complex cases to a human operator
Where
ID, Southeast Asia
Segment
enterprise
Problem
call-centre load and wait time on routine requests
Technology
Veronika virtual assistant (Kata.ai, later rebuilt on Azure OpenAI), WhatsApp, Messenger, LINE, Telegram
Who uses it
Telkomsel, Indonesia
Outcome
96 % of incoming Messenger inquiries handled by the bot and about 30 % operating-cost saving (Facebook for Developers success story, vendor-reported; the saving is called 'potential' in one section); the 97 % / 5 % Kata.ai figures could not be re-verified the company’s own figure
Effort
medium
Risks
Vendor figures; wrong answers on billing; customer frustration if handoff fails.
How to adopt
Build an assistant on existing app and messaging channels with an escalation path to human operators.
How a room runs it
the ROOM for the customer-service lead: the service head, the shift supervisors and a front-line teammate that answers from the company's knowledge base; the agent GENERATES the supervisors' tasks as plan steps with owners and due dates (review the ten escalations of the day, fix the answer that caused them, brief the evening shift) and files in the head's Inbox a decision on every refund or exception above the envelope with the SLA as deadline; the humans keep the exception, the empathy and the sign-off; the receipt records each handover, each refund and the cost per resolved case
The engagement on top
a custom engagement: the company's own helpdesk, billing and knowledge systems and WhatsApp number wired to the room through connectors loaded on this pull; a customer-experience professional joins as the paid professional on the match; the customer-service pack carries escalation and macro-review templates
A telco gave employees a virtual HR assistant that handles administrative requests and routine HR questions
Where
ID, Southeast Asia
Segment
enterprise
Problem
HR staff answering the same administrative questions repeatedly
Technology
ASTRID virtual HR assistant
Who uses it
Indosat, Indonesia
Outcome
none stated
Effort
medium
Risks
No outcome data; personal-data access; limited scope.
How to adopt
Add a virtual assistant to the staff app for HR requests and routine questions; connect to HR records.
How a room runs it
the ROOM for the people-operations desk: the HR manager, the HR officers and a staff-help teammate that answers from the HR handbook; the agent GENERATES the desk's tasks as plan steps with owners and due dates (process the three leave exceptions, update the policy answer staff keep asking about) and files in the HR manager's Inbox a decision on each exception with the payroll cut-off as deadline; the humans keep every personal or sensitive case; the receipt records each request, answer and approval, with clearance applied so staff see only their own rows
The engagement on top
a custom engagement: the company's own HRIS, payroll and IT-ticket system wired to the room through connectors loaded on this pull (an internal IT helpdesk is the same room with a different handbook); an HR or IT service professional joins as the paid professional on the match; the corporate pack carries request-routing templates
A telecom group launched an in-house AI legal intelligence system that searches and summarises internal policy documents for staff, with later updates planned to help draft and update internal rules
Where
ID, Southeast Asia
Segment
enterprise
Problem
staff lose hours searching and reading policy documents
Telkom Indonesia (Telkom Group legal unit first), Indonesia
Outcome
40 minutes saved per employee and 24x faster analysis (company's internal figures)
Effort
medium
Risks
Incorrect summaries of binding rules; internal figures unaudited; access controls.
How to adopt
Index internal policies and regulations; start with search and summary in one legal unit; add drafting later.
How a room runs it
the ROOM for a corporate function (legal, finance ops, procurement): the function head, the officers and a policy teammate that reads the rules the company already has; the agent GENERATES the function's recurring work as plan steps with owners and due dates (check the new contract against clause 4, refresh the expiring policy, answer the three open staff questions) and files in the head's Inbox a decision when two policies conflict (which wins) with the review date as deadline; the humans keep the ruling and the approval, and the agent assists, not replaces; the receipt records the source clause, the answer given and who approved a changed rule
The engagement on top
a custom engagement: the company's own policy repository, contract store and ERP wired to the room through connectors loaded on this pull; an in-house counsel or procurement professional joins as the paid professional on the match; the corporate pack carries policy-review and approval-routing templates
A bank gave staff AI agents: a relationship-manager agent that ties market sentiment to client milestones, and a contact-centre agent that surfaces procedures and product facts during live calls and flags interaction trends for training
Where
ID, Southeast Asia
Segment
enterprise
Problem
service staff search internal knowledge during calls; advisors need context for consultations
Technology
Google Cloud (Gemini-based agents) with Artefact, an AI knowledge-management system
Who uses it
CIMB Niaga, Indonesia
Outcome
none stated
Effort
high
Risks
No measured results; mis-selling risk in banking; data privacy.
How to adopt
Define staff use cases first; connect agents to product and procedure knowledge; pilot with a small contact-centre team.
How a room runs it
the ROOM for a service team: the contact-centre manager, the team leads, the trainer and a knowledge teammate that answers from the bank's procedures; the agent GENERATES the team's weekly tasks as plan steps with owners and due dates (update the procedure that agents searched for most, schedule a coaching slot on the trend it flagged) and files in the manager's Inbox a proposal when a procedure should change (publish or hold) with the policy review date as deadline; the staff keep the customer conversation and the compliance sign-off; the receipt records every change to a procedure and who approved it
The engagement on top
a custom engagement: the company's own knowledge base, ticketing and telephony systems wired to the room through connectors loaded on this pull; a compliance professional joins as the paid professional on the match; the customer-service pack carries procedure-review and coaching templates
Source: vendor · 2026 · vendor only; to corroborate
A telecom and technology group committed to deploy a supplier's enterprise AI to all group companies and to automate over 100 million internal workflows
Where
JP, East Asia
Segment
enterprise
Problem
Group companies run many manual workflows and want agents trained on their own data
Technology
ChatGPT Enterprise, a secure data-training and fine-tuning environment, AI agents, a 50:50 joint venture to sell to Japanese companies
Who uses it
SoftBank Group and SoftBank Corp., Japan
Outcome
none stated (a plan to automate over 100 million workflows, not a result)
Effort
high
Risks
A plan, not a result; large spend; data security when training on internal systems.
How to adopt
Agree a group-level supply deal, give staff the assistant first, inventory the workflows, then build agents on the group's own data in a controlled environment.
How a room runs it
the ROOM for a group workflow programme: the group CIO, each company's operations lead and an automation teammate; the teammate GENERATES the workflow backlog as plan steps with owners and due dates (which process, which data, which check) and files each go-live as a decision in the owning executive's Inbox with a deadline; the humans keep the choice of what is automated and the exceptions; the receipt records each workflow, its owner and its sign-off
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the internal workflow systems and data stores), and an automation architect joins as the paid professional on the match
Source: vendor · 2025 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 7 adoptions here, 0 were low effort, 5 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Index past rulings; flag similar cases to prevent conflicting decisions; judges keep the decision; pilot in a few courts.
Firm-wide Harvey with LexisNexis content for research, litigation preparation, transcript review, contract and M&A diligence analysis, after a 100+ attorney pilot
Where
US, North America
Segment
mid
Problem
manual transcript review, clause analysis and research
Technology
Harvey, Ask LexisNexis
Who uses it
Stinson (US law firm)
Outcome
Pilot group of more than 100 attorneys before firm-wide adoption (Stinson press release); no time-saved figure
Effort
high
Risks
Single release source; accuracy; client confidentiality.
How to adopt
Pilot with 100+ attorneys; add licensed legal content; firm-wide with review rules.
How a room runs it
the litigation room: the partner, associates, a paralegal and a teammate that turns transcripts into deposition outlines and tasks with owners and due dates; legal-hold or filing deadlines are tasks with dates; strategy choices go to the partner as decisions; the receipt records the source pages behind every outline line
The engagement on top
connectors to the firm's DMS and research subscriptions, a paralegal or lawyer on the match, deposition and diligence templates
Source: press · 2026 · vendor only; to corroborate
Firm-wide adoption of Harvey for contract review, drafting, due diligence, data extraction and summaries, hosted onshore in Australia
Where
AU, Oceania
Segment
mid and enterprise
Problem
time-intensive review work; data residency
Technology
Harvey agentic AI, secure AU environment
Who uses it
Clayton Utz (Australia)
Outcome
Pilot of over 100 lawyers before the firm-wide adoption (Clayton Utz's head of AI); efficiencies not quantified
Effort
high
Risks
Firm's own release only; confidentiality; accuracy.
How to adopt
Pilot with 100+ lawyers; host data in-country; then firm-wide with review rules.
How a room runs it
the same matter room with the residency promise visible: the review stays in the firm's chosen jurisdiction; the agent turns a contract batch into review tasks with owners and due dates and the exceptions into the partner's Inbox; the humans keep the advice; the receipt records which clauses were flagged and who cleared them
The engagement on top
connectors to the firm's DMS, a lawyer on the match, contract-review task templates; where data must stay onshore, the deployment region is part of the engagement
Firm-wide legal AI after a 6-week experiment with 200+ lawyers, rolled out to 1,200+ lawyers in six offices
Where
AU;SG, Oceania
Segment
enterprise
Problem
slow drafting, research and document review
Technology
Harvey (legal AI)
Who uses it
King & Wood Mallesons Australia (Australia and Singapore offices)
Outcome
Experiment with about 200 lawyers before rollout to 1,200+ lawyers across 6 offices; firm said it surfaced hundreds of use cases (Lawyers Weekly, May 2025)
Effort
high
Risks
Hallucinated citations; client confidentiality; professional liability.
Next: the firm calls the rollout the beginning of a broader push including possible client collaborations and co-development source
How to adopt
Run a time-boxed lawyer experiment with ten criteria; train everyone at launch; require lawyer review of output.
How a room runs it
the matter room: the partner, the associates, the client's counsel and a drafting teammate; the agent drafts the clause analysis and chronology and GENERATES the review tasks per associate with due dates, and each risk finding becomes a proposal in the partner's Inbox with a deadline; the humans keep the review and the advice; the receipt records the draft, the reviewer and the changes
The engagement on top
connectors to the firm's DMS and document vault, a lawyer on the match, a matter-review task template set
Firm-wide generative AI for contract analysis, due diligence, litigation and regulatory compliance after a beta with 3,500 lawyers and 40,000 queries
Where
GB, Europe
Segment
enterprise
Problem
slow first drafts and diligence
Technology
Harvey (OpenAI-backed legal AI)
Who uses it
Allen & Overy (43 offices; now A&O Shearman after the 2024 combination with Shearman & Sterling)
Outcome
Beta with about 3,500 lawyers and 40,000 queries before firm-wide access (firm's figures, Feb 2023); in April 2025 A&O Shearman and Harvey announced agents for antitrust filing analysis, cybersecurity, fund formation and loan review
Effort
high
Risks
Hallucinated citations; client confidentiality; professional liability.
How to adopt
Run a lawyer beta first; require lawyer review of all output; deploy firm-wide with guidance.
How a room runs it
the matter room: the partner, the associates, the client's counsel and a drafting teammate; the agent drafts the clause analysis and diligence checklist and GENERATES the review tasks per associate with due dates, and each risk finding becomes a proposal in the partner's Inbox with a deadline; the humans keep the review and the advice, as the firm itself requires; the receipt records the draft, the reviewer and the changes
The engagement on top
connectors to the firm's DMS and the data room, a lawyer on the match, a due-diligence task template set
Legal AI for drafting, contract review and summarising, tied to a workforce-skills programme, at a Singapore firm with 1,000+ fee earners
Where
SG, Southeast Asia
Segment
enterprise
Problem
routine drafting and review slowing lawyers
Technology
Harvey (legal AI), Microsoft 365 Copilot
Who uses it
Rajah & Tann Singapore (Rajah & Tann Asia)
Outcome
Firm began with Microsoft 365 Copilot in 2023, then partnered with Harvey in Sep 2024 across a firm of over 1,000 fee earners; the firm appointed a CISO to oversee data handling (firm release) the company’s own figure
Effort
medium
Risks
Client confidentiality; hallucinated authorities; uneven adoption between teams.
Next: Harvey said in March 2026 it will open a Singapore office in June 2026, listing seven Singapore law firms as customers source
How to adopt
Start with an assistant on the existing suite, add a legal-specific tool, appoint a security owner, train the workforce.
How a room runs it
the matter room: the partner, the associates, the knowledge lawyer and a drafting teammate; the agent drafts from the firm's own precedents and GENERATES the review tasks per associate with due dates, and each non-standard clause becomes a proposal in the partner's Inbox with a deadline; the humans keep the advice; the receipt records the draft, the precedent used and the reviewer
The engagement on top
connectors to the firm's precedent bank and DMS, a knowledge lawyer on the match, a drafting task template set
AI-first strategy: a generative legal AI on a pilot basis and a second legal platform for lawyers (due diligence, fact cross-checking, litigation support, first drafts, research), and Copilot and ChatGPT Plus for business services (knowledge management, business development, resourcing, financial analysis)
Where
IN, South Asia
Segment
enterprise (full-service law firm)
Problem
lawyers spend hours on diligence, drafting and research, and business teams lack a shared knowledge system
Technology
Harvey (pilot), Lucio, Microsoft Copilot, ChatGPT Plus
Who uses it
Cyril Amarchand Mangaldas (India)
Outcome
none stated
Effort
medium
Risks
Client confidentiality; hallucinated citations; no measured result yet.
How to adopt
Pilot a legal AI with a few practice groups on matters the firm can review; set confidentiality and review rules first; extend to business services with a general assistant under the firm's data policy.
How a room runs it
the ROOM for a law-firm engagement: the partner, the associates, the knowledge-management lead and a research teammate (an agent definition with the firm's playbooks); the agent reads the data-room documents it is allowed to see and GENERATES the diligence plan as steps with owners and due dates (extract the change-of-control clauses by Wednesday, cross-check the facts in the draft), and puts each draft and each flagged risk into the partner's Inbox as a proposal with a deadline; the humans keep the legal judgement and the sign-off; the receipt records the sources used and who accepted each deliverable
The engagement on top
connectors to the document management system and the matter database, loaded on this pull; a paid legal-operations practitioner on the match
Source: press · 2025 · vendor only; to corroborate
Supreme Court decision-support system that flags similar cases early to prevent conflicting rulings; a separate AI app assigns cassation and review panels
Where
ID, Southeast Asia
Segment
public
Problem
conflicting rulings on similar cases; panel assignment
Technology
AI-based decision support system, case-similarity
Who uses it
Mahkamah Agung (Indonesia), piloted in the general, religious, military and administrative courts in Jakarta and surroundings
Outcome
none stated
Effort
high
Risks
Bias; judicial independence; national use not verified since.
How to adopt
Index past rulings; flag similar cases to prevent conflicting decisions; judges keep the decision; pilot in a few courts.
How a room runs it
the ROOM for a legal department or a court registry: the head of legal, the lawyers and a research teammate (an agent definition with the matter vocabulary); the agent reads the cases and files the office already has and GENERATES the matter plan (check for similar prior positions, draft the memo, file by the deadline) as steps with owners and dates, and any inconsistency it finds becomes a flag in the lawyer's Inbox, not a conclusion; the humans keep the judgement and the signature; the receipt records the authorities consulted
The engagement on top
connectors to the document management system and the case register, a lawyer on the match, matter-intake and precedent-check task templates
Your own team builds it. Across the 7 adoptions here, 0 were low effort, 2 medium and 5 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Build or buy a giving app with recurring debit, volunteer sign-up and public reporting.
Member-only AI advisor built on the association's own business guides and articles; each question is reported to the local association for follow-up and in-person consultation
Where
IT, Europe
Segment
mid (membership body)
Problem
members need quick advice; staff cannot answer all queries
Technology
Le Bussole AI on Microsoft Azure AI, retrieval over the association's own library
Who uses it
Confcommercio (Italy, retail associations)
Outcome
none stated
Effort
medium
Risks
Out-of-date guidance; liability for advice; no usage figures.
How to adopt
Build a member-only assistant on the association's own guides; log questions for local follow-up.
How a room runs it
the ROOM for a membership body: the secretary-general, the member-services staff and an advisory teammate that answers from the association's own guides; a question it cannot settle becomes a task for a named staffer with a due date and the member sees the follow-up; the board decides on any new service as a proposal; the receipt records question, answer and handler
The engagement on top
connectors to the member register (CRM or Excel) and the document library, WhatsApp as the member door, an association professional on the match; member-enquiry task templates
Literature review of AI in NGOs: adoption is uneven and biased towards larger organisations; use cases fall in engagement, creativity, decision-making, prediction, management and optimisation
Where
multi, global
Segment
all (NGOs)
Problem
small NGOs lack funds, skills and infrastructure
Technology
various AI uses across 65 studies
Who uses it
NGOs globally (review of 65 studies)
Outcome
65 studies reviewed (2020-2025); finding: adoption among NGOs is uneven and biased towards larger organisations (the authors, a literature review)
Effort
low
Risks
Uneven adoption favouring large NGOs; beneficiary data; unreviewed preprint.
How to adopt
Start with low-risk admin and engagement uses; build staff skills; pool tools across NGOs.
How a room runs it
the room fits the small NGO that cannot hire a data team: the director, the finance officer, the field coordinators and a teammate for the work the review lists as management and engagement (donor updates, volunteer rosters, report drafts); every task has an owner and a date, every spend above the envelope is the director's decision; the budget degrades before it blocks
The engagement on top
connectors to whatever the NGO runs (accounting, a spreadsheet, a donor CRM), a nonprofit professional on the match, grant-report and volunteer templates
Zakat and donation organisation launches an app for recurring giving by automatic debit, volunteer sign-up, program visits and public reports; digital channels reached 42% of donations in 2023
Where
ID, Southeast Asia
Segment
mid (Indonesia)
Problem
donor acquisition among younger givers; transparency
Technology
mobile app, recurring e-wallet debit, digital reporting
Who uses it
Dompet Dhuafa (Indonesia)
Outcome
Digital channels 42 % of donations in 2023 (CNN Indonesia report)
Effort
medium
Risks
Not AI; donor data privacy; trust and religious compliance.
How to adopt
Build or buy a giving app with recurring debit, volunteer sign-up and public reporting.
How a room runs it
the ROOM for a foundation's back office: the programme manager, the finance officer, the volunteer coordinator and a fundraising teammate; the agent reads the donor list and programme plan the foundation holds and GENERATES the month's tasks with owners and dates (thank and report to the recurring donors, confirm the visit roster, file the grant report), and each grant-spending proposal goes into the board member's Inbox as a decision with a deadline; the humans keep the beneficiaries' care and the sign-off; the receipt records every donation use and who approved it
The engagement on top
WhatsApp for donors and volunteers (the channel is built), connectors to the foundation's donor database and accounting, a nonprofit finance professional on the match, a reporting task template
Workshops that show nonprofits how to use AI to scale, streamline operations and improve service delivery; one partner updates a database of government schemes in local languages in 24-48 hours instead of years
Where
IN, South Asia
Segment
all (200+ nonprofits in four cities)
Problem
nonprofits serve thousands of beneficiaries with limited staff and fast-changing programme information
Technology
AI tools from OpenAI used in workshops; partners Karya and Wadhwani AI
Who uses it
OpenAI India; Haqdarshak; Wadhwani AI; Karya (India)
Outcome
Haqdarshak: a 7,000-scheme database in 14 languages took 4.5 years (2016-2020) to build; AI now updates all local versions in 24-48 hours (co-founder's statement, reported by Storyboard18)
Effort
low
Risks
Wrong information reaching beneficiaries; language quality; sensitive beneficiary data; vendor-led workshops.
How to adopt
Pick one repetitive information task (updating programme content, reports); train staff on the tool with their own material; keep a person approving anything beneficiaries see; measure update time before and after.
How a room runs it
the ROOM for a nonprofit engagement: the director, the programme officers, a volunteer coordinator and an operations teammate (an agent definition with the organisation's programme rules); the agent keeps programme information current and GENERATES the work as steps with owners and due dates (update the scheme list in four languages by Friday, draft the donor report), and puts each beneficiary-facing change into the director's Inbox as a proposal with a deadline; the humans keep beneficiary decisions and funder commitments; the receipt records sources and who approved each publication
The engagement on top
connectors to the programme database and the donor system, loaded on this pull; a paid nonprofit-operations practitioner on the match
Prototype of a multi-agent assistant that answers people seeking help with disaster relief, blood services, military family support and lifesaving training, tested for empathy and guidance
Where
US, North America
Segment
large (national humanitarian organisation)
Problem
people in stressful moments need accurate help across several services at any hour
Technology
Clara AI multi-agent platform built with AWS generative AI support
Who uses it
American Red Cross (US); Amazon Web Services
Outcome
none stated
Effort
high
Risks
Harm from a wrong answer to a person in distress; sensitive personal data; empathy not proven; prototype only.
How to adopt
Prototype on one service line with approved guidance; test the tone and the hand-off to a person with staff and volunteers before opening it to the public; define what the agent must never answer.
How a room runs it
the ROOM for a humanitarian-service engagement: the service lead, caseworkers, volunteers and a help-desk teammate (an agent definition with the organisation's service rules); the agent answers from approved guidance and GENERATES the follow-ups as plan steps with owners and due dates (call back the family case by 5 pm, update the shelter list), and puts every escalation and every case needing judgement into the caseworker's Inbox as a proposal with a deadline; the humans keep anyone in distress and every decision about help; the receipt records what the agent said and who took over
The engagement on top
connectors to the case system and the service directory, loaded on this pull; a paid nonprofit-operations practitioner on the match
Your own team builds it. Across the 5 adoptions here, 2 were low effort, 2 medium and 1 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Add AI questionnaire drafting to a survey tool; let researchers edit; keep method guidance.
A European bioinformatics institute and an AI lab run an open database of more than 200 million predicted protein structures for scientists to use
Where
GB, Europe
Segment
public body
Problem
Experimental structures exist for only a small share of known proteins
Technology
AlphaFold predictions curated and linked to PDB and UniProt
Who uses it
EMBL-EBI (United Kingdom) with Google DeepMind
Outcome
Over 200 million predicted structures, nearly all catalogued proteins (EMBL-EBI page)
Effort
low
Risks
Predictions are not experiments; low-confidence regions; licence terms.
How to adopt
Look up predictions for your targets, check confidence scores and compare with experimental structures before relying on them.
How a room runs it
the ROOM for a structural-biology project: the lead scientist, the lab manager and a literature teammate; the teammate GENERATES the target list as plan steps with owners and due dates (which proteins, which check against experimental structures) and files each go to the wet lab as a decision in the lead scientist's Inbox with a deadline; the humans keep the experiments and the interpretation; the receipt records each prediction used and what it was checked against
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the lab notebook and the sequence stores), and a structural biologist joins as the paid professional on the match
A pharmaceutical company built a self-service generative-AI platform on a cloud service so employees create their own chatbots and agents for non-regulated processes (information retrieval, document drafting, virtual critic)
Where
multi, global
Segment
enterprise
Problem
long drafting and retrieval tasks; months-long development cycles for small tools
Technology
Amazon Bedrock, serverless AWS services
Who uses it
Novo Nordisk, over 25,000 employees, global
Outcome
Over 25,000 employees used it; chatbots for 2,500+ use cases; about $10 per use case per month; a day of work to 30 minutes (AWS case study) the company’s own figure
Effort
high
Risks
Vendor case study; regulated processes excluded; governance of employee-built bots.
How to adopt
Stand up a self-service platform on managed cloud services with guardrails; limit to non-regulated processes.
How a room runs it
the ROOM for an R&D project team: the project lead, the scientists, the regulatory affairs person and a literature-and-documents teammate; the agent reads the documents the team already has and GENERATES the tasks as plan steps with owners and due dates (summarise the 40 new papers, draft section 3 of the protocol, list open questions for the steering meeting) and files in the lead's Inbox a decision whenever a draft would enter a regulated process (hold for sign-off); the scientists keep every experiment and every conclusion; the receipt records sources used and who signed off; budgets degrade before they block
The engagement on top
a custom engagement: the lab's own document store, ELN or LIMS and literature subscriptions wired to the room through connectors loaded on this pull; a scientific writer or regulatory professional joins as the paid professional on the match; the research pack carries review and drafting templates
Source: vendor · 2024 · vendor only; to corroborate
A survey-research platform added AI that drafts a questionnaire from a topic and goals and helps interpret results, aimed at small business owners and first-time academic researchers
Where
ID, Southeast Asia
Segment
SMB and mid
Problem
designing a valid questionnaire and reading results takes expertise that small teams lack
Technology
generative AI questionnaire drafting and result summarisation in the Poplite self-service platform
Who uses it
Populix, Indonesia
Outcome
none stated
Effort
low
Risks
Bad question design; bias; small-scale evidence.
How to adopt
Add AI questionnaire drafting to a survey tool; let researchers edit; keep method guidance.
How a room runs it
the ROOM for a research project: the owner or lead researcher, a fieldwork coordinator and a research teammate (an agent definition with questionnaire and sampling skills); the agent GENERATES the project plan as tasks with owners and due dates (draft questionnaire by Monday, pilot with 20 respondents, clean the data, write the findings page) and files in the lead's Inbox a decision on each design choice that changes validity (sample size, question wording) with the fieldwork start as deadline; the humans keep the research question, the interpretation and the sign-off; the receipt records what was asked, of whom and what was concluded
The engagement on top
a custom engagement: the client's own survey tool, panel export and spreadsheets wired to the room through connectors loaded on this pull; a research methodologist joins as the paid professional on the match; the research pack carries questionnaire, pilot and findings templates
A non-profit launched web and API agents for literature search, deep review and prior-art checks, benchmarked against frontier models and researchers
Where
US, North America
Segment
enterprise
Problem
Scientists cannot read the volume of papers and trials published
Technology
FutureHouse Platform agents Crow, Falcon, Owl and the experimental Phoenix
Who uses it
FutureHouse, United States (non-profit)
Outcome
Claims better retrieval precision than PhD-level researchers on LitQA, citing an external arXiv paper (FutureHouse's measure) the company’s own figure
Effort
low
Risks
Vendor benchmark; experimental agents may err; scope limited to literature.
How to adopt
Use the agents for first-pass literature search through the web or API, and have a researcher check every cited source.
How a room runs it
the ROOM for a literature review: the principal investigator, a postdoc and a literature teammate; the teammate GENERATES the review plan as steps with owners and due dates (question, sources, who checks the claims) and files the conclusions as a decision in the investigator's Inbox with a deadline; the humans keep the reading of key papers and the claims made; the receipt records each source behind each statement
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the reference manager and the lab notebook), and a domain scientist joins as the paid professional on the match
Start with a narrow field that has code templates, sandbox every run, keep a human reading of every result, and treat the automated reviewer as a filter only.
How a room runs it
the ROOM for a research programme: the principal investigator, the lab manager and a research teammate; the teammate GENERATES the experiment queue as plan steps with owners and due dates (idea, run, check) and files each paper for release as a decision in the investigator's Inbox with a deadline; the humans keep the choice of question and the verification of results; the receipt records each run, its code and its reviewer
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the compute, code and reference stores), and a research engineer joins as the paid professional on the match
A global research group tested GPT-4 synthetic respondents against about 5,000 real survey respondents and concluded they are not yet good enough to replace human sample, expecting blended human-plus-synthetic designs
Where
multi, global
Segment
enterprise
Problem
survey cost and time; whether machine-generated respondents are valid
Technology
GPT-4 conditioned on respondent demographics; comparison against real survey data
Who uses it
Kantar (Chief AI Scientist Ashok Kalidas), global
Outcome
Synthetic respondents judged not yet good enough vs about 5,000 real respondents (Kantar's own test) the company’s own figure
Test synthetic samples against real survey results before use; blend with human samples.
How a room runs it
the ROOM for a methodology review: the head of research, the analyst, the client lead and a method teammate; the agent GENERATES the validation tasks as plan steps with owners and due dates (run the synthetic arm next to the human arm, compare sub-groups, write the caveat note) and files in the head's Inbox a decision (use synthetic sample for this study or not) with the client proposal date as deadline; the humans keep the methodological judgment and what is told to the client; the receipt records the comparison and the choice, so the claim can be audited
The engagement on top
a custom engagement: the firm's own survey data, analysis scripts and panel tools wired to the room through connectors loaded on this pull; a statistician joins as the paid professional on the match; the research pack carries validation-protocol templates
A national research agency is building its AI ecosystem (people, infrastructure, data, governance), reports about 27 BRIN-built AI products and plans AI regulation for its researchers
Where
ID, Southeast Asia
Segment
public
Problem
AI literacy, reskilling and governance across a public research body
Technology
BRIN-built AI products (e-voting, data management, fire and vessel monitoring, education); AI governance framework
Who uses it
BRIN (head Arif Satria), Indonesia
Outcome
About 27 BRIN-built AI products (agency's own count)
Effort
high
Risks
No tool use named; governance not yet in place; talent.
How to adopt
Build people, infrastructure, data and governance in parallel; publish researcher AI regulation.
How a room runs it
the ROOM for a research programme: the programme head, the research group leads, the data steward and a programme teammate; the agent GENERATES the programme's recurring work as tasks with owners and due dates (collect each group's agenda, track the AI-use rules, prepare the quarterly review) and files in the head's Inbox the decisions that need a ruling (permit this tool on this data) with the review date as deadline; the researchers keep the research and its integrity; the receipt records every ruling and who made it
The engagement on top
a custom engagement: the institute's own project registry, grant system and shared drive wired to the room through connectors loaded on this pull; a research-administration or data-governance professional joins as the paid professional on the match; the research pack carries agenda, ruling and review templates
Your own team builds it. Across the 7 adoptions here, 3 were low effort, 1 medium and 3 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Feed sales data into a planning system; pilot visit planning and item recommendations with a few field reps.
An Indonesian conversational-platform company uses an LLM sales agent on its own SDR team to screen inbound leads from WhatsApp, web forms and social, follow up automatically and hand over to a rep with the history
Where
ID, Southeast Asia
Segment
SMB and mid
Problem
leads arrive on several channels at once; early-funnel follow-up is slow and uneven
Technology
Qiscus AgentLabs (LLM AI agent), WhatsApp and web channels
Who uses it
Qiscus (own SDR team), Indonesia
Outcome
none stated
Effort
low
Risks
Vendor's own team only; hallucinated promises; lead-data privacy.
How to adopt
Connect an LLM agent to inbound channels; screen leads, follow up and hand over with history.
How a room runs it
the ROOM for a small sales desk: the owner, the salesperson and a lead-desk teammate; the agent reads the inbound conversations the business already has and GENERATES the follow-ups as tasks with owners and due dates (reply to the quote request today, call the lead who asked for a price), and files a decision in the owner's Inbox when a lead asks for terms above the envelope (give the discount or not) with the day the quote lapses as deadline; the salesperson keeps the negotiation and the close; the receipt records every handover with its conversation
The engagement on top
a custom engagement: the company's own WhatsApp Business number, web forms and CRM wired to the room through connectors loaded on this pull; a sales trainer joins as the paid professional on the match; the sales pack carries lead-qualification and follow-up templates
Source: vendor · 2026 · vendor only; to corroborate
An insurer deployed a remote-first sales distribution platform that captures sales activity automatically and gives managers visibility into performance
Where
ID, Southeast Asia
Segment
enterprise
Problem
agents' manual daily system updates; managers' blind spots across a distributed force
Technology
Vymo Sales Excellence Cloud (automation and AI, features unspecified)
Who uses it
AIA Indonesia (PT AIA Financial), over 7,000 salespeople and managers; Vymo, Indonesia
Outcome
none stated
Effort
medium
Risks
AI features not specified; staff surveillance concerns; vendor lock-in.
How to adopt
Deploy a sales platform that logs activity automatically; train managers on dashboards.
How a room runs it
the ROOM for a branch: the branch manager, the unit leaders, the agents and a sales-desk teammate; the agent reads the activity the company's system already holds and GENERATES each agent's follow-ups as tasks with owners and due dates (call the lead from last week's quote, renew the policy due on the 20th) and files in the manager's Inbox the cases that need a decision (exception on a proposal, a transfer of a lead) with the month-end as deadline; the agents keep the conversation and the sale; the receipt records the follow-ups done and the exceptions approved
The engagement on top
a custom engagement: the insurer's own sales and policy systems and WhatsApp wired to the room through connectors loaded on this pull; a licensed insurance professional joins as the paid professional on the match; the sales pack carries follow-up and renewal templates
A technology company's sales team used a revenue-intelligence tool to score calls, build a call library for onboarding and give sellers time back
Where
US, North America
Segment
enterprise
Problem
Managers cannot review enough calls; new sellers ramp slowly
Technology
Gong Scorecards, Smart Tracker and Call Library
Who uses it
Google sales team, United States
Outcome
Sellers get back about 12 % of their time a year (estimate by Google's Director of Product Strategy; method not given) the company’s own figure
Effort
medium
Risks
Vendor case study; recorded-call consent; time estimate not verified.
How to adopt
Pilot call scoring with managers' ratings as the baseline, train the model on effective calls, then use the library in onboarding.
How a room runs it
the ROOM for a sales team: the sales manager, the sellers and a coaching teammate; the teammate GENERATES the week's coaching tasks for each seller (calls to review, tracker misses) as plan steps with owners and due dates and files the changes to the scorecard as decisions in the sales director's Inbox with a deadline; the humans keep the selling and the coaching conversation; the receipt records each reviewed call and the feedback given
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the CRM and the call recordings), and a sales-operations professional joins as the paid professional on the match
Source: vendor · 2023 · vendor only; to corroborate
An AI sales-management system for FMCG manufacturers and wholesalers plans customer visit schedules, recommends items and cross-sells to field sales staff selling to warung
Where
ID;JP, Southeast Asia;East Asia
Segment
SMB and mid
Problem
hundreds of small outlets to visit, incomplete customer data, slow order processing and sales aggregation
Technology
SENRI AI order management (visit planning, item recommendation, cross-sell, demand forecasting)
Who uses it
PT Mustika Ratu (IDX-listed cosmetics and herbal maker, demonstration partner) and other consumer-goods manufacturers; SENRI Ltd. (Japan) with PT SENRI Teknologi Indonesia, Indonesia
Outcome
97 % forecast accuracy, +10 % visits (supplier's self-reported figures; +35 % at a Japanese client in six months)
Effort
medium
Risks
Supplier-only figures; small stores' data quality; sales staff acceptance.
How to adopt
Feed sales data into a planning system; pilot visit planning and item recommendations with a few field reps.
How a room runs it
the ROOM for a regional sales team: the sales manager, the area supervisors, the salespeople and a sales teammate; the agent reads the order and visit records the company already has and GENERATES each salesperson's week as plan steps (outlets to visit, items to offer, an outlet overdue) with owners and due dates, and files in the manager's Inbox a decision to approve a discount or credit above the envelope with the delivery date as deadline; the salespeople keep the visit and the relationship; the receipt records visits done, orders taken and every credit approved
The engagement on top
a custom engagement: the company's own order system, field app or spreadsheet and WhatsApp ordering wired to the room through connectors loaded on this pull; a sales operations consultant joins as the paid professional on the match; the sales pack carries visit-plan and order-follow-up templates
Your own team builds it. Across the 4 adoptions here, 1 were low effort, 3 medium and 0 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Roll out a coding assistant to a pilot group; measure acceptance; set code-review and IP rules.
AI coding tools used across about 3,000 technology staff, reported in the 2025 results as more than 20 % more code changes
Where
DE, Europe
Segment
enterprise (e-commerce technology teams)
Problem
a large engineering organisation needs more output per engineer
Technology
AI coding tools (the tool is not named in the source)
Who uses it
Zalando (Germany)
Outcome
More than 20 % more code changes within the technology teams (Zalando, reported by Retail Gazette; output count, not a controlled productivity measure)
Effort
low
Risks
Code-change counts can rise without value; quality regressions; the tool and method are not disclosed.
How to adopt
Roll the assistant out to all engineers with review and security rules; track change counts and quality together (defects, rollbacks), not output alone.
How a room runs it
the ROOM for an engineering engagement: the engineering director, team leads, developers and a delivery teammate (an agent definition with the team's standards); the agent proposes changes and GENERATES the work as steps with owners and due dates (clear the dependency upgrades by Friday, add tests for the checkout path), and puts each production change into the team lead's Inbox as a proposal with a deadline; the humans keep the merge and the release; the receipt records what the agent proposed and who approved it
The engagement on top
connectors to the source host and the ticket system, loaded on this pull; a paid engineering practitioner on the match
AI pair-programming assistant rolled out across a global services firm after a randomised controlled trial
Where
IE, global
Segment
enterprise
Problem
slow code delivery and developer toil
Technology
GitHub Copilot
Who uses it
Accenture (global professional services), with GitHub
Outcome
Randomised trial: 84% increase in successful builds and 15% higher pull-request merge rate; developers accepted about 30% of suggestions and kept 88% of the generated characters (GitHub, 2024) the company’s own figure
Effort
medium
Risks
Insecure or licensed code suggestions; over-trust of generated code; vendor-reported metrics.
Next: in Dec 2025 Accenture and Anthropic announced about 30,000 Accenture professionals to be trained on Claude and Claude Code for tens of thousands of its developers source
How to adopt
Run a randomised pilot with telemetry; keep human review on every merge; measure build success, not lines written.
How a room runs it
the engineering room: the tech lead, the developers, the QA owner and a coding teammate; the agent drafts changes and review notes and GENERATES the follow-up tasks per developer with due dates, and each risky change becomes a proposal in the lead's Inbox with a deadline; the humans keep the review and the merge; the receipt records the suggestion, the reviewer and the build result
The engagement on top
connectors to the repository and CI, a tech-lead match, a review-and-merge task template set
AI agent answering customer support questions on the vendor's own help site, with escalation to support engineers
Where
US, North America
Segment
enterprise
Problem
millions of support requests a year and slow answers
Technology
Agentforce (Salesforce AI agents)
Who uses it
Salesforce (USA), its own Help site as customer zero
Outcome
About 76% of customer queries resolved without a human; 1.7 million+ conversations handled; 5% of inquiries escalated; 94% of global case volume covered by language support (Salesforce, Dec 2024) the company’s own figure
Effort
medium
Risks
Wrong answers at scale; customer frustration when escalation is slow; vendor-reported figures.
Next: by Sep 2025 Salesforce said about half of service interactions were with agents, support roles were no longer being backfilled and hundreds of staff were redeployed to professional services, sales and customer success source
How to adopt
Ground the agent in the knowledge base, route low-confidence cases to humans, track escalation and satisfaction.
How a room runs it
the support room: the support lead, the support engineers, the product owner and a support teammate; the agent answers from the knowledge base and GENERATES the escalation tasks per engineer with due dates, and each refund or exception becomes a proposal in the lead's Inbox with a deadline; the humans keep the hard cases; the receipt records the answer, the sources and the hand-off
The engagement on top
connectors to the help centre and ticketing, a support lead on the match, an escalation task template set
Generative AI services offered to Indonesian businesses with Microsoft, aligned to ISO 27001 and with data-loss-prevention work
Where
ID, Southeast Asia
Segment
enterprise (IT services)
Problem
secure, agile adoption of generative AI
Technology
Microsoft generative AI, data loss prevention
Who uses it
Telkomsigma (Telkom subsidiary, Indonesia)
Outcome
none stated
Effort
medium
Risks
Client-facing offer, not internal adoption; data leakage.
How to adopt
Offer generative AI with data-loss-prevention and ISO 27001 alignment to clients.
How a room runs it
the IT-services room is the delivery team's and the client's together: the account lead, the engineers, the client's security officer and a delivery teammate; the agent plans the data-loss checks as tasks with owners and dates and files any exception as a decision in the client's Inbox; the receipt is the evidence the client's auditor reads
The engagement on top
connectors to the client's ticketing and cloud tenant, a security professional on the match, security-review task templates
Meta Business Agent on WhatsApp answers customers outside business hours; a digital creative agency saw sales conversion up 65.4%
Where
ID, Southeast Asia
Segment
SMB (digital agency)
Problem
after-hours customer questions
Technology
Meta Business Agent on WhatsApp
Who uses it
Mastung (Indonesian digital creative agency); WhatsApp Indonesia
Outcome
Sales conversion +65.4 % at a digital agency (Meta's figure, no method stated) the company’s own figure
Effort
low
Risks
Unexplained figure; wrong promises to customers.
How to adopt
Enable a WhatsApp business agent with FAQs and catalogue; review answers outside business hours.
How a room runs it
the small agency's room: the owner, the account manager and an enquiry teammate on the WhatsApp number the customers already message; the agent answers from the agency's price list and turns a serious enquiry into a task with an owner and a decision for the owner (send the quote by Friday); the humans keep price and promise; the receipt records every quote sent
The engagement on top
WhatsApp is the door and is built; connectors to the agency's price list, project tracker and invoicing; a sales professional on the match
Source: vendor · 2026 · vendor only; to corroborate
Claude and Claude Code made available to up to 350,000 associates for engineering, delivery and corporate functions
Where
US, North America
Segment
enterprise
Problem
engineering productivity and legacy code modernisation at IT-services scale
Technology
Claude Enterprise, Claude Code, MCP and the Agent SDK, with Cognizant's own multi-agent platform
Who uses it
Cognizant (US-headquartered IT services), with Anthropic
Outcome
Claude access for up to 350,000 associates across corporate functions, engineering and delivery (company figure, Nov 2025) the company’s own figure
Effort
high
Risks
Code quality and security regressions; client data leakage; over-reliance on one model vendor.
Next: initial focus is software engineering productivity, legacy modernisation and agent building, with industry solutions starting in financial services, taken to clients through workshops and reference patterns source
How to adopt
Deploy to engineering first; add reference patterns per client industry; keep human review on all shipped code.
How a room runs it
the delivery room: the delivery lead, the engineers, the client's technical owner and a coding teammate; the agent drafts code, tests and documentation and GENERATES the migration tasks per engineer with due dates, and each risky change becomes a proposal in the lead's Inbox with a deadline; the humans keep the review and the release; the receipt records the change, the reviewer and the test result
The engagement on top
connectors to the client's repository, ticketing and CI, a delivery lead on the match, a legacy-migration task template set
AI coding assistant for engineers: nearly 1,000 adopted in three months, full rollout planned; engineers report saving over seven hours a week and 30% suggestion acceptance
Where
ID, Southeast Asia
Segment
enterprise
Problem
developer productivity and code quality
Technology
GitHub Copilot, via Microsoft Indonesia and partner Insight
Who uses it
GoTo Group (Indonesia)
Outcome
Engineers report saving over 7 hours a week, 30 % suggestion acceptance (GoTo's self-reported figures) the company’s own figure
Effort
low
Risks
Self-reported savings; code licensing and IP leakage; security flaws.
How to adopt
Roll out a coding assistant to a pilot group; measure acceptance; set code-review and IP rules.
How a room runs it
the ROOM for a software house: the delivery manager, the tech lead, the developers, the client's product owner and a delivery teammate; the agent reads the tickets and repos the house already uses and GENERATES the sprint's tasks with owners and dates (triage the bug list, write the test plan, prepare the release notes), and each release or scope change lands in the client owner's Inbox as a decision with a deadline; the developers keep the code and the merge; the receipt records what shipped and who approved
The engagement on top
connectors to the house's tracker and repository, a senior engineer on the match, sprint and release task templates; the coding assistant stays the developers' own tool
Source: vendor · 2024 · vendor only; to corroborate
Broad rollout of an AI coding assistant to developers across the software lifecycle: code generation, migration and optimisation, documentation and test automation
Where
IN, South Asia
Segment
enterprise (consulting and engineering services)
Problem
engineering teams need faster delivery across the whole lifecycle
Technology
GitHub Copilot
Who uses it
Synechron (India developers; US-headquartered)
Outcome
Productivity gains of up to 40 % across the software lifecycle in internal experiments run by Synechron Labs, with developer surveys (Synechron's own figures) the company’s own figure
Effort
low
Risks
Self-reported gain; insecure or licensed code in suggestions; over-reliance.
How to adopt
Licence the assistant for a pilot group, set code-review and security rules, run a measured internal experiment against a baseline, then widen.
How a room runs it
the ROOM for an engineering engagement: the engineering lead, the developers, the QA lead and a delivery teammate (an agent definition with the team's standards); the agent reads the repositories and tickets it is allowed to see and GENERATES the work plan as steps with owners and due dates (migrate module X by sprint end, write the missing tests), and puts each change that touches production into the lead's Inbox as a proposal with a deadline; the humans keep the merge and the release; the receipt records what the agent proposed and who approved it
The engagement on top
connectors to the source host and the ticket system, loaded on this pull; a paid engineering practitioner on the match
Source: vendor · 2024 · vendor only; to corroborate
Survey of Indonesian software development leaders (75 of 2,000+ in Asia Pacific) on AI agents: 80% expect agents to speed up development, 71% of AI-using developers use low-code/no-code, 83% say infrastructure needs upgrading, 57% not confident in skills
Where
ID, Southeast Asia
Segment
mid and enterprise (Indonesia survey)
Problem
data quality, testing and skills gaps
Technology
AI agents, low-code/no-code
Who uses it
75 Indonesian software development leaders (Salesforce State of IT)
Outcome
80 % expect agents to speed up development; 57 % not confident in skills (Salesforce survey, n=75) the company’s own figure
Effort
medium
Risks
Small vendor survey; skills gap.
How to adopt
Upgrade infrastructure, train developers, then trial agents on low-risk tasks.
How a room runs it
for the house whose developers want agents but fear the gaps: the room puts the agent at the edge where it is safe, on the work around the code (test plans, release notes, status to the client), with each task owned and dated and each deploy a decision a human accepts; the budget caps token spend per sprint and degrades first
The engagement on top
connectors to the house's repo and tracker, a QA or security professional on the match, release and testing templates
Source: vendor · 2025 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 9 adoptions here, 4 were low effort, 4 medium and 1 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Connect a chatbot to the booking system on the website and WhatsApp; start with periodic-service booking; hand off to staff.
Dealer AI voice agents book service, run recall outreach and pitch add-ons; a consumer investigation finds them spreading across US dealership networks
Where
US, North America
Segment
SMB and mid
Problem
unanswered service lines and recall follow-up; side effect: upsell pressure on customers
Technology
voice agents, call transcription, handoff to a human advisor
Who uses it
dealerships (US) using AutoAlert, DealerIQ, Spyne tools; reported via Consumer Reports
Outcome
none stated
Effort
medium
Risks
Aggressive upselling; consumer-protection and disclosure rules; poor handling of complex calls.
How to adopt
Pilot a voice agent for booking and recall outreach; set rules on upselling; disclose to callers that it is AI.
How a room runs it
the room as the neutral side of the same call: the teammate books and follows up recalls, but every pitch beyond the manufacturer schedule is a proposal with its reason shown to the service manager before any customer sees it; the customer can ask for a person and the handoff carries the transcript; budgets degrade before they block; the receipt shows what was offered and who approved it
The engagement on top
a custom engagement: the dealer's service platform and call recordings wired through connectors, a workshop-operations professional to set the offer rules, the workshop pack
AI agent front end for used-car selling: instant reply, 30-minute inspection with an AI inspection app, payment within 24 hours; routes to a person when needed
Where
SG;MY;HK, Southeast Asia;East Asia
Segment
enterprise
Problem
slow, opaque used-car sale; lost leads
Technology
agentic AI chat, AI inspection app, ML valuation
Who uses it
Carro (online used-car marketplace), Singapore, Malaysia, Hong Kong
Put an AI agent in front of seller enquiries with a fast inspection and payment flow; route to a person when needed.
How a room runs it
the room for a used-car dealer: owner, sales lead, inspector and a trade-in teammate; the teammate replies to the lead, GENERATES the inspection visit and the paperwork steps with owners and dates, and files the price offer as a decision for the owner with a deadline; the humans keep the inspection, the offer and the handover; the receipt records each step a seller can see
The engagement on top
a custom engagement: the dealer's stock list and listing portal wired in, a used-car valuation professional on the match, the dealer pack's templates
Source: vendor · 2026 · vendor only; to corroborate
Automotive joint venture buys chatbot and 24/7 voice customer service with human handoff for website, app and messaging enquiries; contract renewed
Where
CN, East Asia
Segment
enterprise
Problem
high volume of owner and prospect enquiries to answer around the clock
Technology
conversational chatbot (iBot Pro), smart voice customer service with call routing and human handoff
Who uses it
an unnamed Sino-European automotive joint venture (China) and Xiao-I (Shanghai)
Outcome
none stated
Effort
medium
Risks
Client unnamed and no figures, so the benefit is unproven; poor handoffs frustrate callers; recorded-call consent.
How to adopt
List the enquiry types by channel, start the chatbot on the top few and the voice line after-hours, route to a person on low confidence, and review the transcripts weekly.
How a room runs it
the ROOM: the customer-care manager, two agents and a service-quality teammate (an agent definition with the automotive skills); the agent reads the enquiry and call logs and GENERATES the week's tasks for the care team (which intents to add, which handoffs to review) as plan steps with owners and due dates, and an escalation rule change lands in the care manager's Inbox as a proposal with a deadline; what the humans keep: the hard cases, refunds and complaints, and the sign-off on scripts; the receipt records every finished task and each decision with who decided and when
The engagement on top
a custom engagement: the CRM, the call platform and the messaging channels stay the client's own — the agent is wired to them through connectors loaded on this pull, a customer-care professional joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates; the client receives the reports and gives direction in the room
Source: vendor · 2025 · vendor only; to corroborate
Carmaker puts a ChatGPT-based assistant into the in-car voice assistant (IDA) by cloud update to cars already on the road and new models; unanswered questions go anonymously to the model
Where
DE, Europe
Segment
enterprise
Problem
drivers cannot get open-ended answers from fixed voice commands
Technology
large language model chatbot (Cerence Chat Pro), hybrid on-board/cloud voice assistant, over-the-air update
Who uses it
Volkswagen Group (Volkswagen, Cupra, Seat, Skoda) with Cerence, Germany; European lineup
Outcome
none stated
Effort
high
Risks
Wrong or invented answers read aloud while driving; data-protection questions on voice data; dependence on one supplier and one model provider.
How to adopt
Pick a voice-assistant supplier with a language-model add-on, keep vehicle commands on the head unit and send only unanswered questions to the cloud, ship by over-the-air update, and start in the languages with the most owners.
How a room runs it
the ROOM: the product owner, the customer-experience lead and a voice-quality teammate (an agent definition with the automotive skills); the agent reads the anonymised unanswered-question log and GENERATES the fortnight's tasks for the voice team (which question classes to cover, which answers to review) as plan steps with owners and due dates, and the owner's rollout-by-language decision lands in the product lead's Inbox as a proposal with a deadline; what the humans keep: the answer policy, the legal sign-off and every change to what the assistant may say; the receipt records every finished task and each decision with who decided and when
The engagement on top
a custom engagement: the question log, the telemetry store and the release pipeline stay the client's own — the agent is wired to them through connectors loaded on this pull, a voice-UX and compliance professional joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates; the client receives the reports and gives direction in the room
Voice AI answers dealership service calls around the clock and writes appointments straight into the service-department platform (myKaarma) as a service advisor would
Where
US, North America
Segment
SMB and mid
Problem
missed service calls, hold times and advisors tied to the phone
Technology
conversational voice AI, read-write integration with a dealer service platform
Who uses it
STELLA Automotive AI and myKaarma, US dealerships
Outcome
Service booking reduced to under two minutes (vendor's announcement; not a measured result)
Connect a voice AI to the dealership's service-scheduling platform; test after-hours calls first; review transcripts.
How a room runs it
the same workshop room, the call as the entry: the teammate reads the shop's own appointment book, books the job and GENERATES the advisor's callback list and the next morning's job cards as plan steps; a call that needs judgement becomes a proposal in the service manager's Inbox, not a guess; the humans keep the diagnosis, the price and the customer relationship; the receipt keeps the call, the booking and who changed it
The engagement on top
a custom engagement: their scheduling or DMS system wired to the room through connectors loaded on this pull, telephony or WhatsApp as the channel, an experienced service advisor on the match, the workshop pack
Dealer-group chatbot (TASIA) books periodic service at the workshop or Toyota Home Service (technician comes to the customer) over the website and WhatsApp, 24 hours
Where
ID, Southeast Asia
Segment
enterprise
Problem
after-sales booking handled by phone and counter; customers lost outside opening hours
Technology
chatbot on website and WhatsApp, service-booking back end, home-service dispatch
Who uses it
Auto2000 (Toyota dealer network, Astra group), Indonesia
Outcome
none stated
Effort
medium
Risks
Booking errors; WhatsApp channel dependency; no usage evidence.
How to adopt
Connect a chatbot to the booking system on the website and WhatsApp; start with periodic-service booking; hand off to staff.
How a room runs it
the ROOM for one workshop: the service manager, the service advisors, the foreman and a booking teammate (an agent definition with the workshop's service skills); the teammate takes the WhatsApp request, drafts the booking and GENERATES the day's job cards for the foreman (bay, mechanic, parts to pull) as plan steps with owners and due dates; a decision lands in the manager's Inbox when a slot or a part clashes (move the job / order the part, deadline end of day); the humans keep the repair, the quote sign-off and any spend above the envelope; the receipt records each job from request to hand-over
The engagement on top
a custom engagement: the workshop's own dealer management system and parts stock wired in through connectors loaded on this pull, WhatsApp as the door (built), a service-advisor professional on the match, the workshop pack's job-card and reminder templates
Your own team builds it. Across the 6 adoptions here, 0 were low effort, 5 medium and 1 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Embed a skin-diagnosis and virtual try-on tool on the brand's site; connect to the product catalogue.
How a room runs it
the brand's own counter room: the shop manager, the beauty advisors and an advisor teammate; the teammate GENERATES the follow-up for each diagnosis (sample, visit, reorder) as tasks for the advisor and files a stock or promotion proposal for the manager; the humans keep the advice at the counter; the receipt records what was recommended
The engagement on top
a custom engagement: the brand's e-commerce and loyalty systems wired in, a beauty advisor on the match
Source: vendor · 2020 · vendor only; to corroborate
Dermatology brand runs an AI skin advisor inside WhatsApp: QR code starts a chat, the analyser runs in WhatsApp's browser, results and product recommendations come back in the chat
Where
ID, Southeast Asia
Segment
mid and enterprise
Problem
customers unsure which treatment or product fits skin of Fitzpatrick types III-V
Technology
WhatsApp Business chatbot, AI skin analyser (Revieve), QR entry
Who uses it
ERHA (Arya Noble), Indonesia, 112+ clinics
Outcome
none stated
Effort
low
Risks
Skin images are sensitive data; wrong recommendations; WhatsApp dependency.
How to adopt
Run a skin-analysis tool inside WhatsApp linked to the product catalogue; start with a QR code at clinics.
How a room runs it
the ROOM for a clinic or salon: the owner, the front desk, the therapist or dermatologist and a client-care teammate; the teammate carries the WhatsApp chat, hands the analysis to the therapist and GENERATES the follow-ups (book the consultation, reminder before the visit, check-in after treatment) as plan steps with owners and dates; a treatment recommendation is a proposal in the dermatologist's Inbox, never sent as advice by the agent; the humans keep the diagnosis and the treatment; the receipt records what was suggested and who signed it
The engagement on top
a custom engagement: the clinic's booking system and customer list wired through connectors loaded on this pull, WhatsApp as the door (built), a dermatologist or senior therapist on the match, the beauty pack's follow-up templates
Salon marketplace launches an AI receptionist that answers calls, books, cancels and reschedules in the salon's calendar, with transcripts and WhatsApp alerts
Switch on an AI receptionist tied to the salon calendar; test answers on common bookings; review transcripts.
How a room runs it
the salon room: owner, stylists, reception teammate; the teammate answers the call, changes the calendar and GENERATES the stylist's day list and the reminders (client, time, service) with owners and times; a request it cannot settle is a proposal to the owner by a stated time; the humans keep the chair work and the price changes; the receipt records each booking change
The engagement on top
a custom engagement: their booking tool or calendar wired through connectors, WhatsApp as the channel (built), a salon-operations professional on the match, the beauty pack
Beauty retailer adds AR virtual try-on (GLAMlab) and an AR and AI skin-analysis tool that recommends products and saves assessments over time, with QR codes in stores replacing testers
Where
US, North America
Segment
enterprise
Problem
shoppers could not use testers during the pandemic and needed guidance on products
Technology
AR virtual try-on, AI skin analysis, QR codes on shelf strips
Who uses it
Ulta Beauty, United States
Outcome
none stated
Effort
medium
Risks
Facial-image privacy; skin-tone accuracy; recommendations steering to house products; no result attributed to the tool.
How to adopt
Add a try-on and skin-analysis feature to the retailer app linked to the catalogue, put QR codes on shelves, and train store staff to steer shoppers to it.
How a room runs it
the ROOM: the digital-merchandising lead, the store-experience manager and a skin-advice teammate (an agent definition with the beauty skills); the agent reads the tool's usage and the product sell-through the retailer already has and GENERATES the month's tasks for the store teams (which shelf to promote, which advice to refresh) as plan steps with owners and due dates, and a change to the recommendation rules lands in the merchandising lead's Inbox as a proposal with a deadline; what the humans keep: product claims, the brand relationships and the in-store consultation; the receipt records every finished task and each decision with who decided and when
The engagement on top
a custom engagement: the app, the catalogue and the sales data stay the client's own — the agent is wired to them through connectors loaded on this pull, a beauty-retail professional joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates; the client receives the reports and gives direction in the room
AI innovation lab with Microsoft bringing generative AI to 20+ brands, an employee chatbot, and ChatGPT Enterprise early adoption
Where
US, North America
Segment
enterprise
Problem
slow trend response, product development and information search across many brands
Technology
Azure OpenAI Service, ChatGPT Enterprise, Azure AI voice assistant
Who uses it
The Estee Lauder Companies (USA), with Microsoft
Outcome
AI Innovation Lab announced for 20+ brands including Clinique, MAC and Too Faced; became an early ChatGPT Enterprise adopter after reviewing employee-submitted use cases (company statements, Apr 2024)
Effort
medium
Risks
Unsubstantiated product claims; brand-voice drift; consumer data handling.
Next: the employee chatbot for project, region and campaign information is being built and generative AI is being explored for research and development source
How to adopt
Collect employee use cases first, set up a joint lab, build one internal chatbot, then extend to brand teams.
How a room runs it
the brand room: the brand manager, the product developer, the regional marketer and an insights teammate; the agent collects consumer signals and drafts the brief and GENERATES the follow-up tasks per team with due dates, and each claim or campaign change becomes a proposal in the brand manager's Inbox with a deadline; the humans keep the creative and claims decisions; the receipt records the sources and the approver
The engagement on top
connectors to the brand-feedback and campaign tools, a brand lead on the match, a trend-to-brief task template set
Enable the concierge in the booking platform; configure services and payment rules; review conversations weekly.
How a room runs it
the same salon room with the concierge as the teammate: it reads the business's own services, team and availability, GENERATES the daily call-back and deposit steps for the front desk, and files any exception (a refund, a double booking) as a decision for the owner; the humans keep the service and the money decisions; the receipt records every booking change
The engagement on top
a custom engagement: the salon's booking system wired in, a beauty-business professional on the match, the pack's deposit and reminder templates
Source: vendor · 2026 · vendor only; to corroborate
Beauty retailer pilots its own app inside ChatGPT for advice, recommendations and loyalty rewards
Where
FR;US, global
Segment
enterprise
Problem
customers ask AI assistants where to find the right product
Technology
ChatGPT apps integration, Beauty Insider profile data
Who uses it
Sephora (LVMH group, France; US pilot)
Outcome
Pilot in the United States launched 24 Mar 2026; Sephora's community has more than 80 million active members (company figure); no performance figures yet the company’s own figure
Effort
medium
Risks
Skin and allergy advice errors; profile data consent; dependence on a third-party assistant channel.
Next: in-app payments and checkout for US users, with a stated goal to expand the experience globally source
How to adopt
Pilot in one market, link the loyalty profile with consent, add checkout after measuring recommendations.
How a room runs it
the beauty-advice room: the consultant, the store lead, the customer and a shopping teammate; the agent recommends from the customer's linked profile and GENERATES the follow-up tasks (samples, restock, appointment) with due dates, and each discount or reward use becomes a proposal in the lead's Inbox; the humans keep the personal advice; the receipt records the recommendation, the profile data used and the offer
The engagement on top
connectors to the catalogue, the loyalty system and the booking tool, a consultant on the match, a follow-up task template set
Beauty group builds a generative-AI beauty counsellor on its online shop that recommends products from purchase history, reviews and skin-diagnosis data; its existing skin-diagnosis tool has 2.5 million uses
Where
KR, East Asia
Segment
enterprise
Problem
online skin diagnosis converts less than in-store advice from a sales assistant
Technology
GPT-4o and GPT-4o-mini on a cloud platform, AI search, data pipeline, skin-diagnosis tool
Who uses it
Amorepacific Group (South Korea) with Microsoft
Outcome
Existing skin-diagnosis tool used 2.5 million times in four years (Amorepacific, via Microsoft) the company’s own figure
Effort
high
Risks
Skin and health claims; sensitive customer data; recommendations that favour stock over fit; unlaunched at publication.
How to adopt
Join purchase history, reviews and diagnosis data in one store, ground a language model on the catalogue, launch inside the shop app, and compare conversion against assisted in-store sales.
How a room runs it
the ROOM: the e-commerce lead, the beauty-advisor team lead and a counsellor teammate (an agent definition with the beauty skills); the agent reads the shop's chat and diagnosis logs and GENERATES the week's tasks for the advisors (which product answers to correct, which customers to follow up) as plan steps with owners and due dates, and a change to what the counsellor may recommend lands in the e-commerce lead's Inbox as a proposal with a deadline; what the humans keep: claims about skin and health, the product catalogue and the sign-off on recommendations; the receipt records every finished task and each decision with who decided and when
The engagement on top
a custom engagement: the shop, the order history and the review data stay the client's own — the agent is wired to them through connectors loaded on this pull, a beauty-retail professional joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates; the client receives the reports and gives direction in the room
Source: vendor · 2025 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 8 adoptions here, 3 were low effort, 4 medium and 1 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Let drivers report suspicious orders; add detection of tampered apps and fake location; graded penalties with appeal.
Driver-side AI: Turbo mode to optimise earnings and a Driver AI Assistant; adopters earned 23% more per online hour
Where
multi, Southeast Asia
Segment
enterprise
Problem
driver idle time and earnings transparency
Technology
AI routing and earnings optimisation, driver assistant, digital earnings tracker
Who uses it
Grab, Southeast Asia
Outcome
Adopters earned 23 % more per online hour (Grab's remarks, adopter vs non-adopter, not causal)
Effort
high
Risks
Not causal; worker autonomy and algorithmic management; Driver AI Assistant not sourced.
How to adopt
Offer drivers earnings-optimisation tools and an assistant in the driver app; measure against non-adopters.
How a room runs it
the room around a driver's week: the driver or family fleet, the household and a teammate that reads the earnings records the driver already has and GENERATES the week's tasks (service the bike on Sunday, file the tax, renew the licence) with owners and dates; a decision lands for the owner when a cost crosses the envelope; the humans keep the driving; the receipt records costs and filings
The engagement on top
a custom engagement: the fleet's bookkeeping and the platform export wired in through connectors, an accountant on the match, the gig pack
Machine-learning system predicts when food will be ready and assigns each order to the best-placed courier, using order data from three apps, weather and traffic
Where
GB, Europe
Segment
enterprise
Problem
slow pickups and idle couriers
Technology
machine learning order assignment and preparation-time prediction
Who uses it
Deliveroo (acquired by DoorDash, completed 2 Oct 2025), United Kingdom (140 cities)
Outcome
In France average delivery time fell about 20 % (32 to 26 minutes) January-June 2017 and couriers made about 20 % more deliveries (Deliveroo's figures, reported by Journal du Net; other growth factors not separated)
Effort
high
Risks
Gains confounded by other changes; worker pay and autonomy; data quality of preparation times.
How to adopt
Collect order, preparation and courier-location data, predict preparation time, assign by best placement, and compare delivery times with a control zone.
How a room runs it
the ROOM: the operations lead, the city managers and a dispatch teammate (an agent definition with the logistics skills); the agent reads the assignment and delay data the platform already holds and GENERATES the week's tasks for the city managers (which zones to add couriers, which restaurants to review) as plan steps with owners and due dates, and a change to dispatch priorities lands in the operations lead's Inbox as a proposal with a deadline; what the humans keep: courier pay rules, appeals and the sign-off on algorithm changes; the receipt records every finished task and each decision with who decided and when
The engagement on top
a custom engagement: the order platform, the courier app and the maps feed stay the client's own — the agent is wired to them through connectors loaded on this pull, a logistics-operations professional joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates; the client receives the reports and gives direction in the room
Real-time intelligent dispatch combining optimisation and machine learning assigns orders to couriers across a network of over 60 million orders a day and more than five million couriers
Where
CN, East Asia
Segment
enterprise
Problem
assigning orders to couriers in seconds at very large scale
Average delivery time down 20.96 % and courier distance per order down 23.77 % since 2019, about USD 0.23 billion saved a year (Meituan's own authors) the company’s own figure
Effort
high
Risks
Self-reported by the operator; very high data and engineering needs; pressure on couriers from time targets.
How to adopt
Model the assignment problem with both cost and courier-experience goals, run the new rule beside the old one, and expand by area once the gain is measured.
How a room runs it
the ROOM: the operations director, the regional dispatch managers and a network-planning teammate (an agent definition with the logistics skills); the agent reads the dispatch and courier logs the platform already holds and GENERATES the week's tasks for the regional managers (which areas to rebalance, which rules to test) as plan steps with owners and due dates, and a change to the assignment objective lands in the operations director's Inbox as a proposal with a deadline; what the humans keep: courier welfare rules, the objective weights and the sign-off on live changes; the receipt records every finished task and each decision with who decided and when
The engagement on top
a custom engagement: the order platform and the courier app stay the client's own — the agent is wired to them through connectors loaded on this pull, a logistics-optimisation professional joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates; the client receives the reports and gives direction in the room
Generative AI voice support for couriers (Dashers) answering calls at 2.5 seconds or less
Where
US, North America
Segment
enterprise
Problem
couriers calling support while driving need fast answers; call volume in the hundreds of thousands a day
Technology
Amazon Bedrock with Claude 3 Haiku, Amazon Connect and Lex, knowledge-base retrieval
Who uses it
DoorDash (USA), with the AWS Generative AI Innovation Center
Outcome
Answers in 2.5 seconds or less; hundreds of thousands of Dasher calls handled daily; escalations to live agents down by thousands per day; built to live testing in about 2 months; testing capacity up 50x (AWS case study) the company’s own figure
Effort
high
Risks
Wrong answers about pay or deactivation; latency while driving; guardrail failures.
Next: DoorDash plans to integrate its event-driven logistics workflow service so the system can take actions for couriers, not only answer questions source
How to adopt
Build a retrieval-grounded voice assistant, test thousands of calls per hour, escalate to a human on low confidence.
How a room runs it
the courier support room: the support lead, the dispatch owner, the courier and a support teammate; the agent answers from the knowledge base and GENERATES the escalation tasks per support engineer with due dates, and each payout exception becomes a proposal in the lead's Inbox with a deadline; the humans keep the exceptions; the receipt records the answer, the source and the hand-off
The engagement on top
connectors to the contact centre and the knowledge base, a support lead on the match, a courier-issue task template set
Marketplace AI for freelancers: Upwork's Uma shortlists talent, summarises work history and drafts contracts; Fiverr's Neo assists sellers and matching
Where
multi, global
Segment
SMB
Problem
time spent on proposals, shortlisting and contracts
Technology
AI work agent, AI matching, contract generator
Who uses it
Upwork, Fiverr, global
Outcome
none stated
Effort
low
Risks
Secondary blog source; bias in shortlists; platform lock-in.
How to adopt
Switch on a marketplace assistant for shortlisting and contract drafting; keep final hiring with the client.
How a room runs it
the freelancer's one-person room: the freelancer, their client and an admin teammate; the teammate GENERATES the proposal, the invoice and the contract steps with dates, and files the price decision for the freelancer; the humans keep the work and the sign-off; the receipt records the scope and the deliverable
The engagement on top
a custom engagement: the freelancer's calendar and invoicing wired in, a business-admin professional on the match
Source: press · 2026 · vendor only; to corroborate
Internal 'Karma' score reviews the past records and complaint history of both delivery partner and customer to settle disputes; repeated dishonesty ends a partner's access
Where
IN, South Asia
Segment
enterprise
Problem
food theft and refund fraud on both sides of a delivery
Technology
internal behaviour scoring of delivery partners and customers
Who uses it
Zomato, India
Outcome
About 5,000 gig workers terminated a month; in about 50-70 % of disputed cases the customer is refunded and the partner keeps working (CEO's statement)
Effort
medium
Risks
Wrongful removal of income; opaque scoring; self-reported figures; bias against partners.
How to adopt
Record both sides of every dispute, score repeat behaviour, set a refund-and-keep-working rule for unclear cases, and give partners an appeal.
How a room runs it
the ROOM: the trust-and-safety lead, the support managers and a disputes teammate (an agent definition with the marketplace skills); the agent reads the dispute and score data the platform holds and GENERATES the week's tasks for the support managers (which cases to re-review, which appeals to answer) as plan steps with owners and due dates, and a change to the score thresholds lands in the trust-and-safety lead's Inbox as a proposal with a deadline; what the humans keep: the final say on removing a partner, appeals and any cost the company absorbs; the receipt records every finished task and each decision with who decided and when
The engagement on top
a custom engagement: the order, refund and partner systems stay the client's own — the agent is wired to them through connectors loaded on this pull, a marketplace trust-and-safety professional joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates; the client receives the reports and gives direction in the room
Driver-protection suite: drivers report fake orders in the app and the system cancels them in about two minutes; automatic detection of illegal tools (fake GPS, modified apps) with graded penalties
Where
ID, Southeast Asia
Segment
enterprise
Problem
fictitious orders and cheating tools that cost honest drivers income
Technology
machine learning and AI fraud detection, face verification, phone masking
Who uses it
Gojek (GoTo), Indonesia
Outcome
92 % of partners feel safer (Gojek's own statement, unverified)
Effort
high
Risks
False positives penalising drivers; no independent audit; adversarial workarounds.
How to adopt
Let drivers report suspicious orders; add detection of tampered apps and fake location; graded penalties with appeal.
How a room runs it
the ROOM for a driver fleet owner or cooperative: the coordinator, the drivers and a driver-support teammate; the teammate reads the platform notices and income statements the drivers share and GENERATES the case steps (dispute a deactivation, file the evidence, appeal by the date) as tasks with owners and due dates; the appeal is a proposal in the coordinator's Inbox; the humans keep the appeal and the platform relationship; the receipt records each case
The engagement on top
a custom engagement: the fleet's own records and the group's WhatsApp wired in through connectors, a labour or transport-law professional on the match, the gig pack's templates
AI blocks fictitious orders before they reach driver accounts and detects fake-GPS use (claimed up to 98% accuracy); police action on a fake-order syndicate
Where
ID, Southeast Asia
Segment
enterprise
Problem
fake orders and fake GPS
Technology
AI anti-fraud detection, driver reminders
Who uses it
Gojek, Indonesia, with Polda Metro Jaya
Outcome
Fake-GPS detection claimed up to 98 % accuracy (Gojek's claim, unverified)
Effort
high
Risks
Unverified accuracy; wrongful sanctions; arms race with fraudsters.
How to adopt
Block fictitious orders before they reach driver accounts; detect fake GPS; coordinate with police on syndicates.
How a room runs it
the same driver-support room: the teammate turns a detection or a sanction into tasks (collect evidence, report, appeal) for the coordinator and the driver, and the report to the police is a proposal the humans sign; the receipt keeps the evidence chain
The engagement on top
as above, with the fleet's evidence folder wired in and a lawyer on the match
Ride-hail platform turns its worker network into a task platform that supplies data and AI-training work to labs and enterprises
Where
US, global
Segment
enterprise
Problem
idle time and income gaps for drivers and couriers; demand for human-checked AI data
Technology
Uber AI Solutions task platform, annotation and data-foundry tools
Who uses it
Uber (USA)
Outcome
Global digital task platform live in 30 countries; translation across 100+ languages from Uber's own operations (Uber, Jun 2025) the company’s own figure
Effort
high
Risks
Worker classification and pay transparency; data quality; client data confidentiality.
Next: Uber plans an AI-powered interface where clients describe data needs in plain language and the platform sets up, assigns and quality-checks the work source
How to adopt
Pilot with one task type, add reviewer sampling, publish pay rules, expand by country.
How a room runs it
the work-pool room: the programme owner, the quality reviewers, the workers and a coordinating teammate; the agent matches tasks to workers and GENERATES the review tasks per reviewer with due dates, and each payout or rejected batch becomes a proposal in the owner's Inbox with a deadline; the humans keep quality and pay decisions; the receipt records the task, the checker and the payment
The engagement on top
connectors to the task tool and the payment rail, a quality lead on the match, annotation and review task templates
Claude-based support for driver onboarding and rider issues, starting from onboarding and widening to charge disputes
Where
US, North America
Segment
enterprise
Problem
riders and drivers waited 30 to 40 minutes for an agent
Technology
Claude (Anthropic)
Who uses it
Lyft (USA)
Outcome
Support resolution time down more than 87% and decision-making accuracy up more than 30%; previous waits of 30 to 40 minutes; savings reinvested in upskilling agents (Lyft via Anthropic) the company’s own figure
Effort
medium
Risks
Wrong dispute outcomes; vendor-reported metrics; vulnerable riders mishandled by automation.
Next: Lyft's stated direction is to have Claude handle increasingly complex issues while human agents focus on cases needing empathy, and to expand programmes like Lyft Silver for older riders source
How to adopt
Start with one narrow flow such as driver onboarding, measure resolution time, then widen to disputes.
How a room runs it
the support room: the support lead, the support agents, the policy owner and a support teammate; the agent resolves the routine case and GENERATES the human-care tasks per agent with due dates, and each charge dispute becomes a proposal in the lead's Inbox with a deadline; the humans keep the cases needing empathy; the receipt records the answer, the policy used and the hand-off
The engagement on top
connectors to the support desk and the policy base, a support lead on the match, a dispute-handling task template set
Your own team builds it. Across the 10 adoptions here, 1 were low effort, 2 medium and 7 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Gather demand data by micro-market, forecast, and allocate professionals with a model; invest in training centres.
Marketplace matches demand and supply by micro-market, forecasts demand and allocates professionals with ML; trains them in its own centres
Where
IN, South Asia
Segment
enterprise
Problem
matching customers and professionals reliably
Technology
ML demand forecasting, micro-market matching, training centres
Who uses it
Urban Company, India and 4 international markets
Outcome
54,347 professionals as of Q1 FY26 (analyst IPO explainer, based on the company's filings)
Effort
high
Risks
Source says machine learning, not AI; gig-worker welfare; quality control.
How to adopt
Gather demand data by micro-market, forecast, and allocate professionals with a model; invest in training centres.
How a room runs it
the ROOM for a household-services business: the owner, the dispatcher, the technicians and a scheduling teammate; the teammate reads the bookings the business already takes and GENERATES the day's jobs per technician (visit window, address, tools) as plan steps with owners and dates, and files a reschedule or a refund as a decision for the owner; the humans keep the visit and the price; the receipt records each job and what the customer confirmed
The engagement on top
a custom engagement: their booking sheet or app wired through connectors, WhatsApp as the door (built), a service-operations professional on the match, the household pack
Home-cleaning, housekeeping and (from H1 2025) ironing booked through the app by verified, trained taskers; target 1,000 taskers in Indonesia in 2025
Where
ID, Southeast Asia
Segment
SMB and mid
Problem
finding trusted household help
Technology
booking app, tasker verification
Who uses it
bTaskee (Vietnam-founded), Indonesia
Outcome
none stated
Effort
medium
Risks
Matching method undescribed; company's own targets; worker classification.
How to adopt
Recruit and verify taskers, train them, and book through an app; add services stepwise.
How a room runs it
the household room: the home owner, the helper and a household teammate; the teammate GENERATES the week's chore schedule and supplies list for the helper and the owner, and files the booking or the spend as a proposal for the owner; the humans keep the work and the payment; the receipt records the visit
The engagement on top
a custom engagement: the owner's calendar and the booking app wired through connectors, WhatsApp as the door
Insurer-group digital platform books home repair (handyman, plumbing, electrical, locksmith) and cleaning on demand, routed to repair networks and a cleaning marketplace, offered through business partners
Where
DE;FR, Europe
Segment
enterprise
Problem
households struggle to find and trust urgent home repair
Technology
digital booking platform, routing to provider networks
Who uses it
Allianz Partners (with Helpling for cleaning), Germany and France
Outcome
none stated
Effort
high
Risks
Release does not describe AI; provider quality and guarantee cost; reliance on partner channels.
How to adopt
Sign up repair and cleaning networks, expose one booking flow to partner channels, confirm cleaning within 24 hours, and keep a human support line.
How a room runs it
the ROOM: the service-operations lead, the network manager and a job-routing teammate (an agent definition with the home-services skills); the agent reads the requests and provider capacity the platform holds and GENERATES the day's tasks for the network manager (which jobs to reassign, which providers to chase) as plan steps with owners and due dates, and a change to provider routing rules lands in the operations lead's Inbox as a proposal with a deadline; what the humans keep: the guarantee terms, provider vetting and the human support call; the receipt records every finished task and each decision with who decided and when
The engagement on top
a custom engagement: the booking platform and provider networks stay the client's own — the agent is wired to them through connectors loaded on this pull, a home-services professional joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates; the client receives the reports and gives direction in the room
Source: vendor · 2024 · vendor only; to corroborate
AI Helper turns a homeowner's plain-language description into a precise service request through follow-up questions, then lists qualified pros
Where
US, North America
Segment
enterprise
Problem
homeowners cannot describe the problem or the pro they need
Technology
large language models, job-assessment questions built with pros
Who uses it
Angi, United States
Outcome
Match accuracy about 30 % better than form-based requests (Angi's claim, no method given) the company’s own figure
Effort
medium
Risks
Vendor-only figure; wrong job classification; lead-sale incentives shape the matches.
How to adopt
Replace the request form with a guided description, build job-specific follow-up questions with working pros, and compare quote quality with the old form.
How a room runs it
the ROOM: the marketplace operations lead, the pro-success manager and a request-quality teammate (an agent definition with the home-services skills); the agent reads the requests and quote outcomes the platform holds and GENERATES the week's tasks for the pro-success team (which job types need better questions, which pros to onboard) as plan steps with owners and due dates, and a change to the follow-up questions lands in the operations lead's Inbox as a proposal with a deadline; what the humans keep: pro vetting, quote disputes and the sign-off on matching rules; the receipt records every finished task and each decision with who decided and when
The engagement on top
a custom engagement: the request form, the pro database and the quote flow stay the client's own — the agent is wired to them through connectors loaded on this pull, a home-services professional joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates; the client receives the reports and gives direction in the room
Source: vendor · 2026 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 4 adoptions here, 0 were low effort, 2 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: a childcare management platform already holding children, parent and invoice records. First step: configure one parent-and-staff app on that platform for messages, activity updates, invoices and session booking. Running: staff training and a parent onboarding drive.
Childcare ICT apps for attendance check-in and check-out, contact books and daily records, with an AI feature suggesting sample sentences for lesson plans and contact books, rolled out across public nurseries under a national 100 % ICT target
Where
JP, East Asia
Segment
SMB and public (nursery schools)
Problem
nursery staff shortage and paperwork load; long hours
Technology
childcare ICT apps, tablets, generative text suggestions
Who uses it
Tokyo Jichirouren Nursery Section (union of local-government employees) surveying 647 nursery staff in nine Tokyo wards, Japan
Outcome
Union survey of 647 staff: only about 10 % clearly agreed admin time was shortened, about 40 % positive including "somewhat agree"; the share spending under one hour a day on admin rose 12.2 points (staff-side survey, as reported)
Effort
medium
Risks
Devices too few and systems slow; entry can take longer than handwriting; eye and posture strain; AI text seen as unhelpful by experienced staff; the survey is union-side; the root issue is staffing, not software.
How to adopt
Fix the shared-device count before rollout, start with attendance and contact books, let staff review any AI-suggested text, and ask staff after a month whether admin time fell.
How a room runs it
the ROOM for a nursery: the head, the class leads and a records teammate; the agent teammate drafts the daily record and contact-book entries from the attendance and photos the nursery already captures; GENERATES each class lead's review of the drafts before end of day as plan steps with owners and due dates, and the head's choice to keep, change or drop a tool after a month of staff feedback as a decision block with a deadline in the Inbox of the head; what the humans keep: what is written to parents, the care itself, and staffing levels; the receipt records who approved each entry and the staff feedback that fed the decision; the client receives the reports and gives direction in the room
The engagement on top
the client's own systems stay theirs, wired to the room through connectors loaded on this pull (the nursery app, the attendance feed, the parent messaging channel); the paid professional (an early-years operations adviser) joins on the match
A childcare management suite for online enrolment, waitlists, automated subsidy submissions, family payments and QR-code sign-in and sign-out across a state-wide operator
Where
AU, Oceania
Segment
public-sector-adjacent (outside-school-hours care operator)
Problem
enrolment and payment administration across many services
Technology
childcare management SaaS, family app, payment automation, QR attendance
Who uses it
YMCA South Australia (46 services, 9,544 children, 500 staff), Australia; Xplor Education
Outcome
Enrolments typically activated within 24 hours of receipt; payments through PayNow arrive within half an hour; Mondays and Fridays no longer busier than other days (vendor's case study, no baseline) the company’s own figure
Effort
medium
Risks
Vendor-reported with no baseline; lock-in to one suite; subsidy-rule changes need ongoing upkeep.
How to adopt
Migrate families and fee schedules to the suite, switch on online enrolment and payment, and measure enrolment turnaround before and after.
How a room runs it
the ROOM for an operator office: the operations manager, the service coordinators and an enrolment teammate; the agent teammate watches the waitlist and enrolment queue the suite already holds; GENERATES each coordinator's follow-up on a stalled enrolment or an unpaid fee as plan steps with owners and due dates, and the manager's choice to open or close places at a service as a decision block with a deadline in the Inbox of the operations manager; what the humans keep: the placement decision, fee exceptions and family conversations; the receipt records each enrolment activated, each exception approved and by whom; the client receives the reports and gives direction in the room
The engagement on top
the client's own systems stay theirs, wired to the room through connectors loaded on this pull (the management suite, the payments feed, the subsidy portal export); the paid professional (an early-years operations adviser) joins on the match
Source: vendor · 2021 · vendor only; to corroborate
Government co-funds digital solutions for licensed preschools: management systems with e-forms for enrolment and subsidy, vacancy and waitlist management, video analytics for safety, and AI-assisted lesson planning and portfolio tools
Where
SG, Southeast Asia
Segment
public (all licensed operators)
Problem
administrative load, enrolment, subsidy paperwork, teacher time
Technology
preschool management systems, e-forms, waitlist tools, AI lesson-planning and observation tools
Who uses it
Early Childhood Development Agency (ECDA) under the Early Childhood Digitalisation Grant, Singapore
Outcome
none stated
Effort
medium
Risks
Child data protection; grants cover only first-year costs, leaving recurring fees; AI-drafted learning content needs teacher review.
How to adopt
Prerequisite: an operator licensed under the regulator and a shortlist of eligible digital tools. First step: choose a management or e-form system and apply for the co-funding grant. Running: staff time for enrolment and subsidy data entry and training on any AI lesson tools.
How a room runs it
the same centre room under a regulator's calendar: the principal, the admin officer, the teachers and the teammate; the agent GENERATES the tasks the grant-listed systems leave to people (enrolment forms to chase, subsidy paperwork to complete, the waitlist offer to send, the lesson-plan draft for a teacher to edit) as plan steps with owners and due dates, and a place offered to a waitlisted child is a proposal in the principal's Inbox; what the humans keep: the admission decision, the child's assessment and the lesson itself; the receipt records each offer and each form
The engagement on top
a custom engagement: connectors loaded on this pull to the centre's grant-approved management system and the regulator's centre portal where the centre is allowed; an early-childhood education consultant as the paid professional on the match
Preschool software for the Singapore market adds a parent app with QR attendance, temperature checks and automated invoices, plus real-time staff-to-child ratio tracking to meet the regulator's requirements
Where
SG, Southeast Asia
Segment
SMB and mid
Problem
regulatory ratio compliance, invoice chasing, parent updates
Technology
parent and teacher app, QR attendance, automated invoicing, ratio tracking
Who uses it
preschool and childcare centres in Singapore; illumine (vendor)
Outcome
none stated
Effort
low
Risks
Vendor-only evidence, no named adopter; ratio data errors could mask a real compliance breach; child and parent data privacy.
How to adopt
Prerequisite: staff and child rosters and a tablet or phone at the entrance. First step: switch on QR attendance and automated invoicing for one centre. Running: keep rosters current so ratio tracking is right.
How a room runs it
the owner's week inside a room: the centre owner, the teacher in charge, the part-time admin and the teammate; the agent reads attendance and the roster and GENERATES the compliance and money tasks (a ratio shortfall at 3 pm becomes a call-in task for the owner, unpaid invoices become reminders for the admin) with owners and due dates, and a late-fee waiver is a proposal for the owner with a deadline; what the humans keep: the care of the children, the waiver and the safeguarding call; the receipt records every reminder sent
The engagement on top
a custom engagement: connectors loaded on this pull to the centre's attendance and accounting systems; an early-years consultant as the paid professional on the match; WhatsApp to the parents
Source: vendor · 2026 · vendor only; to corroborate
Childcare management software with automated billing and tuition collection, a parent mobile app, online enrolment and lead tracking, used by a three-centre owner to cut back-office work
Where
US, North America
Segment
SMB (multi-site independent operator)
Problem
invoicing, enrolment tracking and parent communication workload
Technology
childcare management SaaS, mobile app, automated billing
Who uses it
Bright Beginnings Preschools (three centres, 300+ children, 75 staff, Charlottesville, Virginia), United States; Procare Solutions
Outcome
none stated
Effort
low
Risks
Vendor-authored with no figures; data-migration and training time; parent data privacy duties.
How to adopt
Load enrolment and billing data into the software, train staff, switch on the parent app and automatic payments, then track leads and capacity weekly.
How a room runs it
the ROOM for an owner-run set of centres: the owner, the directors and an admissions teammate; the agent teammate tracks the enrolment pipeline and capacity the software already shows; GENERATES each director's follow-up with a prospective family and the weekly staffing check as plan steps with owners and due dates, and the owner's choice on opening a waitlist or adding a room as a decision block with a deadline in the Inbox of the owner; what the humans keep: the tours, the family relationships and every spend decision; the receipt records each follow-up done and each capacity decision with its figures; the client receives the reports and gives direction in the room
The engagement on top
the client's own systems stay theirs, wired to the room through connectors loaded on this pull (the management software, its parent app, the billing feed); the paid professional (a child-care business consultant) joins on the match
Source: vendor · 2024 · vendor only; to corroborate
Nursery group builds one parent-and-staff app with its software platform to replace several tools: messages, activity ideas, invoices and extra-session booking across its UK settings
Parent data privacy and safeguarding; dependence on one platform vendor.
How to adopt
Prerequisite: a childcare management platform already holding children, parent and invoice records. First step: configure one parent-and-staff app on that platform for messages, activity updates, invoices and session booking. Running: staff training and a parent onboarding drive.
How a room runs it
the ROOM for one daycare or preschool: the owner or centre manager, the room leaders, the finance clerk and a centre-admin teammate (an agent definition with the early-years skills); the agent reads the attendance, sessions and fee schedule the centre already holds and GENERATES the week's tasks (invoices to send, parents to remind about extra-session fees, the staffing gap for Thursday afternoon, the daily-report drafts for the room leader to approve) as plan steps with owners and due dates, and a fee-waiver or ratio exception lands in the manager's Inbox as a proposal with a deadline; what the humans keep: the care of the children, the safeguarding judgement and every waiver; parents are reached over WhatsApp; the receipt records what was sent to whom
The engagement on top
a custom engagement: connectors loaded on this pull to the centre's own management and billing system and accounting package; an early-years consultant or centre-operations professional as the paid professional on the match; the childcare pack's attendance, fees and parent-update templates
Your own team builds it. Across the 6 adoptions here, 2 were low effort, 4 medium and 0 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: core hospital systems moved to a cloud platform. First step: use generative AI to summarise medical records for doctors and to capture inventory data. Running: IT and clinical review of summaries.
Public healthcare cluster rolling a generative-AI scribe progressively across its institutions from September 2024 (stated goal: all clinicians by end 2024): live transcript and summary of the consultation in four languages, clinician reviews and edits before upload to the electronic medical record
Where
SG, Southeast Asia
Segment
public
Problem
clinical documentation burden and doctor-patient eye contact
Technology
Note Buddy on Synapxe's Tandem platform (Azure OpenAI Service)
Who uses it
SingHealth including Singapore General Hospital; Synapxe, Singapore
Outcome
none stated
Effort
medium
Risks
Transcription errors in clinical records; language coverage; no measured savings stated.
How to adopt
Prerequisite: an electronic medical record and a cloud AI platform. First step: pilot a scribe with a few clinicians, then extend to all institutions. Running: clinician review before upload and support staff.
How a room runs it
the ROOM for a specialty centre: the clinic lead, the nurse coordinator, the registrar and the teammate; with consent captured by the clinician, the agent GENERATES the after-visit tasks from the approved note (patient reminders in the patient's language, the referral to book, the discharge paperwork for the coordinator) as plan steps with owners and due dates, and any item that changes a patient's care is a proposal in the registrar's Inbox; what the humans keep: the consent conversation, the clinical judgement and the edit of every note; the receipt records consent, review and each closed task
The engagement on top
a custom engagement: connectors loaded on this pull to the centre's EMR and appointment book; a nurse coordinator or clinical-informatics lead as the paid professional on the match; WhatsApp reminders in the patient's language over the built channel
Generative AI scribe that transcribes visits and drafts the note for the physician to review, rolled out across a regional medical group
Where
US, North America
Segment
enterprise
Problem
documentation burden and after-hours charting
Technology
ambient speech capture, generative note drafting, EHR integration
Who uses it
The Permanente Medical Group (Kaiser Permanente), Northern California, United States; Kaiser Permanente Division of Research
Outcome
7,260 physicians used it across 2,576,627 encounters; users saved nearly 16,000 hours of documentation time; 88 % of 102 surveyed physicians reported a positive effect on visit interactions (Kaiser Permanente researchers, NEJM Catalyst)
Effort
medium
Risks
Draft errors if review is skipped; patient consent and recording rules; uneven benefit across users.
How to adopt
Run a short pilot with consenting clinicians, tell patients the tool is in use, require clinician review of every draft, measure note time and after-hours charting, then widen.
How a room runs it
the ROOM for a clinic group: the clinicians, the practice manager and a documentation teammate; the agent teammate checks which notes are still unsigned and the template quality the scribe tool produces; GENERATES each clinician's note sign-off and a weekly template review for the lead as plan steps with owners and due dates, and the medical director's choice to widen the tool to another specialty as a decision block with a deadline in the Inbox of the medical director; what the humans keep: every clinical judgement and the signature on each note; the receipt records who signed each note and the review that fed the rollout decision; the client receives the reports and gives direction in the room
The engagement on top
the client's own systems stay theirs, wired to the room through connectors loaded on this pull (the EHR, the scribe tool, the clinic survey results); the paid professional (a clinical informatics adviser) joins on the match
AI-assisted inquiry and consultation with family-doctor and clinic services, where AI helps doctors handle online consultation requests and cuts the cost of serving each family-doctor customer
Where
CN, East Asia
Segment
enterprise
Problem
doctor capacity and service cost per customer
Technology
medical large language model, consultation records, AI family doctor and clinic products
Who uses it
Ping An Good Doctor (1833.HK), China
Outcome
AI helped doctors handle up to 4.0 million consultation requests a day; average family-doctor service cost per customer down about 52 % year on year; AI-assisted inquiry accuracy about 98 % (company's own figures, six months to 30 June 2025) the company’s own figure
Effort
high
Risks
Self-reported figures with undefined accuracy; clinical safety; data-protection rules for consultation records.
How to adopt
Start with low-risk inquiry types, define what counts as an accurate answer, keep a clinician on every escalation, and track cost per customer against a baseline.
How a room runs it
the ROOM for a health-service provider: the clinical lead, the doctors and a triage teammate; the agent teammate sorts incoming consultation requests the service already receives and drafts the follow-up; GENERATES each doctor's review of flagged requests and the follow-up calls as plan steps with owners and due dates, and the clinical lead's choice on which request types the agent may answer first as a decision block with a deadline in the Inbox of the clinical lead; what the humans keep: diagnosis, treatment decisions and every escalation; the receipt records each reviewed answer, who approved it and the escalations raised; the client receives the reports and gives direction in the room
The engagement on top
the client's own systems stay theirs, wired to the room through connectors loaded on this pull (the consultation platform, the records system, the call-centre log); the paid professional (a clinical safety adviser) joins on the match
Source: vendor · 2025 · vendor only; to corroborate
Two NHS hospital trusts agreed the largest secondary-care rollout of ambient voice technology: an AI scribe drafts the clinic note and letter, the clinician reviews and edits before it enters the record
Where
GB, Europe
Segment
public
Problem
documentation time after clinics; letters taking two to three days
Technology
Accurx Scribe (powered by Tandem), ambient voice capture, EPR
Who uses it
University Hospitals of Leicester and University Hospitals of Northamptonshire NHS trusts, England; Accurx
Outcome
Pilot saved eight minutes per patient and about an hour of admin a day (the trust's and supplier's own figures)
Effort
medium
Risks
Mis-transcribed notes; patient consent; savings figures are not independent.
How to adopt
Prerequisite: an electronic patient record and consent processes. First step: pilot an ambient AI scribe in some clinics, with the clinician editing every note. Running: training, governance and licences.
How a room runs it
the ROOM for a clinic team: the consultant, the clinic nurse, the medical secretary and the documentation teammate; the agent takes the clinician-approved note and GENERATES the follow-through as plan steps with owners and due dates (the letter to the GP for the secretary to send, the test the consultant ordered for the nurse to book, the patient's next appointment for scheduling), and a letter needing a consultant's wording lands in the consultant's Inbox as a proposal with a deadline; what the humans keep: the consultation, the note and its sign-off; the agent never writes into the record unreviewed; the receipt records who approved which note and which follow-up closed
The engagement on top
a custom engagement: connectors loaded on this pull to the trust's electronic patient record and appointment system; a medical secretary or clinical-informatics professional as the paid professional on the match; the outpatient pack's follow-up and letter task templates; the scribe itself stays the trust's own
Hospital group moved core systems to AWS and put generative AI on two jobs: auto-summarising medical records for doctors and capturing inventory data; operational forecasting, patient engagement and decision support are named as next
Where
ID, Southeast Asia
Segment
enterprise
Problem
clinician paperwork, nurse administration, legacy IT
Technology
AWS cloud, generative AI record summarisation, AI inventory capture
Who uses it
Siloam Hospitals Group (CEO Caroline Riady), Indonesia; AWS Indonesia
Outcome
none stated
Effort
high
Risks
Wrong summary in a medical record; patient-data privacy; time-saved figures not given.
How to adopt
Prerequisite: core hospital systems moved to a cloud platform. First step: use generative AI to summarise medical records for doctors and to capture inventory data. Running: IT and clinical review of summaries.
How a room runs it
the ROOM for a hospital's outpatient week: the unit head, the ward nurse manager, the procurement officer and an administration teammate (an agent definition with the hospital back-office skills); the agent reads the record summaries and stock counts the hospital already holds and GENERATES the week's tasks (restock lines for the ward store, summaries awaiting the doctor's review before they are filed) as plan steps with owners and due dates, and a purchase proposal above the ward's envelope lands in the procurement officer's Inbox with a deadline; what the humans keep: every clinical note and its sign-off, the order itself, the spend above the envelope; the receipt records which summary the doctor accepted, edited or rejected and which stock task closed
The engagement on top
a custom engagement: the hospital's own systems stay theirs — connectors loaded on this pull to the hospital information system (HIS), the pharmacy and stores module and the cloud file store; a hospital administrator or health-informatics professional joins as the paid professional on the match; the sector pack carries the ward-round, discharge and restock task templates
A clinical AI platform lets physicians query in natural language across electronic records, notes, pathology reports and imaging at the point of care, starting with prostate cancer
Where
AE, Middle East
Segment
enterprise (private hospital)
Problem
clinical data sits in separate systems and takes days or weeks to assemble
Technology
clinical AI scientist platform (Aila) combining records, pathology and imaging
Who uses it
Cleveland Clinic Abu Dhabi (UAE), Owkin (France)
Outcome
none stated
Effort
high
Risks
No results published; patient privacy and consent; answers drawn from incomplete records.
How to adopt
Start with one disease area and four data sources, keep clinicians checking every answer, and agree who may query what before launch.
How a room runs it
the ROOM for a hospital's oncology service: the lead physician, a data steward, the privacy officer and a clinical-data teammate (an agent definition with the hospital's data policy); the agent teammate reads the cohort definitions and de-identified extracts the hospital already holds and GENERATES the data steward's tasks (validate the registry fields by Friday) as plan steps with owners and due dates, and each new data-use request for the privacy officer (approve / decline, with the purpose attached) as a decision block in the Inbox with a deadline; what the humans keep: the diagnosis, the treatment decision and every consent question; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the electronic record export, the pathology report store, the imaging archive index), a clinical-informatics specialist joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Nine-site London trial of an ambient AI scribe across hospitals, GP practices, mental health services and ambulance teams: more than 17,000 encounters, clinicians check and edit every note
Where
GB, Europe
Segment
public
Problem
note-taking load, short appointments, A&E throughput
Technology
TORTUS ambient voice scribe
Who uses it
Great Ormond Street Hospital (GOSH) with NHS England sponsorship; nine London NHS sites, England
Outcome
23.5 % more direct interaction time, 8.2 % shorter appointments and 92 % patient consent across 17,000+ encounters (hospital-led trial)
Effort
medium
Risks
Patient consent in sensitive settings; error in notes; findings are from a trial and may not hold at scale.
How to adopt
Prerequisite: consent and a clinical-safety review. First step: trial an ambient scribe at a few sites. Running: clinicians check each note; track evaluation measures.
How a room runs it
the ROOM for an emergency or outpatient department: the department lead, the shift nurse in charge, the clinicians and the teammate; the agent GENERATES the shift's administrative tasks around the approved notes (letters to send, results to chase, patient follow-up calls) as plan steps with owners and due dates, and flags to the lead a pile-up of unsigned notes as a decision with a deadline; what the humans keep: the clinical decision, the consent and the sign-off of every note; the receipt records review time and closed follow-ups
The engagement on top
a custom engagement: connectors loaded on this pull to the department's EPR and rota; a clinical-operations or medical-secretary professional as the paid professional on the match; the sector pack's shift-handover and follow-up task templates
Memorandum to bring AI into radiology, pathology, interventional cardiology and clinical operations, plus AI in telemedicine, electronic medical records and remote monitoring, with training for clinicians, IT staff and medical students
Where
ID, Southeast Asia
Segment
enterprise
Problem
clinical workflow and decision support at scale; skills gap
Technology
Philips AI solutions, EMR, telemedicine, remote patient monitoring
Who uses it
Siloam Hospitals Group, Royal Philips, Universitas Pelita Harapan Medical Sciences Group, Indonesia
Outcome
none stated
Effort
high
Risks
MoU only; clinical safety of AI results; skills gap; cost.
How to adopt
Prerequisite: electronic medical records and imaging systems. First step: pilot AI in radiology or pathology with a vendor and train clinicians. Running: clinical validation and IT staff.
How a room runs it
the same hospital room, on the operations side: the radiology head, the clinical-operations manager and the teammate; the agent GENERATES the non-clinical tasks that surround an imaging or cardiology workflow (booking the slot, chasing the referral letter, preparing the pre-procedure checklist for the nurse, the follow-up call list) and files a decision for the manager when a queue breaches its target, with a deadline; what the humans keep: the reading of every image, the diagnosis and every patient-facing decision stay with the licensed specialist; the receipt records each task's owner and close time
The engagement on top
a custom engagement: connectors loaded on this pull to the hospital's RIS/PACS worklist and appointment system (read-only on the images), a clinical-operations manager as the paid professional on the match, and the hospital-operations pack's task templates; the MoU is intent only, so the engagement starts from what the hospital already runs
Source: vendor · 2024 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 8 adoptions here, 0 were low effort, 4 medium and 4 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: appointment system and clinical triage rules. First step: replace the 8 am phone rush with an online request form and automated triage. Running: clinicians review flagged requests; slot lengths adjusted.
Ambient AI note-taking for clinicians across two academic health systems, studied for burnout
Where
US, North America
Segment
enterprise
Problem
documentation burden and clinician burnout
Technology
ambient documentation technology (AI scribe) in the electronic record
Who uses it
Mass General Brigham (Massachusetts) and Emory Healthcare (Georgia), United States
Outcome
More than 1,400 clinicians tracked; at Mass General Brigham burnout fell from 52.6% to 30.7% after 84 days (JAMA Network Open, as reported Aug 2025)
Effort
medium
Risks
Self-reported survey with low response rates; patient consent and recording privacy; note errors.
Next: burnout had already fallen to 29.4% by day 42, so gains show early source
How to adopt
Pilot with volunteers, measure burnout at day 0, 42 and 84, require clinician sign-off on every note, then widen.
How a room runs it
the clinic room: the clinician, the patient with consent, and a scribing teammate; the agent drafts the note from the visit and GENERATES the follow-up tasks (orders to place, letters to send) with due dates, and each draft is a proposal in the clinician's Inbox that only the clinician can sign; the receipt records the draft, the edits and the signer
The engagement on top
connectors to the record system and the visit audio, a consent template, a clinician who owns the pilot
Head-to-head scribe selection, then rollout to ambulatory physicians and advanced practice providers
Where
US, North America
Segment
enterprise
Problem
documentation time and clinician retention
Technology
Ambience Healthcare AI Scribe, Epic
Who uses it
Cleveland Clinic, United States
Outcome
More than 4,000 of about 6,000 eligible clinicians active within 15 weeks; 1 million encounters; note time down 2 minutes per appointment and 14 minutes per day (clinic's figures, Aug 2025)
Effort
high
Risks
Exclusive multi-year vendor tie; patient consent and opt-out; self-reported figures.
Next: the clinic's associate CMIO expects AI scribes to become something physicians expect their employer to offer source
How to adopt
Score several vendors in a pilot, keep use optional, require physician approval, roll out in waves.
How a room runs it
the clinic room: the clinician, the patient and a scribe agent; the agent drafts a structured note and after-visit summary and GENERATES coding and follow-up tasks with due dates, each a proposal the clinician approves; use stays optional and patients can opt out, recorded on the receipt
The engagement on top
connectors to the record, a selection scorecard, an opt-out flow
Enterprise licences for an AI clinical-documentation assistant after a controlled study with primary-care clinicians
Where
US, North America
Segment
enterprise
Problem
documentation burden in primary care
Technology
Nuance DAX Copilot (Microsoft), Epic integration
Who uses it
Advocate Health (Southeast region), United States, with Microsoft
Outcome
125 clinicians studied; about 17 minutes saved per day on average (control group saved about 11), nearly 20% saved an hour or more daily; licences then expanded enterprise-wide
Effort
medium
Risks
Modest net time saving once the control group is counted; cognitive-load benefit is self-reported; no revenue gain found.
Next: after the study Advocate expanded to enterprise-wide licences source
How to adopt
Run a controlled study with a comparison group, publish the net effect, then license enterprise-wide.
How a room runs it
the clinic room: the clinician, the patient and a scribing teammate; the agent drafts the note and GENERATES the coding and follow-up tasks with due dates, each as a proposal the clinician approves; the receipt shows draft, reviewer and changes, and a control group stays on the old way so the saving is measured, not assumed
The engagement on top
connectors to the record, a control-group measurement plan, a clinician champion
Ambient AI scribe for GPs and clinicians, used across practices and health groups in several countries
Where
AU;NZ;GB;US, Oceania;Europe;North America
Segment
SMB and mid (GP practices, health groups)
Problem
note-writing time after consultations
Technology
ambient speech capture, generative note drafting
Who uses it
Heidi Health customers including Bupa, Tamaki Health, MaineGeneral Health and Beth Israel Lahey Health (Australia, New Zealand, UK, United States); Tuggerah Family Doctors, Australia
Outcome
2,110,485 patient visits enabled in the week shown on the vendor's page; no outcome study named the company’s own figure
Effort
low
Risks
Vendor-only evidence; patient consent and privacy rules differ by country; draft errors.
How to adopt
Pilot with a few clinicians, confirm patient consent and recording rules, require review of each draft and measure note time.
How a room runs it
the ROOM for a practice: the GP, the practice manager and a notes teammate; the agent teammate checks which consult notes are drafted and which are unsigned; GENERATES each GP's sign-off on drafts and the manager's weekly sample check as plan steps with owners and due dates, and the practice owner's choice to extend the tool to all clinicians as a decision block with a deadline in the Inbox of the practice owner; what the humans keep: all clinical judgement and the signature on each note; the receipt records each signed note and the sample checks; the client receives the reports and gives direction in the room
The engagement on top
the client's own systems stay theirs, wired to the room through connectors loaded on this pull (the practice-management system, the scribe tool); the paid professional (a practice-management adviser) joins on the match
Source: vendor · 2026 · vendor only; to corroborate
A pre-deployment review process (OPTICA, a 77-item checklist) for every AI product, and a model on 25 years of patient records that flags people at high risk of undiagnosed hepatitis C
Where
IL, Middle East
Segment
enterprise (integrated health fund and provider)
Problem
mass screening finds few carriers; unreviewed AI products carry clinical risk
Technology
risk-prediction model on health records, governance checklist, shared imaging platform running several vendors' algorithms
Who uses it
Clalit Health Services (Israel); WHO report cited; Prof. Ran Balicer as quoted
Outcome
38 carriers found among about 500 at-risk people flagged by the model, versus only dozens found in a screening of about 50,000 (reported 100-fold gain; who measured is not stated)
Effort
medium
Risks
Source is a news report of a WHO document and officials, undated; the detection comparison is not independently shown; checklist can slow adoption.
How to adopt
Write the review checklist before buying, require clinical, data and monitoring plans per product, and keep monitoring after launch rather than approving once.
How a room runs it
the ROOM for a health fund's AI review board: the chief innovation officer, a clinical lead, the data lead, a patient representative and a review teammate (an agent definition with the checklist); the agent teammate reads the vendor documents and validation files the fund already holds and GENERATES the checklist sections to complete per product (finish data assessment by the 20th) as plan steps with owners and due dates, and each product's approval for the board chair (deploy / reject, with the checklist attached) as a decision block in the Inbox with a deadline; what the humans keep: the clinical judgement, the approval and the monitoring after launch; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the product intake files, the validation reports, the post-deployment monitoring export), a clinical-informatics specialist joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Oxford University Hospitals NHS Foundation Trust, United Kingdom
Outcome
88% of clinicians measured time savings; 22.2% saved an hour or more per day; 99.7% of patients consented (trust evaluation, Feb 2026)
Effort
medium
Risks
Hallucinated content (44.4% of users saw it at least once); clinical edits needed on 126 of 398 uses; patient consent.
Next: phased deployment with one solution for most use cases and a few licences for specialties source
How to adopt
Compare several tools in one pilot, log every correction, consent every patient, deploy in phases on one supplier.
How a room runs it
the clinic room: clinician, patient and a scribe agent; the agent drafts the note and GENERATES letters and task follow-ups with due dates; every draft is a proposal the clinician edits and signs, and edits (including hallucination flags) are logged on the receipt so the trust can compare tools on evidence
The engagement on top
connectors to the record, a consent script, an error-reporting form
GP practice replaced the 8 am phone rush with an AI-assisted online triage: requests are triaged and 91 % of appointments allocated automatically, GP slots lengthened
Where
GB, Europe
Segment
SMB and mid
Problem
phone queues, long waits to pre-bookable appointments
Technology
Smart Triage (Rapid Health) online request and triage
Who uses it
The Groves Medical Centre (Surrey, England); Rapid Health; evaluation by Unity Insights
Outcome
Pre-bookable wait 11 to 3 days, peak calls down 47 %, no-shows down 14 % (practice evaluation, Oct 2023 to Feb 2024, as reported)
Effort
medium
Risks
Digital exclusion of patients who cannot use online forms; missed clinical urgency; vendor lock-in.
How to adopt
Prerequisite: appointment system and clinical triage rules. First step: replace the 8 am phone rush with an online request form and automated triage. Running: clinicians review flagged requests; slot lengths adjusted.
How a room runs it
the ROOM for a GP practice: the practice manager, the receptionists, the duty GP and a front-desk teammate; the agent takes the requests the practice already receives and GENERATES the day's tasks (book the follow-up, send the form, call back the patient who could not use the online form) as plan steps with owners and due dates, with the rule that any request carrying a red-flag word goes straight to the duty GP's Inbox as an urgent proposal; what the humans keep: the clinical triage decision, the consultation and every prescription; the receipt records who handled which request
The engagement on top
a custom engagement: connectors loaded on this pull to the practice's clinical system and appointment book; a practice manager as the paid professional on the match; WhatsApp for patients who do not use the web form; the clinic pack's recall, referral-chasing and results-letter templates
Ministry drive for every health facility and electronic-medical-record provider to register and integrate with the national SATUSEHAT platform; clinics, hospitals, laboratories and pharmacies are in the target
Where
ID, Southeast Asia
Segment
public and SMB
Problem
fragmented patient data, unconnected facilities
Technology
SATUSEHAT interoperability platform, electronic medical records (RME)
Who uses it
Ministry of Health, Indonesia; clinics, pharmacies and hospital RME providers
Outcome
792 facilities joined at 2023-09-04 (ministry page)
Effort
high
Risks
Poor integration quality; patient data privacy; facility counts per type not published.
How to adopt
Prerequisite: an electronic medical record that supports the national standard. First step: register the facility and integrate the EMR through the platform. Running: IT or vendor support and data-quality staff.
How a room runs it
the ROOM for an independent clinic or a pharmacy: the owner-doctor or apothecary, the nurse or assistant, the admin clerk and a compliance teammate; the agent reads what the clinic already records and GENERATES the tasks the national connection requires (the visits still to be reported, the master-data fixes, the registration steps) as plan steps with owners and due dates, and a decision on which RME vendor to pick lands in the owner's Inbox as a proposal with a deadline and the cost beside it; what the humans keep: the clinical record content, the patient's consent and the vendor choice; the receipt records each submission
The engagement on top
a custom engagement: the clinic's own RME or pharmacy system wired through connectors loaded on this pull to the SATUSEHAT API where it exposes one; a health-information-management professional as the paid professional on the match; the clinic pack's registration and reporting templates
Dental clinic automated recall self-booking, four-step confirm-or-reschedule reminders, waitlist backfilling and after-hours AI triage that escalates urgent symptoms to the on-call dentist
Where
CA, North America
Segment
SMB
Problem
hygiene no-shows, overdue recalls, late cancellations
Technology
AI agent on Open Dental, SMS and email, chat and phone triage
Who uses it
six-chair dental clinic in Toronto, Canada (unnamed); Builts AI
Outcome
No-show rate 15.8 % to 9.3 %, recall conversion 11 % to 31 %, about 18 front-desk hours a week reclaimed (vendor's own figures, 90 days) the company’s own figure
Effort
medium
Risks
Vendor-reported figures from an anonymous clinic; missed emergency triage; patient consent for messaging.
How to adopt
Prerequisite: practice-management system with an integration route, SMS and email. First step: automate recall and reminders, then waitlist backfill. Running: front desk reviews after-hours triage escalations.
How a room runs it
the ROOM for a dental practice: the dentist-owner, the hygienist, the front-desk coordinator and the recall teammate; the agent reads the recall list and the diary and GENERATES tasks (message these 40 overdue patients, offer this cancelled slot to the next on the waitlist, flag the sedation case for a human call) as plan steps with owners and due dates; what the humans keep: the diagnosis, the treatment plan, every urgent-symptom call and the quote for work; insurance checks are tasks the coordinator closes; the receipt records each recall sent and each slot filled
The engagement on top
a custom engagement: connectors loaded on this pull to the practice-management system (Open Dental in the case; in Indonesia the clinic's own appointment and billing tool); a dental practice manager as the paid professional on the match; WhatsApp reminders (the channel is built)
Source: vendor · 2026 · vendor only; to corroborate
Health and fertility clinics run patient chat on one platform: AI agent answers FAQs, books appointments, sends WhatsApp reminders, post-consultation prescription delivery updates and feedback questionnaires
Where
ID, Southeast Asia
Segment
SMB and mid
Problem
scattered chat channels, queues, staff load
Technology
Qiscus omnichannel chat with AI agent over WhatsApp
96 % satisfaction and a 30 % retention-sales figure at one clinic customer (vendor's own) the company’s own figure
Effort
low
Risks
Vendor figures not corroborated; patient data privacy; wrong medical information.
How to adopt
Prerequisite: a messaging channel account and FAQ content. First step: put an AI agent on WhatsApp for FAQs, bookings and reminders. Running: staff handle escalations.
How a room runs it
the ROOM for a clinic front desk: the clinic manager, the registration staff, the doctor and the patient-chat teammate; the teammate answers only from the clinic's own schedule and price list and GENERATES the tasks the chat creates (the slot to confirm, the call-back for a question it may not answer, the prescription to deliver) as plan steps with owners and due dates; a medical question is never answered by the agent, it lands in the doctor's Inbox; what the humans keep: every clinical answer and prescription; the receipt records each handoff
The engagement on top
a custom engagement: connectors loaded on this pull to the clinic's own registration and doctor-schedule system; a clinic manager as the paid professional on the match; WhatsApp is the door (built)
Source: vendor · 2026 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 10 adoptions here, 2 were low effort, 6 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: a self-referral website and a clinical safety review. First step: add a chatbot for intake that, with consent, passes standard questionnaire scores to the referral. Running: clinicians triage and monitor.
Free 24-hour psychologist tele-consultation inside the city's public super-app: the person opens the health section, is connected to an on-duty clinical psychologist within minutes, and is referred onward only if registered in the city
Where
ID, Southeast Asia
Segment
public
Problem
access to mental-health support, queues, travel to a facility
Technology
JAKI city app, tele-consultation
Who uses it
DKI Jakarta provincial government, Jakarta Health Office, Indonesia
Outcome
none stated
Effort
medium
Risks
Human service with no AI claimed; demand outstripping staff; referral only for registered residents.
How to adopt
Prerequisite: a city app and on-duty clinical psychologists. First step: add a free tele-consultation entry in the app. Running: staffed 24 hours with a referral pathway.
How a room runs it
the ROOM behind such a service: the duty-psychologist coordinator, the on-call psychologists, the referral officer and the teammate; the agent prepares each session's admin (the intake form filled before the call, the rota gaps for the coming week, the referral letter drafts) and GENERATES follow-up tasks with owners and due dates (call back after a missed session, send the referral to the facility near the person's address); a rota gap or an unusual call volume lands in the coordinator's Inbox as a decision with a deadline; what the humans keep: the session, the assessment, every safety judgement and the referral decision; the receipt records the hand-offs
The engagement on top
a custom engagement: connectors loaded on this pull to the service's rota, booking and referral directory; a clinical psychologist as the paid professional on the match; the channel the people already use (WhatsApp) alongside the app
AI self-referral chatbot on NHS Talking Therapies websites collects intake and, with consent, PHQ-9/GAD-7/WSAS scores onto the referral so clinicians triage and size the assessment slot
Where
GB, Europe
Segment
public
Problem
long waits to assessment, dropout, long intake calls
Technology
Limbic Access conversational AI chatbot
Who uses it
Everyturn Mental Health, nine NHS Talking Therapies services, England; Limbic
Outcome
Assessment wait 15.2 vs 17.4 days; dropout 21.9 % vs 26.7 % (observational study; authors are vendor employees)
Effort
medium
Risks
Mental-health risk handling and escalation; consent; vendor-authored evidence.
How to adopt
Prerequisite: a self-referral website and a clinical safety review. First step: add a chatbot for intake that, with consent, passes standard questionnaire scores to the referral. Running: clinicians triage and monitor.
How a room runs it
the ROOM for a talking-therapies service: the service lead, the intake clinician, the administrator and the intake teammate; the agent answers the person who writes in only with the service's own information and GENERATES the administrative tasks from the referral (book the assessment slot, send the consent form, chase the missing questionnaire) as plan steps with owners and due dates; any risk word or score over the service's threshold is not handled by the agent but lands at once in the duty clinician's Inbox as an urgent proposal; what the humans keep: the assessment, every risk decision and the treatment plan; the receipt records who saw which referral and when
The engagement on top
a custom engagement: connectors loaded on this pull to the service's patient record and booking system; a clinical lead (psychologist or senior therapist) as the paid professional on the match who sets the escalation rules the agent obeys; WhatsApp or web intake; the mental-health pack's intake and follow-up templates
Ambient AI scribe trialled in mental health services among nine London sites, clinicians review every note; public polling shows consent limits in this setting
Where
GB, Europe
Segment
public
Problem
note-taking in therapy and psychiatric consultations, sensitive data
Technology
TORTUS ambient voice scribe
Who uses it
NHS England-sponsored trial led by GOSH, including mental health services, England; Healthwatch England
Outcome
Trial-wide, not mental-health-specific: 23.5 % more direct interaction time, 8.2 % shorter appointments, 92 % consent across 17,000+ encounters at nine NHS sites (hospital-led trial; page gives no separate mental health figures)
Effort
medium
Risks
Patient comfort is low in this setting (a poll of 4,039 adults found 28 % comfortable); sensitive data; trial does not report mental-health results separately.
How to adopt
Prerequisite: consent and a clinical-safety review. First step: trial an ambient scribe at a mental-health service. Running: clinicians check every note.
How a room runs it
the ROOM for a community mental-health team: the team manager, the care coordinator, the clinician and the teammate; the teammate does the paperwork around a session only when the person has said yes in advance (the consent is a task the clinician closes first): it GENERATES the letter drafts, the care-plan review dates and the appointment reminders as plan steps with owners and due dates; what the humans keep: the session, the clinical record, consent and every risk decision; where the person says no the room works with no recording at all; the receipt records the consent and the review
The engagement on top
a custom engagement: connectors loaded on this pull to the team's record system and diary; a clinical lead as the paid professional on the match who decides what may be recorded; the mental-health pack's care-plan-review and letter templates
A generative-AI therapy chatbot app tested in a randomised trial for depression, anxiety and eating-disorder concerns
Where
US, North America
Segment
research (randomised trial)
Problem
shortage of mental health clinicians; access to support
Technology
generative AI chatbot, smartphone app
Who uses it
Dartmouth Geisel School of Medicine and Center for Technology and Behavioral Health (Michael Heinz, Nicholas Jacobson), United States; Therabot
Outcome
Average symptom reduction 51 % for depression, 31 % for generalised anxiety and 19 % for body-image and weight concerns versus controls (Dartmouth researchers, randomised trial, NEJM AI)
Effort
high
Risks
Safety in a crisis; short trial; generalisation beyond the study group; regulatory status of therapy chatbots.
How to adopt
Treat it as supervised adjunct support, set clear crisis hand-off rules, test on a small group with clinician oversight, and measure symptom scales.
How a room runs it
the ROOM for a counselling service: the clinical lead, the counsellors and a support-companion teammate; the agent teammate handles between-session check-ins within limits the lead set and flags risk phrases; GENERATES each counsellor's follow-up on a flagged check-in the same day as plan steps with owners and due dates, and the clinical lead's choice on which topics the companion may handle as a decision block with a deadline in the Inbox of the clinical lead; what the humans keep: all clinical judgement, risk assessment and crisis response; the receipt records each flag raised, who followed up and when; the client receives the reports and gives direction in the room
The engagement on top
the client's own systems stay theirs, wired to the room through connectors loaded on this pull (the booking system, the service's own messaging channel); the paid professional (a licensed clinical supervisor) joins on the match
Your own team builds it. Across the 4 adoptions here, 0 were low effort, 3 medium and 1 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: consent from clients and a secure recording process. First step: a pilot with a small group of social workers using AI case-note drafting. Running: worker edits every draft; licences.
Social workers record meetings and AI drafts the case notes and assessment text, which the worker edits; used across local authorities
Where
GB, Europe
Segment
public
Problem
hours lost to case recording and assessment writing
Technology
Magic Notes (Beam) meeting transcription and summarisation
Who uses it
Kent County Council, Warrington, Somerset and more than 100 local authorities in England; Beam
Outcome
About 7 hours a week saved per staff member in the Kent pilot (vendor-sourced) the company’s own figure
Effort
low
Risks
Privacy and consent of vulnerable clients; wrong case notes; the evaluation was not opened.
How to adopt
Prerequisite: consent from clients and a secure recording process. First step: a pilot with a small group of social workers using AI case-note drafting. Running: worker edits every draft; licences.
How a room runs it
the ROOM for an adult-social-care team: the team manager, the social worker, the occupational therapist and the case-admin teammate; with the person's consent, the agent drafts the record and GENERATES the follow-ups from the worker's approved note (the referral to send, the equipment request, the review visit in six weeks) as plan steps with owners and due dates, and a change to a care package is a proposal in the manager's Inbox with a deadline; what the humans keep: the assessment, the eligibility decision, the conversation and the signing of every note; the receipt records edit and approval
The engagement on top
a custom engagement: connectors loaded on this pull to the council's case-management system and the shared calendar; a registered social worker as the paid professional on the match; the social-care pack's assessment, review and referral templates
Source: vendor · 2025 · vendor only; to corroborate
Digital home-care provider books live-in caregivers for elderly people with dementia through the web or phone, with video interviews of candidate caregivers
Where
ID, Southeast Asia
Segment
SMB
Problem
finding trusted caregivers, family distance
Technology
website booking, video calls, caregiver training
Who uses it
Insan Medika (about 10,000 patients since 2013), Indonesia
Outcome
Nearly 10,000 patients served (company claim as reported by Warta Ekonomi; no outcome measure)
Effort
low
Risks
No AI claimed; trust and vetting of caregivers; quality of care.
How to adopt
Prerequisite: a pool of trained caregivers. First step: web or phone booking with video interviews. Running: caregiver training and vetting.
How a room runs it
the ROOM for a home-care provider and one family: the provider owner, the care coordinator, the live-in caregiver, the client's family and the teammate; the agent GENERATES the matching and handover tasks (shortlist caregivers for the family's video call, the shift handover note, the weekly family update drafted for the coordinator's check) as plan steps with owners and due dates, and a change of caregiver is a proposal for the family in their Inbox with a deadline; what the humans keep: the care itself, medication given and any call to a doctor; the receipt records visits and updates
The engagement on top
a custom engagement: connectors loaded on this pull to the provider's booking and payroll tools; a nurse or care manager as the paid professional on the match; family updates over WhatsApp
Wearable-sensor and AI monitoring of residents in senior living, detecting changes in routine so staff can act early
Where
US, North America
Segment
mid and enterprise (senior living operators)
Problem
late detection of decline, infections and falls; staff workload
Technology
wearable sensors, behavioural analytics, alerts
Who uses it
Edgewood Healthcare, Lifewell Senior Living, Bickford, Paradise Pointe and Maple Knoll Communities, United States; CarePredict
Outcome
Operators quoted on the vendor's page: 39 % fewer hospitalisations (Lifewell), 37 % gain in staff efficiency (Paradise Pointe), 60 % faster UTI recovery through earlier detection (Bickford), $320k+ unbilled care captured (Edgewood); vendor-reported, no method the company’s own figure
Effort
high
Risks
Vendor-reported figures without method; resident consent and privacy; alert fatigue; hardware upkeep.
How to adopt
Pilot on one floor with consent from residents and families, agree who acts on each alert, and compare hospitalisations and staff time to a baseline.
How a room runs it
the ROOM for a senior-living community: the director of care, the nurses and a resident-watch teammate; the agent teammate reads the alerts the sensor platform already raises and groups them by resident; GENERATES each nurse's same-day check on a resident whose routine changed as plan steps with owners and due dates, and the director's choice on a care-plan change or a family call as a decision block with a deadline in the Inbox of the director of care; what the humans keep: every care and clinical judgement and the family conversation; the receipt records each alert, who checked and the care-plan change; the client receives the reports and gives direction in the room
The engagement on top
the client's own systems stay theirs, wired to the room through connectors loaded on this pull (the sensor platform, the care-record system, the staff rota); the paid professional (a registered nurse adviser) joins on the match
Source: vendor · 2026 · vendor only; to corroborate
AI scheduling and routing for home-care workers adapts rotas to absences and travel; used by 56 private and government homecare providers
Where
GB, Europe
Segment
SMB, mid and public
Problem
scheduling complexity, absences, carer time pressure
Technology
optimisation algorithms (Webroster, OptifAI)
Who uses it
56 UK homecare providers; Webroster/Access Group, NDG AI Ltd; University of Nottingham
Outcome
Used by 56 UK homecare clients of Webroster/NDG AI; the case study quotes the company that providers can deliver up to 20 % more home care with current resources (supplier's claim, submitted by the University of Nottingham) the company’s own figure
Effort
medium
Risks
No quantified savings in the part read; rota quality; carer well-being.
How to adopt
Prerequisite: rota, visit and travel-time data. First step: use optimisation software on rotas for one provider. Running: schedulers review proposals.
How a room runs it
the ROOM for a home-care agency: the agency owner, the care coordinator, the carers and the rota teammate; the agent takes the visits the agency has and the carers' availability and GENERATES the week's rota proposal and the swap tasks when a carer is sick (who covers, who tells the client) as plan steps with owners and due dates; the coordinator approves the rota in her Inbox with a deadline; what the humans keep: the visit, the care itself and the call on a client who looks unwell; the receipt records which visit was done by whom
The engagement on top
a custom engagement: connectors loaded on this pull to the agency's care-management and payroll systems; a registered manager as the paid professional on the match; carers receive the rota over WhatsApp (built)
Flemish residential care homes pilot voice-activated reporting to cut record-keeping, environmental monitoring and smart incontinence products so staff do not wake residents for routine checks
residential care homes in West Flanders, East Flanders and Flemish Brabant, Belgium; Artevelde University College; Alzheimer's Research Foundation
Outcome
none stated
Effort
high
Risks
Pilots with no results; resident privacy and consent; false alerts.
How to adopt
Prerequisite: care homes willing to pilot and sensors and voice devices installed. First step: pilot voice reporting and environmental monitoring in some homes. Running: staff training and evaluation.
How a room runs it
the ROOM for a care home: the home manager, the head nurse, the carers and the teammate; the carers' spoken reports become draft entries the nurse approves, and the agent GENERATES the follow-ups (call the GP about the resident's sleep, review the fall-risk plan on Thursday) as plan steps with owners and due dates; a sensor alert the home already receives is a task for the night carer with the escalation to the nurse as a proposal; what the humans keep: all care, the nurse's judgement and what goes into the care plan; the receipt records who acted on which alert
The engagement on top
a custom engagement: connectors loaded on this pull to the home's care-planning software and the sensor vendor's alert feed (the agent reads, it does not own sensors); a registered nurse as the paid professional on the match; the care-home pack's shift and care-plan templates
Your own team builds it. Across the 5 adoptions here, 2 were low effort, 1 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: a national teacher platform with a curriculum template library and teacher log-ins. First step: add generative drafting of lesson plans and modules, with the teacher editing every draft. Running: prompt and template upkeep and teacher training.
A ministry of education puts a generative-AI teaching assistant inside the national teacher platform to draft lesson plans, teaching modules and analyse student results
Where
ID, Southeast Asia
Segment
public
Problem
teacher time lost to administrative documents
Technology
generative AI inside the Platform Merdeka Mengajar (PMM), NLP drafting of RPP and modul ajar, automatic analysis of learning results
Who uses it
Kementerian Pendidikan Dasar dan Menengah (Mendikdasmen), Indonesia
Outcome
none stated
Effort
medium
Risks
Teacher adoption can lag (a community paper cites low platform usage); generic or inaccurate drafts; student-data exposure.
How to adopt
Prerequisite: a national teacher platform with a curriculum template library and teacher log-ins. First step: add generative drafting of lesson plans and modules, with the teacher editing every draft. Running: prompt and template upkeep and teacher training.
How a room runs it
the ROOM for a school's teaching week: the head teacher, the teachers and a school-admin teammate (an agent definition with the curriculum-paperwork skills); the agent drafts the term's lesson-plan and module paperwork from the curriculum the school already uses and GENERATES tasks for each teacher (review and sign your module by Friday) and for the head (approve the term calendar by the 15th) as plan steps with owners and due dates; the head's approvals land in the Inbox as proposals with a deadline; what the humans keep: every grade, every judgement of a student, the sign-off on each document; the receipt records which draft was accepted, edited or rejected and by whom
The engagement on top
a custom engagement: the school's own systems (the student information system, the timetable, the ministry reporting portal) wired to the room through connectors loaded on this pull; an education administrator joins as the paid professional on the match; the education pack carries the module, calendar and report templates
A national ministry rolls out AI-driven digital textbooks for maths, English and computer science in schools, then makes them voluntary and withdraws their textbook status after complaints
Where
KR, East Asia
Segment
public
Problem
uniform lessons for pupils of different levels; teacher time on tracking progress
Technology
AI adaptive digital textbooks from about a dozen approved publishers
Who uses it
Ministry of Education and approved publishers, South Korea
Outcome
School use fell from 37 % to 19 % in the second semester, to 2,095 schools (as reported by Rest of World); no learning result stated
Prerequisite: a pilot with teachers before a mandate and enough time for content review. First step: let willing schools try one subject and log problems. Running: keep use voluntary until errors and workload are measured.
How a room runs it
the ROOM for a school's curriculum team: the head of maths, two teachers and a pilot teammate; the agent collects what teachers log about each digital lesson and GENERATES tasks (check the three questions flagged as wrong in unit 4 by Friday; list the pupils who fell behind) and a proposal to the head to continue, narrow or stop a tool for next semester, with a deadline; what the humans keep: the choice of tool, every grade, any decision to drop it; the receipt records the issues raised, the evidence and who decided
The engagement on top
a custom engagement: the school's own learning platform and teacher feedback forms wired through connectors loaded on this pull; a curriculum adviser as the paid professional on the match; the education pack's tool-review task templates
A school district connects an AI attendance assistant to its student information system: it texts families within 15 minutes of an unexpected absence in 100+ languages, takes replies and documents, marks the absence excused and flags only the hard cases to staff
Where
US, North America
Segment
public
Problem
office staff chasing absences by phone; families who do not read English
Technology
AI messaging assistant linked to the student information system, multilingual SMS, document photos, escalation flags (Edia)
Who uses it
Raton Public Schools, New Mexico, USA (named); other New Mexico districts use the same tool
Outcome
none stated
Effort
medium
Risks
Incorrect auto-excusing of absences; mistranslation; storing family documents and photos; one district claim could not be confirmed.
How to adopt
Prerequisite: a student information system with an integration route and verified family phone numbers. First step: connect an AI messaging assistant for unexcused-absence texts in the families' languages. Running: staff handle only the flagged cases; review translations and the excusal rules.
How a room runs it
the ROOM for a school office: the attendance clerk, the vice principal and an attendance teammate; the agent reads the absence list the school already holds and GENERATES tasks (message the family of student X, in their language) and, when a parent stops replying or mentions bullying or transport, a task for the counsellor by the next day; the vice principal's decision to open a case lands in the Inbox as a proposal with a deadline; what the humans keep: any call to a family in distress, the decision to excuse beyond policy, every sanction; the receipt records each message, each excused absence and who approved it
The engagement on top
a custom engagement: the school's own student information system and parent contact list wired through connectors loaded on this pull; WhatsApp is the channel in Indonesia (built); an education administrator as the paid professional on the match; the attendance task templates in the education pack
Two-way translated school-to-family messaging with AI plus human translators; families join by text with no app
Where
US, North America
Segment
public and mid
Problem
families who speak other languages cannot follow teacher messages
Technology
AI translation with human reviewers, context glossary, readability flag, SMS
Who uses it
TalkingPoints (US nonprofit) with school districts; the interview reports 10 of the 20 largest US districts as partners
Outcome
Founder-reported study with a third-party researcher: maths gains equal to 7 months of extra learning and 2,000 more classroom hours from higher attendance, in one large urban district (TalkingPoints' own account, method not published)
Effort
medium
Risks
Mistranslation of sensitive school messages; consent and data privacy; vendor and counts are the author's own.
How to adopt
Prerequisite: family phone numbers and consent to text. First step: run two-way messages through AI translation with human translator review for the main languages. Running: maintain a glossary and the reviewer pool.
How a room runs it
the ROOM for parent communication: the class teacher, the school secretary and a family-liaison teammate; the agent drafts the weekly class note and each family message in the family's language, and GENERATES the tasks (send the field-trip consent reminder to 14 families by Wednesday, check three unanswered replies) with owners and due dates; any message that commits the school or touches a student's conduct lands in the teacher's Inbox as a proposal before sending; what the humans keep: the tone of every difficult message, the meeting with the parent; the receipt records what was sent, in which language, and the replies
The engagement on top
a custom engagement: the school's parent contact list and student records wired through connectors loaded on this pull; WhatsApp, the channel parents already use, is built; a bilingual education liaison as the paid professional on the match; the parent-note templates in the education pack
A national body gives primary teachers an AI lesson assistant that drafts lesson plans, quizzes and resources from the national curriculum for the teacher to edit, tested in a randomised controlled trial
Where
GB, Europe
Segment
public
Problem
teachers spending 3+ hours a week preparing lessons and resources
Technology
AI lesson assistant on a national curriculum resource bank (Aila)
Who uses it
Oak National Academy (public body, England); Education Endowment Foundation and NFER (evaluators), United Kingdom
Outcome
Randomised trial of Year 3-6 teachers over ten weeks: about 49 minutes a week saved (2.5 h against 3.3 h, 76 % of the comparison group's time), materials no worse in quality (independent assessment, EEF/NFER, as reported)
Effort
low
Risks
Uneven use (about 7 % of lessons); errors in drafted content; trial covered primary only; the page does not give the sample size.
How to adopt
Prerequisite: a curriculum and scheme of work in digital form and a rule for who checks drafts. First step: give a few teachers an AI lesson assistant for planning and compare preparation time. Running: teachers edit every draft; track use, which fell during the trial.
How a room runs it
the ROOM for a primary school's planning team: the head of year, three class teachers and a planning teammate; the agent reads the term's scheme of work and GENERATES tasks (draft the Year 4 fractions sequence for the teacher by Thursday, check it against the school's marking policy) with owners and due dates, and a proposal to the head to adopt a shared resource set, with a decision date; what the humans keep: every lesson as taught, every assessment judgement, every pupil-facing resource they sign off; the receipt records which resource the agent drafted, who edited it and who approved it
The engagement on top
a custom engagement: the school's own scheme of work and curriculum documents wired through connectors loaded on this pull; a former headteacher or curriculum lead as the paid professional on the match; the lesson-planning task templates in the education pack
A state government funds an AI platform for teachers (lesson sequences, plans, resources, tailored student material), piloted across public, Catholic and independent schools, then extended
Where
AU, Oceania
Segment
public
Problem
teacher workload from lesson planning and preparing materials
Technology
generative AI chat platform with editable outputs and tailoring tools (WA ClassmAIte)
Who uses it
Western Australian Department of Education with Catholic Education WA, the Association of Independent Schools of WA and the School Curriculum and Standards Authority, Australia
Outcome
60 schools in the pilot, planned to exceed 100 by end of 2026 (WA Government media statement; count unattributed, no time-saved figure)
Effort
medium
Risks
No time-saved figure published; inaccurate or biased lesson content; student-data handling.
How to adopt
Prerequisite: an agreement with schools of each sector and an approved, privacy-reviewed generative-AI platform. First step: pilot with volunteer teachers and editable outputs. Running: teacher training, evaluation of time saved, and a path to widen the pilot.
How a room runs it
the same school room: the teachers, the head and a planning teammate that reads the school's scope-and-sequence and GENERATES the week's preparation as tasks (draft Year 5 unit 3, adapt for two students with plans) with owners and due dates; the head's decision on a shared resource or a timetable change lands in the Inbox with a deadline; what the humans keep: the teaching, every student judgement, every edit before anything is used; the receipt records what was drafted and what the teacher changed; the school-group office sees the same room across campuses without opening a dashboard
The engagement on top
a custom engagement: the school group's own curriculum documents and timetable system wired through connectors loaded on this pull; a curriculum lead as the paid professional on the match; the education pack's unit and lesson templates
A national ministry gives upper-secondary students and their teachers free access to general AI tools with teacher training first, through a public-private foundation, widening to vocational schools the next year
Where
EE, Europe
Segment
public
Problem
keeping education competitive and personalised while students already use AI; teachers unprepared for it
Technology
education editions of general AI assistants, teacher training programme (AI Leap / TI-Hüpe)
Who uses it
Ministry of Education and Research with the AI Leap Foundation, Estonia
Outcome
none stated
Effort
high
Risks
Teachers trained too late; student over-reliance; data protection for minors; no published evaluation.
How to adopt
Prerequisite: an agreed school or national rule on permitted AI use and trained teachers. First step: train teachers first, then give one cohort of students supervised access. Running: review the rules each term against what students actually do.
How a room runs it
the ROOM for a school's AI-use team: the deputy head, the subject leads and an adoption teammate; the agent GENERATES tasks (run the staff training session for the maths department by the 15th; collect the questions students raised in grade 10) with owners and due dates, and a proposal to the deputy head on which assignments may use AI and which may not, with a decision date; what the humans keep: the assessment rules, every grade, any case of suspected misuse; the receipt records the policy version in force and who approved each change
The engagement on top
a custom engagement: the school's own learning platform and staff list wired through connectors loaded on this pull; a teacher-trainer as the paid professional on the match; the education pack's AI-use policy and training task templates
Your own team builds it. Across the 7 adoptions here, 1 were low effort, 4 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: a knowledge base of student procedures and a messaging channel account (WhatsApp Business). First step: put an AI agent on the busiest channel and route unresolved cases to staff. Running: keep the knowledge base current and monitor wrong answers.
Systemwide ChatGPT Edu for students, faculty and staff across all campuses, with an AI commons, faculty tools and workforce apprenticeships
Where
US, North America
Segment
public
Problem
uneven access to AI and unprepared graduates
Technology
ChatGPT Edu (OpenAI)
Who uses it
California State University (23 campuses), United States
Outcome
More than 460,000 students and over 63,000 staff and faculty across 23 campuses given access (Feb 2025)
Effort
high
Risks
Academic integrity; privacy; cost under budget pressure.
Next: CSU said other AI technologies would follow and set up an AI Workforce Acceleration Board to run AI challenges source
How to adopt
Deploy systemwide with training; give faculty course-specific assistants; set an AI board to steer challenges.
How a room runs it
the course room: the instructor, the TAs, the students and a tutoring teammate; the agent drafts the study guide and GENERATES the practice tasks per student with deadlines, and each grade-affecting use becomes a proposal in the instructor's Inbox; the receipt records the draft, the sources and the instructor's changes
The engagement on top
connectors to the learning-management system and library, an instructor teammate, a course-GPT template set
A distance university moves student service from phone and ticketing to an AI agent over WhatsApp, chat, Instagram and TikTok
Where
ID, Southeast Asia
Segment
public
Problem
500 to 600 calls a day, many unanswered, tickets missing a 3-day SLA, no overnight cover for 760,000 students
Technology
AI agent, omnichannel chat, WhatsApp Business API (Qiscus AgentLabs)
Who uses it
Universitas Terbuka, Indonesia
Outcome
900,000+ AI-handled conversations in one month (vendor's published case study) the company’s own figure
Effort
medium
Risks
Wrong answers on grades or fees; vendor lock-in; privacy of student records; figure comes from the vendor.
How to adopt
Prerequisite: a knowledge base of student procedures and a messaging channel account (WhatsApp Business). First step: put an AI agent on the busiest channel and route unresolved cases to staff. Running: keep the knowledge base current and monitor wrong answers.
How a room runs it
the ROOM for a student-service desk: the head of student affairs, two service officers and a registrar teammate (an agent definition with the registration and exam-administration skills); the agent answers from the university's own rules and GENERATES tasks for the officers when a case needs a person (re-issue transcript for student X by Tuesday) and a weekly list of recurring questions for the head as a proposal to change the FAQ, with a deadline; what the humans keep: every exception, every grade dispute, every refund; budgets degrade before they block: a spent envelope lowers the agent to drafting for officers instead of replying; the receipt records each handed-over case and its outcome
The engagement on top
a custom engagement: the university's own academic information system and ticketing wired through connectors loaded on this pull; WhatsApp is built; a student-affairs administrator as the paid professional on the match; the education pack's service-desk task templates
Source: vendor · 2026 · vendor only; to corroborate
AI teaching assistant for teachers (lesson plans, rubrics, groupings), free to teachers via a donated cloud, expanded to more than 180 countries
Where
US, global
Segment
nonprofit
Problem
teacher preparation consuming more than half of the workload
Technology
Khanmigo for Teachers (Azure OpenAI Service)
Who uses it
Khan Academy (nonprofit), United States, with Microsoft
Outcome
Available to teachers in more than 180 countries and territories, with 20+ teacher activities and 30+ experimental languages (Khan Academy, Dec 2025) the company’s own figure
Effort
low
Risks
Inaccurate content; student data; unequal access by country.
Next: 30+ experimental languages were added in December 2025 beyond English source
How to adopt
Start with teacher-facing tools; keep teacher review of anything students see; expand by language.
How a room runs it
the staff-room: the teacher, the co-teacher, the parent liaison and a planning teammate; the agent drafts the lesson, rubric and groupings and GENERATES the prep tasks per class with due dates, and each student-facing document stays behind the teacher's review; the receipt records the draft, the standard it targets and the teacher's changes
The engagement on top
connectors to the school's curriculum and roster, a teacher-planning teammate, a rubric template set
Open call where faculty, staff and students propose generative-AI projects; licensed ChatGPT Edu to hundreds of funded projects
Where
US, North America
Segment
public
Problem
finding which AI uses actually help teaching, research and work
Technology
ChatGPT Edu (OpenAI)
Who uses it
Arizona State University, United States
Outcome
Over 200 projects activated: nearly 175 proposals in spring 2024 with 105 licensed, and 200 proposals in summer 2024 with 114 approved the company’s own figure
Effort
medium
Risks
Uneven quality; data privacy; student integrity.
Next: a fall 2024 call for proposals stayed open through 6 September 2024 with licences starting 19 August and 16 September source
How to adopt
Run a funded challenge in rounds; license winning projects; extend what works.
How a room runs it
the innovation room: the faculty lead, the research staff, the students and an advisor teammate; the agent drafts the proposal review and GENERATES the project tasks per team with dates, and each licence award becomes a proposal in the committee's Inbox; the receipt records the project, the data used and the outcome
The engagement on top
connectors to the university identity and learning systems, a project-review teammate, a proposal template set
A university links a chatbot to its WhatsApp number so applicants and students get round-the-clock answers on applications, registration, fees, timetables and results, with a call-centre agent in office hours
Where
ZA, Africa
Segment
public
Problem
applicant and student enquiries beyond call-centre hours
Technology
WhatsApp chatbot connected to student records (MoUJi)
Who uses it
University of Johannesburg, South Africa
Outcome
none stated
Effort
medium
Risks
Exposing personal records over a messaging channel; wrong results or fee answers; no published numbers.
How to adopt
Prerequisite: records the bot may read after verifying the student, and a WhatsApp business number. First step: answer the top applicant questions and hand the rest to the call centre in hours. Running: review wrong answers and the handover rate.
How a room runs it
the ROOM for a student-service desk: the head of student affairs, call-centre agents and a registrations teammate; the agent answers from the university's own rules and GENERATES tasks for the agents when a case needs a person (fix the fee statement for student X by Tuesday) and a monthly list of unanswered questions for the head as a proposal to change the FAQ, with a deadline; what the humans keep: every exception, every results dispute, every fee waiver; the receipt records each handed-over case and its outcome
The engagement on top
a custom engagement: the university's student information system and call-centre queue wired through connectors loaded on this pull; WhatsApp is built; a student-affairs administrator as the paid professional on the match; the education pack's service-desk task templates
Source: vendor · 2024 · vendor only; to corroborate
A university puts a chatbot on its admissions line so applicants get adjusted-score calculations and answers instantly on results day, cutting the phone queue
Where
AU, Oceania
Segment
public
Problem
phone lines swamped on the day Year 12 results are released
Technology
web chatbot with score calculator
Who uses it
University of Adelaide, Australia
Outcome
Calls fell 47 % in the first three hours and 40 % overall on results day; over 1,000 calls handled (director's figures, as reported)
Effort
low
Risks
Wrong score calculation; one peak-day figure; the source is old (2018).
How to adopt
Prerequisite: published admission rules in a form a bot can apply. First step: a chatbot for the single highest-volume question on peak day. Running: compare call counts with the previous year and staff the rest.
How a room runs it
the ROOM for a results-day desk: the admissions director, the phone team and an enquiries teammate; the agent answers score and eligibility questions from the published rules and GENERATES tasks (call back applicants whose score sits on the cut-off, by 5 pm) with owners and due dates, and a proposal to the director on staffing the line for next year's peak, with a decision date; what the humans keep: every offer, every bonus-points ruling, calls to upset applicants; the receipt records each handed-over case and its outcome
The engagement on top
a custom engagement: the university's applicant records and published admission rules wired through connectors loaded on this pull; an admissions professional on the match; the enrolment task templates in the education pack
A university runs a Facebook Messenger chatbot for clearing, taking prospective students from enquiry to application and offer, later adding voice
Where
GB, Europe
Segment
public
Problem
a short, peak clearing window with thousands of enquiries and too few staff
Technology
Messenger chatbot, later voice assistant (Becky)
Who uses it
Leeds Beckett University, United Kingdom
Outcome
21,000 conversations with prospective students, 200 secured places (as reported by Civil Service World)
Effort
medium
Risks
Wrong place or offer information; the source is old (2019); applicant data protection.
How to adopt
Prerequisite: a current list of course places and entry rules, and a channel applicants use. First step: a bot for the clearing period with staff handover. Running: update places daily and review conversations that ended without an offer.
How a room runs it
the ROOM for a clearing desk: the admissions manager, the course tutors and a clearing teammate; the agent talks to applicants on the channel and GENERATES tasks (call applicant X about the physics offer today, confirm the grade with UCAS by noon) with owners and due dates, and a proposal to the manager to open or close places on a course, with a deadline; what the humans keep: every offer, every exception, the hard conversations; the receipt records each conversation, offer and who approved it
The engagement on top
a custom engagement: the university's applicant system and course-places list wired through connectors loaded on this pull; WhatsApp or the applicant's channel of choice (built); an admissions professional on the match; the enrolment task templates in the education pack
A course business sells corporate programmes with AI learning tools (AI companion, AI role-play) and AI agents for support and sales outreach to named Indonesian companies
Where
ID, Southeast Asia
Segment
SMB and mid
Problem
corporate language and skills training without coaching at scale; support and outreach staffing
Technology
AI companion, AI role-play, AI agents, workflow automation (Cakap corporate programmes)
Who uses it
Cakap (Indonesia); testimonial clients listed: Starbucks Indonesia, Danone Indonesia, GoTo, Kementerian Ketenagakerjaan, Kalbe Farma, PFN
Outcome
Platform-wide: 6,000,000+ participants, 2,000+ partners, 1,600+ courses (Cakap's own page, unaudited, no per-client result) the company’s own figure
Effort
low
Risks
The source describes offerings, not measured adoption; learner data privacy; quality of AI coaching.
How to adopt
Prerequisite: a catalogue of programmes and a customer list. First step: add an AI companion or role-play practice to one corporate programme, and an AI support agent on enquiries. Running: content upkeep and review of agent answers.
How a room runs it
the ROOM for a training provider's client account: the account manager, the trainer, the client's HR lead and a programme-coordination teammate; the agent GENERATES the cohort's schedule, reminders and attendance follow-ups as tasks with owners and due dates, and a proposal to the HR lead at mid-programme (extend the cohort, swap a module) with a deadline; what the humans keep: the teaching, the assessment, the contract and the spend above the envelope; the receipt records sessions held, who attended and who signed
The engagement on top
a custom engagement: the provider's own LMS and booking system plus the client's HRIS wired through connectors loaded on this pull; WhatsApp reminders (built); a learning-and-development professional on the match; the coordination templates in the education pack
Source: vendor · 2026 · vendor only; to corroborate
A multi-campus private university uses generative AI to summarise 10,000+ diploma supplements a year, help staff with administrative documents and allocate classrooms and staff from predicted intake
Where
ID, Southeast Asia
Segment
enterprise
Problem
administrative document load and capacity planning across 14 campuses in 7 cities
Technology
Azure OpenAI Service, Azure Machine Learning (intake prediction reported up to 90 % accurate)
Who uses it
BINUS University, Indonesia, with Microsoft
Outcome
Intake prediction up to 90 % accurate; 10,000+ diploma supplements a year summarised (the university's own claim via Microsoft) the company’s own figure
Effort
medium
Risks
Forecast error leading to wrong room or staffing allocation; generative summaries of official documents need human check; student data in a cloud service.
How to adopt
Prerequisite: historical enrolment and student-document data in a cloud environment. First step: pilot summarisation of one document type and an intake forecast for one campus. Running: data-science and IT skills and review of the summaries.
How a room runs it
the ROOM for the registrar's office: the registrar, the faculty administrators, the dean and an academic-administration teammate; the agent drafts each diploma-supplement summary and GENERATES the review tasks for faculty administrators with due dates before the graduation date; the dean's decision on room and staff allocation for next term lands in the Inbox as a proposal with a deadline; what the humans keep: the sign-off on every document, grades and every student judgement; the receipt records which summary was approved and by whom
The engagement on top
a custom engagement: the university's student records system and timetable wired through connectors loaded on this pull; a registrar-side administrator as the paid professional on the match; the education pack's graduation-paperwork templates
Free ChatGPT Edu to all staff and students after a year-long pilot, with training and an AI ambassador network
Where
GB, Europe
Segment
public
Problem
unequal access and uneven AI skills
Technology
ChatGPT Edu (GPT-5, OpenAI)
Who uses it
University of Oxford, United Kingdom
Outcome
Pilot with 750 staff before access for all staff and students in 2025/26; more than 100 AI Ambassadors (university, Sep 2025) the company’s own figure
Effort
medium
Risks
Academic integrity; data protection; dependence on one vendor.
Next: the AI Competency Centre and a network of over 100 AI Ambassadors support the 2025/26 rollout source
How to adopt
Pilot with staff first; train; appoint ambassadors; roll out to everyone with data kept inside the university.
How a room runs it
the tutorial room: the tutor, the students, the library and a study teammate; the agent drafts the reading summary and GENERATES the practice tasks per student, and each assessed piece keeps the tutor's review step; the receipt records the help given, the sources and the tutor's marks
The engagement on top
connectors to the college systems and library, a tutor teammate, a department training track
AI coach inside an online learning platform for learners and a Socratic coach that instructors can configure
Where
US, global
Segment
enterprise
Problem
learners stuck without timely help and instructors short of time
Technology
Coursera Coach (Google Gemini as first model for instructor tools)
Who uses it
Coursera, United States
Outcome
Over 1 million learners used Coach; first-attempt quiz pass rate 9.5 percent higher and 11.6 percent more lessons completed per hour (Coursera, Sep 2024) the company’s own figure
Next: career guidance by year-end and audio and further interaction types in the coming months source
How to adopt
Launch the learner coach first; measure pass rates; then give instructors configurable activities.
How a room runs it
the learning room: the learner, the instructor, the mentor and a coaching teammate; the agent drafts hints and GENERATES the practice tasks per learner with deadlines, and each graded item keeps the instructor's rubric and review; the receipt records the hint, the attempt and the pass
The engagement on top
connectors to the course content and quiz engine, an instructor teammate, a Socratic-activity template set
A university texts every admitted student through an AI assistant that sends reminders, answers questions (tuition due dates, forms) and hands only the rest to staff, tested in a randomised trial
Where
US, North America
Segment
public
Problem
admitted students who never enrol in the autumn (summer melt)
Technology
conversational AI over SMS (AdmitHub / Pounce)
Who uses it
Georgia State University, USA
Outcome
Randomised trial: 21.4 % lower summer melt and 3.9 % higher enrolment than the control group (study by two university researchers, as reported)
Effort
medium
Risks
Wrong answers on deadlines or aid; student consent to texting; the source is old (2017).
How to adopt
Prerequisite: admitted-student contact data and a list of the admission tasks that block enrolment. First step: text-message AI assistant answering deadline and form questions with staff handover. Running: keep answers current and have staff cover the hand-offs.
How a room runs it
the ROOM for admissions: the admissions director, the counsellors and an enrolment teammate; the agent works the list of admitted students and GENERATES tasks (call the 12 students who have not paid the deposit by the 1st) with owners and due dates, and a proposal to the director to extend a deadline for a group, with a decision date; what the humans keep: every admission decision, every fee waiver, the hard conversations; the receipt records which student was reached, how, and the outcome
The engagement on top
a custom engagement: the university's applicant system and fee records wired through connectors loaded on this pull; WhatsApp in place of SMS (built); an admissions professional on the match; the enrolment task templates in the education pack
A labour ministry signed an MoU (5 May 2026) with a telecom operator and a foundation to integrate employment service platforms, including its national employment portal (SIAPKerja), using digital and AI technology
Where
ID, Southeast Asia
Segment
public
Problem
vocational training that is not tied to a worker's goals or to local industry demand
Technology
Wadhwani JobReady and Genie AI platforms, LMS, integrated into SIAPKerja; connectivity from the operator
Who uses it
Kementerian Ketenagakerjaan (Kemnaker), Wadhwani Foundation, Indosat Ooredoo Hutchison, Indonesia
Outcome
none stated
Effort
high
Risks
An MoU is intent, not use; low user take-up; career advice that ignores local labour demand.
How to adopt
Prerequisite: a national employment portal and a network of training centres. First step: an agreement with a platform provider and a telecom partner, then a pilot in a few centres. Running: connectivity, content localisation and tracking of placements.
How a room runs it
the ROOM for one training centre: the centre head, the instructors, the placement officer and a training-administration teammate; the agent GENERATES the cohort's tasks (confirm trainee attendance sheets by Monday, book the assessor for the 12th) with owners and due dates and puts a proposal to the head to open or close a class with a deadline; what the humans keep: assessment and certification, the workshop floor, the spend; the receipt records each enrolment, attendance and certificate issued
The engagement on top
a custom engagement: the centre's own registration and attendance systems wired through connectors loaded on this pull; WhatsApp for trainees (built); a vocational training manager on the match; the training-centre task templates in the education pack
Your own team builds it. Across the 13 adoptions here, 3 were low effort, 8 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: hazard and exposure datasets for the territory. First step: publish a GIS portal of risk by location for planners. Running: a data team to keep the layers current and agreements with data owners.
InaRISK: BNPB's internet GIS portal of disaster risk for all of Indonesia, used by government, local governments and the private sector, with its data already used for regional spatial plans; risk data is the first step towards a multi-hazard early warning system
Where
ID, Southeast Asia
Segment
public
Problem
knowing hazard exposure by place to plan before a disaster
Technology
GIS server, dynamic risk data
Who uses it
BNPB (national disaster agency), Ministry of Agrarian Affairs and Spatial Planning, local governments
Outcome
none stated
Effort
high
Risks
Out-of-date data gives false confidence; the page describes the portal, not an AI use.
How to adopt
Prerequisite: hazard and exposure datasets for the territory. First step: publish a GIS portal of risk by location for planners. Running: a data team to keep the layers current and agreements with data owners.
How a room runs it
the ROOM for a BPBD: the BPBD head, the preparedness officer and a planning teammate; the agent reads the risk layers the agency already uses and GENERATES the preparedness plan's steps for each office (drill the school in zone X, check the shelter list by the month end) and the budget request as a decision in the head's Inbox; the head keeps every public decision (evacuation, a warning issued), the agent only prepares; the receipt records each step done
The engagement on top
connectors loaded on this pull to the agency's risk layers and shelter lists; a disaster-management adviser as paid professional; the pack carries drill and preparedness task templates
Seattle Fire Department's 911 medical calls are monitored by Corti AI, which since December 2023 prompts dispatchers when a call may fit the nurse line instead of an ambulance; the dispatcher keeps the decision; critics say it ran without disclosure to callers or public review
Where
US, North America
Segment
public (city fire and 911)
Problem
dispatcher load and unneeded ambulance responses
Technology
speech and call analysis AI, live dispatcher prompts
Who uses it
Seattle Fire Department, Corti (Denmark)
Outcome
Corti's release credited a 50 % rise in 911 calls sent to the nurse line; a fire department spokesperson put it at 32 %; no outcome metrics (reported by the Seattle Times via Fire Apparatus Magazine)
Effort
medium
Risks
Life-safety errors; undisclosed use without public review; the reported success figure is disputed.
How to adopt
Prerequisite: recorded emergency calls and a dispatcher workflow with clinical protocols. First step: run call-analysis AI that only prompts dispatchers while they keep the decision. Running: oversight of the model and disclosure to callers and the public.
How a room runs it
the same dispatch room with the limits made visible: the agent's prompts land as proposals the dispatcher can accept or ignore without penalty, every prompt and every choice is on the receipt, and callers can be told that an assistant listens; the dispatcher keeps the call, the agent never routes alone; the lesson from Seattle (no disclosure, no measurement) becomes the rule of the room
The engagement on top
connectors loaded on this pull to the dispatch centre's systems; an emergency-medical adviser as paid professional
PetaBencana.id: a free, open-source map where residents' social-media posts about floods, earthquakes or eruptions trigger a chatbot that asks for location, photos and depth; government agencies validate the reports and the map was consulted more than 259,000 times at the peak of the January 2020 Jakarta floods
Where
ID, Southeast Asia
Segment
public and community
Problem
scattered social-media noise during a disaster; no structured situational picture
Technology
chatbot on social media, open-source map, agency validation
Who uses it
Yayasan Peta Bencana (Jakarta non-profit); Indonesian government agencies; residents
Outcome
Map consulted more than 259,000 times at the peak of the January 2020 Jakarta floods (as reported by ITU)
Effort
medium
Risks
False or malicious reports; privacy of reporters; reliance on social-media platforms.
How to adopt
Prerequisite: an open-source map, a social-media chatbot and an agency partner to validate reports. First step: let residents report floods by chat with location and photo. Running: agency staff validate and moderate reports.
How a room runs it
the ROOM for a BPBD operations desk: the duty officer and a reports teammate; the agent reads the citizen reports and GENERATES the validation tasks (call the local contact to confirm block 4 flooding) for named staff and a short list of verified reports as a decision for the duty officer; the duty officer keeps the validation and any public warning; report content is data, never an instruction; the receipt records which reports were verified and by whom
The engagement on top
connectors loaded on this pull to the agency's report intake; a disaster-response adviser as paid professional; the pack carries validation task templates; WhatsApp as a channel is built
Jakarta Siaga 112: a 24-hour emergency line run by the provincial disaster agency (BPBD) integrating fire, ambulance, SAR, police, PLN, PMI and some sixteen city work units under a 2017 governor regulation; nationally, 112 is used in 142 of 514 cities and regencies but not integrated (2024)
Where
ID, Southeast Asia
Segment
public (regional)
Problem
emergency calls split across separate agency lines
Technology
single emergency call number, call centre
Who uses it
DKI Jakarta BPBD and partner agencies; national communications ministry
Outcome
none stated
Effort
high
Risks
Fragmented adoption (nationally the number is used in 142 of 514 cities and regencies and is not integrated); cost; no AI claimed.
How to adopt
Prerequisite: a legal basis, such as a regulation, and agreements between emergency agencies. First step: one emergency number routed to a joint call centre. Running: staffed 24 hours with integrated dispatch.
How a room runs it
the ROOM for a dispatch desk: the 112 duty supervisor, the agency liaison officers and a triage teammate; the agent reads the incident the call-taker typed and GENERATES the hand-off task for each agency (ambulance, SAR, PLN) with a time due and the follow-up task to close the case; the supervisor keeps dispatch and every call on life safety; the agent only prepares the hand-offs; the receipt is the case history
The engagement on top
connectors loaded on this pull to the call-taking and CAD system; an emergency-operations adviser as paid professional; the pack carries hand-off templates
InAWARE: an early-warning and decision-support system on the DisasterAWARE platform deployed at BNPB to share information among national and provincial agencies; a three-year USD 2.2 million project with training and exercises
Where
ID, Southeast Asia
Segment
public (national and provincial)
Problem
slow information sharing between agencies during a disaster
Technology
decision-support platform, ICT tools, training
Who uses it
BNPB with selected BPBDs, BMKG, TNI, BASARNAS; funded by USAID
Outcome
none stated
Effort
high
Risks
The source is a 2013 project and may not reflect current use; cost; inter-agency data-sharing gaps.
How to adopt
Prerequisite: agencies that agree to share information and a decision-support platform. First step: deploy the platform at the national centre and link provincial agencies. Running: training, exercises and sustained funding.
How a room runs it
the ROOM for an incident: the BPBD commander, the field coordinators and an incident teammate; the agent reads the incident feed and GENERATES the shift's tasks (confirm shelter capacity, send the situation report by 18:00) for named coordinators and files each evacuation or resource request as a decision with a deadline in the commander's Inbox; the commander keeps every warning and order, the agent never issues one; the receipt is the incident log
The engagement on top
connectors loaded on this pull to the agency's incident and logistics systems; a disaster-response adviser as paid professional; the pack carries incident task templates; channel: WhatsApp for field staff
Copenhagen's emergency medical dispatch uses Corti, an AI assistant that listens to 112 calls, transcribes them and alerts the dispatcher in real time when the words and background cues (such as breathing) suggest cardiac arrest; the dispatcher stays on the call
Where
DK, Europe
Segment
public (city emergency medical services)
Problem
cardiac arrest is hard to recognise by phone, and delay costs lives
Technology
speech recognition, machine-learning classification of call content and background sound
Who uses it
Copenhagen Emergency Medical Services, Corti (Denmark)
Outcome
Randomised trial over 169,049 calls (2018-09 to 2019-12): the model's sensitivity was 85.0 % vs 77.5 % for dispatchers, but dispatchers alerted by it recognised cardiac arrest in 93.1 % of confirmed cases vs 90.5 % without (P = .15, not significant); the model's positive predictive value was 17.8 % vs 55.8 % (the trial's authors)
Effort
high
Risks
The flag rate and false alarms were not given on the page; later reporting disputes the benefit; alert fatigue; the dispatcher's judgement must stay primary.
How to adopt
Needs a call recording archive with outcomes to train and test on; first step is a silent shadow run beside the dispatchers, measuring flags against confirmed cases; takes clinical governance and integration with the call system.
How a room runs it
the ROOM for an emergency medical dispatch room: the dispatcher, the medical director and a call-support teammate; the teammate GENERATES the in-call prompts and, after the call, the review tasks for the medical director (calls flagged and missed, to be listened to this week) as plan steps with owners and due dates; a decision block in the medical director's Inbox when the flag rate or the miss rate moves, with a deadline to retune the thresholds; what the humans keep: the dispatch decision and the instructions to the caller; the receipt records every finished task and every decision with who took it
The engagement on top
a custom engagement: the dispatch centre's own call system and call archive stay theirs, wired to the room through connectors loaded on this pull, the clinical-safety reviewer joins as the paid professional on the match, and the sector pack carries the audit-of-missed-calls review
Japan's Fire and Disaster Management Agency plans a fiscal-2027 model project in which AI transcribes 119 emergency calls, advises call-takers on what to confirm (address, whether the patient is conscious) and extends to the #7119 ambulance-advice line; call-takers keep the decision
Where
JP, East Asia
Segment
public (national fire and disaster agency)
Problem
call-taking workload grows with ambulance dispatches; recording call details takes staff time
Technology
speech recognition, vertical AI advice to staff, hospital information lookup
Who uses it
Japan Fire and Disaster Management Agency; local fire headquarters (to be selected)
Outcome
none stated (a planned model project; the article reports no costs or outcomes)
Effort
high
Risks
Transcription errors on dialect and noisy calls; over-reliance by call-takers; no results reported yet; the plan may change with the budget.
How to adopt
Needs recorded calls and a call-taking protocol the AI can follow; first step is a model project in one control centre with the human call-taker deciding; takes procurement, integration with the dispatch system and staff training.
How a room runs it
the ROOM for an emergency control centre: the duty officer, the call-takers and a call-handling teammate (an agent definition with the dispatch skills); the teammate transcribes each call and GENERATES the follow-up tasks for the call-taker (confirm address, ask consciousness) as plan steps with owners and due times, and the shift lead's staffing notes as tasks for the week; a decision block lands in the shift lead's Inbox when a call may fit the advice line instead of an ambulance, with a deadline and the transcript attached; what the humans keep: the dispatch decision, the call itself and the sign-off; the receipt records every finished task and every decision with who took it
The engagement on top
a custom engagement: the centre's own call-taking and dispatch systems stay theirs, wired to the room through connectors loaded on this pull (the call recorder, the dispatch log), the emergency-operations adviser joins as the paid professional on the match, and the sector pack's transcript and advice-line templates are the starting point
Your own team builds it. Across the 7 adoptions here, 0 were low effort, 2 medium and 5 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: a village information system with citizen records. First step: prototype a chatbot that drafts letters and answers common queries. Running: officers validate citizen data and review output.
Pair: a secure AI assistant suite for Singapore's public officers (chat, email-based, search of public records, minutes from audio); more than half of about 150,000 officers use it regularly, and GovTech is building an assistant desk with a registry of who owns each agent and security limits on what agents may do
Where
SG, Southeast Asia
Segment
public (national)
Problem
civil-service drafting, analysis and minute-taking load; need for control over agents in government
Technology
LLM assistants, custom assistants for repeated tasks, agent registry, guardrails
Who uses it
Singapore GovTech / Open Government Products; agencies such as BCA and Singapore General Hospital
Outcome
More than half of about 150,000 public officers use it regularly (as reported by HCA magazine)
Effort
high
Risks
Data leakage of official information; over-reliance on drafts; uncontrolled agents without registry; reference is not local.
How to adopt
Prerequisite: a secure government cloud and an approved set of AI models. First step: roll out chat and drafting assistants to officers with security limits. Running: an agent registry, guardrails and a platform team.
How a room runs it
the ROOM for a department: the head of unit, the officers and a drafting teammate whose definition lists the department's templates; the agent GENERATES the first draft of a memo or the minutes of a meeting as a task for the owning officer (due before the next meeting) and the finished paper as a proposal in the head's Inbox; the agent has a named owner and a limit on what it may send out, as in the registry above; the officer keeps the sign-off, the receipt records who drafted, who changed and who approved
The engagement on top
connectors loaded on this pull to the department's document store and mailbox; a public-administration adviser as the paid professional; the pack carries memo and minutes task templates
One online gateway for 240+ public services (civil registration, driving permits, business services) with digital payment, so most transactions are completed without visiting an office
Where
RW, Africa
Segment
national government
Problem
citizens queue at government offices for routine services
Technology
national digital service platform, e-wallet and digital payments, agent access points
Who uses it
IremboGov, Government of Rwanda
Outcome
About 80% of transactions fully online; nearly 500,000 requests a month; some service times cut from days to two hours; over 1.5 billion working hours saved (as reported; measurer not stated)
Effort
high
Risks
Figures unattributed; the digital divide; dependence on payment rails; service-time claims unaudited.
How to adopt
Start with the highest-volume services, put application, payment and approval on one flow, and add assisted access points for people offline.
How a room runs it
the ROOM for a ministry's service-delivery team: the service owner, the platform lead, a front-line supervisor and a service-queue teammate (an agent definition with the service catalogue); the agent teammate reads the request and completion data the platform already holds and GENERATES the supervisors' tasks (clear the backlog on permit applications over 5 days by Friday) as plan steps with owners and due dates, and each service-redesign proposal for the service owner (change / hold, with the processing times attached) as a decision block in the Inbox with a deadline; what the humans keep: the approvals, the fees and the policy; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the service catalogue, the request-and-payment logs, the agent-access-point roster), a digital-government specialist joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
OpenSID: an open-source, community-driven village information system for population data, letters and certificates, and public publication; installed by more than 10,000 villages and actively used by more than 4,600
Where
ID, Southeast Asia
Segment
public (villages)
Problem
paper-based village administration; weak transparency and service speed
Technology
open-source web village information system
Who uses it
Indonesian village governments; project community on Open Collective
Outcome
Installed by 10,000+ villages and actively used by 4,600+ (the project's own counts) the company’s own figure
Effort
medium
Risks
Install counts are old and the project's own; data quality; weak maintenance in small villages.
How to adopt
Prerequisite: village staff with a computer and population data. First step: install the open-source village information system and move letters and certificates onto it. Running: local IT support and data upkeep.
How a room runs it
the same village room built on what the village already runs: the agent reads the OpenSID population and letter records and GENERATES the clerk's queue (letters to prepare, data gaps to fix) with due dates, and files the village head's monthly publication for approval as a decision with a deadline; the village head keeps the sign-off, residents keep control of their own data, clearance masks personal fields before the model sees them; the receipt records every letter issued
The engagement on top
connector to OpenSID (database or export) loaded on this pull; a village-administration adviser as paid professional; pack templates for the letter types
Source: vendor · 2021 · vendor only; to corroborate
City-wide ChatGPT for office work after a one-month trial with about 4,000 staff in Yokosuka, then a metropolitan rollout and an in-house generative AI platform
Where
JP, East Asia
Segment
public
Problem
routine drafting and document work under labour shortage
Technology
ChatGPT via LoGo chat; Tokyo's open-source Generative AI Platform with GovTech Tokyo
Who uses it
Tokyo Metropolitan Government and Yokosuka City, Japan
Outcome
Yokosuka trial about 4,000 staff (June 2023); Tokyo platform built on open-source software; Sumida ward's ordinance-search app reported cutting time and cost against outsourcing (no figure)
Effort
medium
Risks
Accuracy complaints; confidential data; uneven skills across municipalities.
Next: the platform will be deployed across Tokyo's cities, wards, towns and villages so local staff can build their own tools source
How to adopt
Trial in one city; ban personal data in prompts; then give departments a no-code platform with shared apps.
How a room runs it
the municipal room: the department head, the clerks, the legal reviewer and a drafting teammate; the agent drafts notices and manual searches and GENERATES the review tasks per clerk, and each outgoing text becomes a proposal in the head's Inbox; the receipt records the draft, the reviewer and the personal-data check the city requires
The engagement on top
connectors to the ordinance database and the chat system, a municipal teammate, a no-code app template set
AI assistant inside the national citizen-services app, after the contact centre began resolving inquiries conversationally
Where
UA, Europe
Segment
public
Problem
high volume of routine citizen inquiries to the service desk
Technology
Diia.AI, a national LLM-based assistant
Who uses it
Ministry of Digital Transformation, Ukraine
Outcome
52 percent of inquiries at the Diia contact centre resolved by AI of 53,500 requests in June 2025, the rest handled by human operators (Deputy PM Fedorov)
Effort
high
Risks
Wrong entitlement advice; identity data; trust in a wartime service.
Next: consultations will open to all users and the ministry is developing an AI-agent service for delivering state services source
How to adopt
Beta on the portal; measure resolution share; keep human operators for the rest; add services one by one.
How a room runs it
the citizen-service room: the service officer, the case worker, the data-protection lead and an assistant teammate; the agent answers and GENERATES the follow-up tasks per office with due dates, and each document or entitlement request becomes a proposal in the officer's Inbox; the receipt records the question, the answer and the handoff
The engagement on top
connectors to the registries and the contact-centre tool, a public-service teammate, a handover-to-human playbook
Japan's Digital Agency built Gennai, a generative-AI environment for civil servants with chat, drafting, summarising, translation, a Diet-answer search and legal research support; open-sourced in April 2026 and in a government-wide pilot for about 180,000 employees from May 2026, with full use planned from fiscal 2027
Where
JP, East Asia
Segment
public (national, all ministries)
Problem
administrative work volume; staff time on drafting and analysis
Technology
in-house generative-AI platform, retrieval over administrative documents, open source
Who uses it
Japan Digital Agency; ministries and agencies
Outcome
A Ministry of Agriculture, Forestry and Fisheries analysis of about 8,000 survey responses and 20+ hypotheses took about three days instead of about two months for one employee (reported by the Digital Agency)
Effort
high
Risks
Confidentiality limit on what may be entered; results are the agency's own; reliance on drafts.
How to adopt
Needs a secure environment cleared for the data level, a procurement and use guideline, and an AI officer for governance; first step is one department's drafting and analysis tasks; takes training and ministry funding.
How a room runs it
the ROOM for a ministry bureau: the bureau chief, the officers and a drafting-and-analysis teammate; the teammate GENERATES the first drafts and the analysis steps for the officers (survey responses to code, answers to draft) as plan steps with owners and due dates; a decision block in the bureau chief's Inbox for each answer or analysis that leaves the bureau, with a deadline; what the humans keep: the content of the answer, the confidentiality call and the sign-off; the receipt records every finished task and every decision with who took it
The engagement on top
a custom engagement: the ministry's own documents and systems stay theirs, wired to the room through connectors loaded on this pull (within the confidentiality level), the public-sector AI governance adviser joins as the paid professional on the match
A village administration prototype with a generative-AI model and an NLP chatbot that automates letter and document creation, validates citizen data and answers residents; functional tests reached a 95 % pass rate and usability scored 87.0 (SUS)
Where
ID, Southeast Asia
Segment
public (village)
Problem
manual, slow village administration; officers overloaded with letters and citizen queries
Technology
generative AI model, NLP chatbot, village information system
Who uses it
Popalia Village, Indonesia (research prototype, not a confirmed live service)
Outcome
95 % functional-test pass rate and usability score 87.0 (SUS) in the authors' own tests
Effort
medium
Risks
Prototype only, not a deployment audit; wrong data in official letters; privacy of citizen records.
How to adopt
Prerequisite: a village information system with citizen records. First step: prototype a chatbot that drafts letters and answers common queries. Running: officers validate citizen data and review output.
How a room runs it
the ROOM for a village office: the village head (kepala desa), the village secretary and a village-services teammate (an agent definition with the letter templates and the citizen register the office already keeps); a resident's request arrives over WhatsApp, the agent GENERATES the draft letter and the checklist of missing documents as plan steps for the clerk (due the next working day) and the letter for signature as a decision in the village head's Inbox; the village head keeps the signature and the stamp, the clerk keeps the check of the original papers; the receipt records who requested, who checked, who signed and when
The engagement on top
a custom engagement: the village's own register (OpenSID or the regency's system) wired to the room through connectors loaded on this pull, a village-administration adviser joins as the paid professional on the match, and the pack carries the letter task templates in Bahasa; channel: WhatsApp
Pennsylvania's one-year generative-AI pilot: 175 employees in 14 agencies used a ChatGPT Enterprise licence to draft, brainstorm and simplify information; participants reported saving an average of 95 minutes on days they used it, and over 85 % rated the experience positive
Where
US, North America
Segment
public (state government)
Problem
routine drafting and information work consumes staff time
Technology
generative AI chat assistant, staff training
Who uses it
Pennsylvania Office of Administration, OpenAI, Carnegie Mellon University, SEIU Local 668
Outcome
Average 95 minutes saved on days the tool was used and over 85 % rated the experience positive, from participants' self-reports (reported by the Pennsylvania Office of Administration); the licences, training and support cost USD 108,000 (per press coverage)
Effort
low
Risks
Self-reported savings; data-entry rules; accuracy of drafts; workforce concerns, which the union raised.
How to adopt
Needs a licence, a use policy and training; first step is a cohort of volunteers across several agencies with a survey at the end; takes a rule on what data may be entered and union consultation.
How a room runs it
the ROOM for an agency team of staff using an assistant: the team lead, the staff and a drafting teammate; the teammate GENERATES draft documents and the review tasks for the owner of each as plan steps with owners and due dates; a decision block in the director's Inbox on which tasks the assistant may draft, with a deadline and the sample outputs attached; what the humans keep: the final text, the policy judgement and the sign-off; the receipt records every finished task and every decision with who took it
The engagement on top
a custom engagement: the agency's own document stores and case systems stay theirs, wired to the room through connectors loaded on this pull, the public-sector change adviser joins as the paid professional on the match
Whole-of-government six-month trial of Microsoft 365 Copilot across dozens of agencies with an independent evaluation
Where
AU, Oceania
Segment
public
Problem
time spent summarising, searching and drafting
Technology
Microsoft 365 Copilot; evaluation by Nous Group
Who uses it
Australian Government, Digital Transformation Agency, more than 7,600 staff
Outcome
Over 7,600 staff, six months (H1 2024), cost over 1.2 million AUD; 69 percent faster task completion, 86 percent wanted to continue, about an hour a day saved on summarising and searching, offset in part by checking AI output (DTA evaluation)
Effort
medium
Risks
Review burden; vendor lock-in; uneven adoption.
Next: the evaluation flags specific features and use cases as the places benefits were found, guiding any wider rollout source
How to adopt
Trial with named agencies; evaluate independently; train users; match licences to the use cases that showed value.
How a room runs it
the agency room: the manager, the staff, the records officer and a drafting teammate; the agent drafts the summary or brief and GENERATES the review tasks per officer, and each document leaving the agency becomes a proposal in the manager's Inbox; the receipt records the draft, the reviewer and the changes, which answers the checking burden the trial found
The engagement on top
connectors to the agency's document and records systems, a records-officer teammate, an adoption-training track
Estonia's Bürokratt: a network of virtual assistants on public agencies' sites, launched with the Consumer Protection and Technical Regulatory Authority and then the Police and Border Guard Board and the National Library, with open-source code offered for reuse and a four-year budget of close to EUR 13 million
Where
EE, Europe
Segment
public (national government)
Problem
citizens must find the right agency and service; agency call and email load
Technology
chatbots, speech and language models, open-source code, interoperability layer
Who uses it
Estonian Ministry of Economic Affairs and Communications, State Information System Authority (RIA), Estonian Language Institute; Net Group, Texta, Stacc, Microsoft, Solita
Outcome
none stated (the article gives spend and plans, no usage or results)
Effort
medium
Risks
No usage data in the source; answers must not bind the agency without an officer; reuse across agencies depends on shared standards.
How to adopt
Needs a catalogue of the agency's services and a clean answer base; first step is one agency's most-asked questions; takes language-model support for the local language and an integration with the agency's systems.
How a room runs it
the ROOM for a public agency's service desk: the service manager, the agency's subject experts and a citizen-enquiry teammate; the teammate reads the questions citizens ask and GENERATES the week's tasks for the subject experts (answers to write, pages to fix) as plan steps with owners and due dates; a decision block in the director's Inbox when a topic is asked often enough to become a service, with a deadline; what the humans keep: the answers that bind the agency and the hand-over to an officer; the receipt records every finished task and every decision with who took it
The engagement on top
a custom engagement: the agency's own website, case system and knowledge base stay theirs, wired to the room through connectors loaded on this pull, the public-sector digital adviser joins as the paid professional on the match
Government-built AI tool that groups thousands of consultation responses into themes, first used on a live Scottish Government consultation
Where
GB, Europe
Segment
public
Problem
slow, costly manual analysis of public consultation responses
Technology
Consult, part of the Humphrey AI suite (UK DSIT)
Who uses it
UK Department for Science, Innovation and Technology with the Scottish Government
Outcome
Over 2,000 responses analysed across six questions with an F1 score of 0.76; government estimate of about 75,000 analyst days and roughly 20 million GBP a year across about 500 consultations (projection, not measured) the company’s own figure
Effort
medium
Risks
Subjective theme boundaries; bias in summaries; over-trust in a projected saving.
Next: the Technology Secretary plans to use Consult in his own department, with other Whitehall departments to follow source
How to adopt
Run beside human analysts on one live consultation; measure agreement; publish the theme dashboard; expand only after evaluation.
How a room runs it
the consultation room: the policy lead, the analysts, the minister's office and an analysis teammate; the agent drafts the theme set and GENERATES the review tasks per analyst with due dates, and each theme sign-off becomes a proposal in the policy lead's Inbox; the receipt records the themes, the human check and the changes
The engagement on top
connectors to the consultation portal and document store, a policy-analysis teammate, a theme-review template
Free shared generative-AI evaluation platform for all federal agencies, with usage dashboards, launched as a trial
Where
US, North America
Segment
public
Problem
each agency separately buying, testing and securing AI tools
Technology
USAi.gov chat, code generation and document summarisation
Who uses it
US General Services Administration and federal agencies
Outcome
Launched 14 Aug 2025 at no cost to all federal agencies as a shared services platform for comparing AI systems; no adoption figures published the company’s own figure
Effort
medium
Risks
Data classification; lock-in to a platform; unclear funding after the trial.
Next: dashboards and usage analytics will help agencies track performance and guide adoption source
How to adopt
Offer one secure shared sandbox; compare models side by side; track usage; decide per use case.
How a room runs it
the agency-evaluation room: the CIO, the program owner, the security lead and an evaluation teammate; the agent drafts the comparison and GENERATES the test tasks per team with deadlines, and each tool adoption becomes a proposal in the CIO's Inbox; the receipt records the models compared, the results and the decision
The engagement on top
connectors to the agency identity and audit systems, an evaluation teammate, a model-comparison template
AI information assistant on the Signpost service giving refugees and migrants vetted local information, piloted in several countries
Where
US, global
Segment
NGO
Problem
displaced people cannot reach enough human information officers
Technology
Signpost AI Information Assistant (LLMs from OpenAI, Anthropic and Google; Zendesk)
Who uses it
International Rescue Committee, with Anthropic, OpenAI, HPE and the World Food Programme
Outcome
Six pilot countries (Kenya, Greece, Italy, El Salvador, Mexico, Libya), 9,202 people served, 99 percent uptime, eight languages added including Dari, Pashto and Urdu (data.org impact page) the company’s own figure
Effort
high
Risks
Wrong safety or legal advice; vulnerable users; privacy.
Next: the IRC plans to expand to selected low- and middle-income countries source
How to adopt
Pilot with local partners vetting answers; escalate sensitive queries to humans; add languages gradually.
How a room runs it
the information room: the country lead, the local partner, the protection officer and an answering teammate; the agent drafts the answer from vetted sources and GENERATES the escalation tasks per partner with due times, and each sensitive reply becomes a proposal in the protection officer's Inbox; the receipt records the question, the source and the handoff
The engagement on top
connectors to the partner knowledge base and the messaging channels, a protection-officer teammate, a low-resource-language review set
Your own team builds it. Across the 13 adoptions here, 1 were low effort, 7 medium and 5 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: agreement between ministries and regions on risk classes. First step: one online single window issuing a registration number by risk level. Running: a platform team and rules upkeep.
Risk-based Online Single Submission (OSS): one online licensing service where a business gets a registration number (NIB) if low risk, a standard certificate if medium and a business permit if high
Where
ID, Southeast Asia
Segment
SMB and public (permits)
Problem
slow, duplicated licensing across ministries and regions; MSMEs treated like large firms
Indonesian government (investment coordinating bodies); MSMEs and large businesses as applicants
Outcome
none stated
Effort
high
Risks
The page describes design, not adoption figures; regional compliance gaps; system downtime.
How to adopt
Prerequisite: agreement between ministries and regions on risk classes. First step: one online single window issuing a registration number by risk level. Running: a platform team and rules upkeep.
How a room runs it
the ROOM for an applicant: the business owner, a permit teammate (an agent definition with the checklist for the permit class) and, if the owner wants one, the consultant; the agent reads what the owner already has (deeds, tax number, site plan) and GENERATES the application's task list with owners and due dates (document X from the notary by Friday) and files the final submission as a decision in the owner's Inbox; the permit itself is granted by the named official at the agency and nobody else, the agent never states a permit as granted; the receipt records each filed document and each reply
The engagement on top
connectors loaded on this pull to the applicant's own file store and mail; a licensing consultant (or the notary) as paid professional on the match; the pack carries per-permit-class task templates; channel: WhatsApp
AI checks building-permit applications against code, flags missing information and guides corrections before a planner reviews; Honolulu reports median permit wait down to 2.5 months and review cycles down from 3.4 to 1.4
Where
US, North America
Segment
public (city planning)
Problem
incomplete applications and repeated correction cycles
Technology
AI plan-review software (CivCheck and similar)
Who uses it
Honolulu, Denver, Louisville, Seattle and other US cities
Outcome
Honolulu median permit wait down to 2.5 months, review cycles from 3.4 to 1.4 (as reported by the city, via a newspaper)
Effort
medium
Risks
Wrong flags; applicant over-reliance; a foreign reference; figures as reported.
How to adopt
Prerequisite: digital permit applications and the building code in a machine-readable form. First step: AI plan-review checks before a planner reviews. Running: planners verify flags and code updates are loaded.
How a room runs it
the ROOM for a permit desk: the permit planner, the applicant and a completeness teammate; the agent GENERATES the list of missing items for the applicant as tasks with due dates and the planner's review packet as a proposal with a deadline; the decision to grant stays with the named planner and the agent only fact-checks, as the cities describe; budgets degrade before they block, so a busy week lowers the agent to flagging only; the receipt records each check and each human decision
The engagement on top
connectors loaded on this pull to the department's permit system; a permitting adviser as paid professional; the pack carries intake and correction task templates
Notaries file corporate, fiduciary and legal-entity matters through the ministry's AHU Online and handle electronic land-certificate processing; deeds are still drafted conventionally and the digital steps sit around them
Where
ID, Southeast Asia
Segment
SMB and mid (notaries, PPAT)
Problem
paper deeds and manual filing; need for electronic signatures and electronic documents with legal force
Technology
AHU Online (legal administration), electronic signature, electronic land certificates
Who uses it
Indonesian notaries and PPAT; Ministry of Law; land agency
Outcome
none stated
Effort
medium
Risks
Only an abstract was read; deeds remain paper; document authenticity.
How to adopt
Prerequisite: notaries with electronic signatures and online access. First step: file corporate matters through the online system. Running: staff training.
How a room runs it
the ROOM for a notary office: the notary, the clerks and a filings teammate; for each client matter the agent GENERATES the checklist (identity papers, tax number, draft deed) as tasks for the clerk with due dates and the AHU Online submission as a decision for the notary with a deadline; the notary keeps the authentic deed, the signature and the legal opinion, the agent drafts nothing as final; clearance keeps one client's file out of another's; the receipt records who prepared, who checked and when it was filed
The engagement on top
connectors loaded on this pull to the notary's practice-management files; a notary or paralegal as paid professional; the pack carries deed-type task templates; channel: WhatsApp for client document requests
India's Supreme Court translates its judgments with SUVAS, an AI translation tool launched in 2019 (36,271 judgments into Hindi and 17,142 into 16 other languages by 2024), and has piloted SUPACE, an assistive tool for summarising case files for judges, as part of the e-Courts programme
Where
IN, South Asia
Segment
public (national judiciary)
Problem
judgments in English are inaccessible to most litigants; case files are large
Supreme Court of India, e-Committee, Ministry of Law and Justice
Outcome
36,271 judgments translated into Hindi and 17,142 into 16 other languages by 2024 (Press Information Bureau releases, as cited by the Frontiers review)
Effort
medium
Risks
Translation errors, hence a disclaimer on the outputs; bias in training data; accountability gaps; assistive tools must not decide cases.
How to adopt
Needs a corpus of judgments and a legal glossary; first step is a translator-reviewed pilot in one language; takes a human check on every output and a publication rule.
How a room runs it
the ROOM for a court registry or a legal-aid clinic: the registrar, the translators and a translation-and-summary teammate; the teammate GENERATES the translated drafts and the checking tasks for each translator as plan steps with owners and due dates; a decision block in the registrar's Inbox on which judgments to publish and in which language, with a deadline; what the humans keep: the legal reading, the checking of every translation and the publication; the receipt records every finished task and every decision with who took it
The engagement on top
a custom engagement: the court's or the clinic's own case system and judgment archive stay theirs, wired to the room through connectors loaded on this pull, the legal translator joins as the paid professional on the match
The UK government's Consult tool sorts the free-text responses to a public consultation into themes: it analysed over 50,000 responses to the Independent Water Commission in about two hours for GBP 240, experts then checked the themes in 22 hours, and the department claims it agreed with experts nearly 83 % of the time
Where
GB, Europe
Segment
public (central government policy teams)
Problem
manual analysis of consultation responses costs about GBP 20 million and 75,000 official days a year
Technology
large language model theme analysis, human expert checking
Who uses it
UK Department for Science, Innovation and Technology (DSIT); Scottish Government (first live use)
Outcome
50,000+ responses analysed in about two hours for GBP 240; 22 hours of expert checking; agreed with one or both expert groups nearly 83 % of the time, while the two expert groups agreed with each other 55 % (all measured and reported by DSIT)
Effort
medium
Risks
Theme choice is subjective; missed minority views; savings are the department's estimate; reviewers must read the responses themselves.
How to adopt
Needs the responses in digital form and a method for checking themes; first step is to run the tool beside the analysts on one finished consultation; takes expert review time.
How a room runs it
the ROOM for a policy team running a consultation: the policy lead, the analysts and a consultation-analysis teammate; the teammate GENERATES the theme tables and the checking tasks for each analyst (themes to verify, responses to read) as plan steps with owners and due dates; a decision block in the policy lead's Inbox on the final theme set, with a deadline and the disagreement rate attached; what the humans keep: the reading of the sensitive responses, the recommendation and the sign-off; the receipt records every finished task and every decision with who took it
The engagement on top
a custom engagement: the team's own consultation platform and response exports stay theirs, wired to the room through connectors loaded on this pull, the policy analyst joins as the paid professional on the match
Your own team builds it. Across the 5 adoptions here, 0 were low effort, 4 medium and 1 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: plants with sensor-ready control systems and a central control room. First step: connect plants through sensors to an optimisation centre. Running: data engineers and operators trained on AI forecasts for variable renewable output.
About 500 AI agents across networks, generation, customer service and corporate functions
Where
ES, Europe
Segment
enterprise
Problem
incident response time and admin load
Technology
AI agents, voicebots and chatbots
Who uses it
Iberdrola, Spain
Outcome
Nearly 500 AI agents deployed or in development; supply restoration aimed for 'within minutes' (Iberdrola, Aug 2026)
Effort
high
Risks
Agent sprawl; outage-response errors; no outcome numbers published.
Next: further agents are in development alongside those already deployed source
How to adopt
Let staff propose use cases, train them, register every agent and review each one.
How a room runs it
the control room: the grid operator, the customer team and an incident agent; the agent anticipates incidents and GENERATES restoration and complaint-handling tasks with deadlines, each a proposal a human releases; the receipt records the signal, the task and the approver
The engagement on top
connectors to the network and customer systems, a champion per department, training for staff
AI for predictive maintenance, well planning and seismic interpretation across an energy major
Where
NO, Europe
Segment
enterprise
Problem
unplanned downtime and planning cost
Technology
machine learning on sensor data; copilots, chatbots and agentic AI for staff
Who uses it
Equinor, Norway
Outcome
About USD 130 million of value and savings in 2025 and over USD 330 million since 2020; over 700 rotating machines with 24,000 sensors monitored (Jan 2026 report)
Effort
high
Risks
Savings are the company's own estimate; sensor drift; over-reliance on alerts.
Next: AI is central to holding Norwegian shelf output at about 1.2 million barrels of oil equivalent a day through 2035 source
How to adopt
Monitor critical machines first, count savings against a baseline, then extend to planning.
How a room runs it
the field room: the reliability engineer, the planner and a monitoring agent; the agent reads sensor streams and GENERATES maintenance work orders with due dates, each a proposal the engineer approves; well options are generated for experts to review, and the receipt records the signal and the decision
The engagement on top
connectors to the sensor historian and work-order system, a reliability engineer
Digitised generation fleet: nearly all plants of PLN and its subsidiaries Indonesia Power and PJB connected through about 5,000 sensors, with optimisation centres (i-CORE, REOC) and AI used to handle the variable output of solar and wind plants
Where
ID, Southeast Asia
Segment
public (state utility)
Problem
reliability and efficiency of a national fleet; unpredictable renewable output
Technology
sensors, control-room digitisation, optimisation centres, AI for renewable variability
Who uses it
PLN, PT Indonesia Power, PT Pembangkitan Jawa-Bali (Indonesia)
Outcome
none stated
Effort
high
Risks
No outcome figures reported; cyber exposure of connected control systems; cost.
How to adopt
Prerequisite: plants with sensor-ready control systems and a central control room. First step: connect plants through sensors to an optimisation centre. Running: data engineers and operators trained on AI forecasts for variable renewable output.
How a room runs it
the ROOM for a plant's reliability week: the plant manager, the shift engineer, the maintenance planner and a reliability teammate (an agent definition with the generation skills); the agent reads the readings and work-order history the utility already holds and GENERATES the week's inspection and maintenance tasks for the planner as plan steps with owners and due dates, and the outage-window question as a proposal in the manager's Inbox with a deadline; what the humans keep: every switching and live-equipment action (the named engineer signs it), the outage sign-off, the spend above the envelope; the receipt records each finished task and who signed
The engagement on top
a custom engagement: the utility's own systems stay theirs - the agent is wired through connectors loaded on this pull (the historian, the CMMS or work-order system, the ERP), a reliability engineer joins as the paid professional on the match, and the sector pack carries the task templates for inspections and outage planning
AI-drafted customer replies for a large energy supplier
Where
GB, Europe
Segment
enterprise
Problem
service cost and email volume
Technology
Kraken Human Assist (generative AI)
Who uses it
Octopus Energy (United Kingdom), Kraken Technologies
Outcome
60% of service emails first drafted for agents by AI; 89% of AI-drafted outbound content sent unchanged; agents save up to 4 minutes per query (Kraken case study) the company’s own figure
Effort
medium
Risks
Vendor-reported results; wrong billing facts in drafts; tone drift.
Next: the same suite sits alongside Kraken's field and flexibility platforms across the customer lifecycle source
How to adopt
Draft for agents first, measure the unchanged rate, never auto-send.
How a room runs it
the service room: the customer, the agent on duty and a drafting teammate; the agent summarises the account history and GENERATES the draft reply and any billing follow-up task with a due date, each a proposal the human agent sends or edits; the receipt records the draft, the edits and the sender
The engagement on top
connectors to the billing and ticketing systems, a tone and policy template
AI trained on past failures reads turbine sensors (pressure, vibration, temperature), flags components about to fail and tells technicians which tools and parts to bring, so repairs start before bad weather closes offshore access
Where
DK, Europe
Segment
enterprise (offshore wind operator)
Problem
unplanned downtime and costly offshore visits
Technology
failure-prediction models on turbine sensor data, drones for corrosion checks (testing)
Who uses it
Ørsted (Denmark), Horns Rev 2 offshore wind farm
Outcome
none stated
Effort
high
Risks
No downtime or cost figures published; false alarms trigger needless trips; dependence on sensor quality.
How to adopt
Collect sensor history and the failure log per turbine type; train on past failures for one component first; agree who acts on an alert and how spare parts are staged; expect integration with maintenance planning.
How a room runs it
the ROOM for an offshore wind farm: the site manager, the service-planning lead and a maintenance teammate; the agent teammate reads the data the operator already has and GENERATES the service tasks (inspect gearbox on turbine 14 before Thursday's weather window, parts list attached) as plan steps with owners and due dates, and the dispatch decision for the site manager (send the vessel now / wait) with the sensor evidence as a decision block in the Inbox with a deadline; what the humans keep: the vessel call, the climb and the repair; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (turbine SCADA and condition-monitoring feeds, the maintenance system, the weather window forecast), a wind-turbine reliability engineer joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Predictive failure diagnosis on a geothermal plant: AI reads real-time plant sensor data and flags anomalies (such as unstable steam flow into the turbine) before they cause a shutdown
Where
ID, Southeast Asia
Segment
mid (independent power producer)
Problem
unplanned shutdowns lower the plant's utilisation rate
Technology
IoT service, AI anomaly and failure-sign diagnosis, performance monitoring (EtaPRO suite)
Who uses it
PT Geo Dipa Energi (Persero), Patuha plant, with Toshiba Energy Systems & Solutions (Indonesia, Japan)
Outcome
problem rate reduced by over 20 % in the NEDO demonstration (reported in Toshiba's release; baseline and method not given) the company’s own figure
Effort
medium
Risks
Release is the vendor's with no independent corroboration; false alarms; vendor lock-in.
How to adopt
Prerequisite: plant sensor data streams and a performance-monitoring suite. First step: add AI anomaly and failure-sign detection on the turbine. Running: engineers to act on alerts and keep models tuned.
How a room runs it
the room for the plant: the plant manager, the O&M lead, the control-room operator and a diagnostics teammate; when the vendor's service raises a sign of anomaly, the teammate turns it into a TASK for the O&M lead (inspect the steam line before Thursday's window) and a DECISION for the manager (derate, wait, or schedule the stop) with the evidence attached and a deadline; the operator keeps the control action, the manager the stop; the receipt records the alert, the decision and who acted
The engagement on top
the plant's own diagnostics service and historian wired to the room through connectors loaded on this pull; a rotating-equipment engineer joins as the paid professional on the match; the sector pack carries the anomaly-to-task templates
Neural-network forecast of wind output 36 hours ahead on 700 MW of wind farms, used to recommend hourly delivery commitments to the grid a day in advance
Where
US, North America
Segment
enterprise (technology company as wind owner)
Problem
scheduled power is worth more to the grid than unpredictable output
value of the wind energy up about 20 % versus no time-based commitments (Google's own figure, not independently audited) the company’s own figure
Effort
high
Risks
Value uplift is the owner's figure with an undefined baseline; forecast error costs money when commitments are missed; depends on a market that rewards scheduled delivery.
How to adopt
Have turbine output history and a weather-forecast feed; train a short-horizon forecast on one farm; compare scheduled against unscheduled revenue before committing; expect data-science effort and a market that pays for firm delivery.
How a room runs it
the ROOM for a wind portfolio desk: the portfolio manager, the grid-scheduling analyst and a forecasting teammate; the agent teammate reads the data the operator already has and GENERATES the day-ahead tasks for the scheduler (confirm tomorrow's hourly commitment per farm by 10:00) as plan steps with owners and due dates, and the commitment recommendation for the portfolio manager (commit / hold back) with the forecast and its error band as a decision block in the Inbox with a deadline; what the humans keep: the commitment to the grid and the risk limit; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (turbine output feed, the weather-forecast subscription, the scheduling system), an energy-markets analyst joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Source: vendor · 2019 · vendor only; to corroborate
In-house industrial generative AI model trained on company history to advise on wells and refined-product forecasts
Where
SA, Middle East
Segment
enterprise
Problem
scattered drilling and market data
Technology
Metabrain (250-billion-parameter model)
Who uses it
Saudi Aramco, Saudi Arabia
Outcome
250 billion parameters trained on seven trillion data points from more than 90 years of company history (Offshore Technology, Mar 2024)
Effort
high
Risks
Model errors in drilling decisions; no published outcome yet; data leakage.
Next: a 1-trillion-parameter version planned by end of 2024 source
How to adopt
Train on owned records, release to engineers as advice, add forecasting after.
How a room runs it
the planning room: the petroleum engineer, the economist and a recommendation agent; the agent reads the well history and GENERATES ranked well options and product forecasts with review dates, each a proposal the engineer accepts; the receipt records the data used and the decision
The engagement on top
connectors to the well and market data stores, an engineer on the match
Smart metering, leak detector and command-centre services offered to the 427 PDAMs of the national association; trials at Perumda Tirta Raharja (Cimahi) and a proof of concept at PDAM Tirta Benteng Tangerang
Pilot scope only with no loss-reduction figures; capital cost; vendor lock-in.
How to adopt
Prerequisite: a water utility willing to replace meters and a connectivity network. First step: trial LoRaWAN smart meters and a leak detector in one zone. Running: meter fleet, network and command-centre staff.
How a room runs it
the room for a water utility district: the director, the technical manager, the field crew lead and a network teammate; the agent reads the meter and pressure readings the utility already receives and GENERATES the leak-check tasks for the crew lead as plan steps with owners and due dates, and a repair-priority or tariff question as a proposal in the director's Inbox; what the humans keep: the dig and the repair on live mains, the shutdown notice to customers, the spend above the envelope; the receipt records each closed task
The engagement on top
the utility's billing system and meter-data feed wired through connectors loaded on this pull; a water-network engineer joins as the paid professional; the sector pack carries the district-check templates; crew tasks go out over WhatsApp (the channel is built)
An agentic AI platform for upstream operations: AI agents trained on tasks from subsurface and seismic analysis to real-time process monitoring, rolled out across more than 28 producing fields
Where
AE, Middle East
Segment
enterprise (national oil company)
Problem
upstream data is large and spread over many assets and specialists
Technology
agentic AI platform (ENERGYai) on cloud with the OSDU subsurface data framework and OpenAI models
Who uses it
ADNOC (UAE), AIQ (a Presight company), G42, Microsoft
Outcome
none stated
Effort
high
Risks
Benefits are expectations, not measurements; data-standard integration cost; operating decisions must stay with engineers.
How to adopt
Begin with one asset and a narrow task such as monitoring or seismic interpretation, keep engineers reviewing every output, and agree on the data standard first.
How a room runs it
the ROOM for an upstream asset team: the asset manager, a reservoir engineer, the HSE lead and a subsurface-analysis teammate (an agent definition with the asset's data standards); the agent teammate reads the subsurface and monitoring data the operator already holds and GENERATES the engineers' follow-up tasks (re-run the history match on well 12 by Thursday) as plan steps with owners and due dates, and each recommended operating change for the asset manager (apply / defer, with the evidence attached) as a decision block in the Inbox with a deadline; what the humans keep: the operating decision, the safety case and the well work; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the subsurface data store, the process-monitoring feeds, the well-work planning system), a reservoir or production engineer joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Your own team builds it. Across the 10 adoptions here, 0 were low effort, 2 medium and 8 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: truck telemetry and maintenance records. First step: AI dispatch scheduling and failure forecasting on one pit. Running: data and planning staff and change in dispatcher routines.
AI recommendations for the copper concentrator circuit at the world's largest copper mine
Where
AU;CL, Oceania
Segment
enterprise
Problem
declining ore grades
Technology
Azure Machine Learning (Microsoft), real-time plant data
Who uses it
BHP (Australia), Escondida mine (Chile), with Microsoft
Outcome
Hourly predictions for the concentrators; BHP expected up to 5% more recovery and up to 10% lower operating cost (projections, May 2023)
Effort
high
Risks
Projections not yet audited; plant-safety limits; operator over-trust.
Next: expected copper recovery up to 5% higher and operating costs up to 10% lower source
How to adopt
Advise only at first, log every recommendation, let operators accept, widen by concentrator.
How a room runs it
the plant room: the shift superintendent, the process engineer and a recommendation agent; the agent reads plant data hourly and GENERATES set-point proposals with expiry times, each a proposal the operations team accepts or declines; the receipt records the data, the proposal and who accepted it
The engagement on top
connectors to the historian and control-room tools, a process engineer on the match
Autonomous haul trucks across five Pilbara mine sites: more than 80 Komatsu trucks, moving roughly a quarter of the material in 2017 and passing one billion tonnes
Where
AU, Oceania
Segment
enterprise (iron ore major)
Problem
manual haulage limits utilisation and raises unit cost and exposure
each autonomous truck ran about 700 more hours than a conventional truck in 2017 and load-and-haul unit costs were about 15 % lower (Rio Tinto's own estimate)
Effort
high
Risks
Company-reported figures; very high capital; workforce displacement; mixed-fleet safety.
How to adopt
Start with one pit with mapped haul roads and a controlled fleet; retrofit or buy autonomy-ready trucks and a remote control room; expect large capital, a long change programme and workforce reskilling.
How a room runs it
the ROOM for a mine operations control room: the superintendent, the dispatch lead and a haulage teammate; the agent teammate reads the data the operator already has and GENERATES the shift tasks (re-route trucks around the blocked ramp, inspect the hauled-tonne shortfall by 06:00) as plan steps with owners and due dates, and the fleet-expansion or re-sequencing proposal for the superintendent with the utilisation numbers as a decision block in the Inbox with a deadline; what the humans keep: the safety sign-off and the capital decision; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (fleet telemetry, the dispatch system, the maintenance system), a mine-planning or autonomy engineer joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Machine learning on the SAG mill at the Los Pelambres concentrator (SIRO platform) identifies and reduces operating constraints on grinding
Where
CL, Latin America
Segment
enterprise (copper miner)
Problem
grinding-mill downtime limits copper throughput
Technology
machine learning on process data, SIRO platform
Who uses it
Antofagasta plc, Los Pelambres (Chile)
Outcome
SAG mill unconstrained operating hours up about 4 % versus 2024 and availability 94.4 % (Antofagasta plc, 2025 full-year results)
Effort
medium
Risks
Company-reported; advice on a mill can cause damage if set points leave the safe envelope; benefit depends on ore variability.
How to adopt
Hold clean mill process data and a historian; run the model advisory-only beside operators on one mill; compare operating hours with and without it; expect process-engineering time to tune it.
How a room runs it
the ROOM for a concentrator control room: the plant manager, the process engineer and a process teammate; the agent teammate reads the data the operator already has and GENERATES the process tasks (test a lower feed-rate set point on mill 2 during the next shift, report the effect by Friday) as plan steps with owners and due dates, and the set-point recommendation for the plant manager with the evidence and a deadline as a decision block in the Inbox with a deadline; what the humans keep: the set-point change and the safety envelope; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the plant historian, the SAG mill controls data, the lab results), a mineral-processing engineer joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
AI for vehicle dispatch scheduling and forecasting of equipment failures at a coal mine
Where
ID, Southeast Asia
Segment
enterprise contractor
Problem
idle trucks and unplanned equipment breakdowns in an open pit
Technology
AI dispatch optimisation, predictive maintenance, fleet telemetry
Who uses it
PT Petrosea Tbk, Tabang coal mine, Kalimantan (Indonesia)
Outcome
none stated
Effort
medium
Risks
Dated source; the company's claim has no figures; dispatch errors affect production and safety.
How to adopt
Prerequisite: truck telemetry and maintenance records. First step: AI dispatch scheduling and failure forecasting on one pit. Running: data and planning staff and change in dispatcher routines.
How a room runs it
the room for a pit's week: the mine manager, the dispatch supervisor, the maintenance planner and an operations teammate; the agent reads the dispatch and maintenance records the contractor already keeps and GENERATES the work-order tasks for the planner and the shift plan for the supervisor as plan steps with due dates, and a repair-or-defer question as a proposal in the manager's Inbox; what the humans keep: stopping a haul truck or a shovel, the safety sign-off, the spend above the envelope; the receipt records the task, the decision and the signer
The engagement on top
the contractor's fleet-management and maintenance systems wired through connectors loaded on this pull; a mining operations engineer joins as the paid professional; the sector pack carries shift-plan and work-order templates
Rock-cutting robots tested in underground mines in South Africa to raise productivity, alongside digital twins used elsewhere in the group
Where
ZA, Africa
Segment
enterprise (global miner, South African underground mines)
Problem
underground mining is slow, costly and dangerous for people
Technology
robotic rock cutting, digital twins
Who uses it
Anglo American (South Africa trials; technical director Tony O'Neill as quoted)
Outcome
none stated as a result; target of 20-30% higher productivity from robotic rock cutting (Anglo American technical director, 2018)
Effort
high
Risks
Source is eight years old and the figure is a target; ground conditions vary; the company's own account.
How to adopt
Trial on one face with a baseline, measure cutting rate and safety events, and extend only when the baseline is beaten.
How a room runs it
the ROOM for an underground mine's technology trial: the mine manager, the safety lead, the trial engineer and a trial-analysis teammate (an agent definition with the mine's trial protocol); the agent teammate reads the trial logs and production reports the mine already holds and GENERATES the engineer's trial tasks (compare cutting rate on both faces by Friday) as plan steps with owners and due dates, and each go / no-go for the next stage for the mine manager (extend / stop, with the data attached) as a decision block in the Inbox with a deadline; what the humans keep: the safety case, the entry into the working area and the capital spend; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the trial telemetry, the production reports, the safety management system), a mining engineer joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Source: analyst · 2018 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 5 adoptions here, 0 were low effort, 2 medium and 3 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: a mapped network with district metering points. First step: install battery-powered meters with data loggers and SMS alerts on a few districts. Running: field crews respond to leak alerts.
AI-based burst and leak detection on district-metered-area flow and pressure data, alerting a 24-hour control room; rolled out across the whole supply area (a fully intelligent network)
Where
GB, Europe
Segment
public (UK regional water company)
Problem
bursts running long before they are found
Technology
district metered area flow and pressure sensors, AI predictive modelling (Aquasuite Burst)
Who uses it
SES Water (United Kingdom), Haskoning
Outcome
run time of critical burst mains down 40 % (SES Water's asset-intelligence lead, via the vendor) the company’s own figure
Effort
medium
Risks
Vendor page without a date or method; false alerts waste crew time; depends on meter coverage.
How to adopt
Have district metering with flow and pressure logging; start on a few DMAs and compare burst run times; agree who is paged and the response time; expect integration with the control room.
How a room runs it
the ROOM for a water-network operations room: the network manager, the leakage technician lead and a leakage teammate; the agent teammate reads the data the operator already has and GENERATES the field tasks (send a crew to DMA 12 by 14:00 with the likely pipe section marked) as plan steps with owners and due dates, and the repair-versus-monitor decision for the network manager with the pressure evidence as a decision block in the Inbox with a deadline; what the humans keep: the excavation, the repair and the customer message; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the DMA flow and pressure telemetry, the work-management system, the control-room alert feed), a leakage engineer joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
A virtual AI assistant answers customer enquiries about electricity and water in English and Arabic around the clock, reachable also through a voice assistant
Where
AE, Middle East
Segment
enterprise (government utility)
Problem
customer enquiries and office visits are costly for the utility and the government
Technology
conversational AI assistant (Rammas), voice assistant channel
Who uses it
Dubai Electricity and Water Authority (UAE)
Outcome
Rammas handled over 1.6 million inquiries in 2025 and more than 12.7 million since its 2017 launch (DEWA's own figures, January 2026)
Effort
medium
Risks
No result figures on the page; wrong answers on billing; language coverage; undated source.
How to adopt
Start with the 20 most common questions, answer in the customers' languages, log every unanswered question, and hand the rest to a person with the conversation attached.
How a room runs it
the ROOM for a utility's customer-service desk: the service manager, two senior agents, the content owner and a customer-assistant teammate (an agent definition with the utility's tariff and service rules); the agent teammate reads the enquiry log the utility already holds and GENERATES the content owner's fixes (rewrite the answer on billing disputes by Wednesday) as plan steps with owners and due dates, and each enquiry the assistant cannot settle for the service manager (refund / escalate, with the conversation attached) as a decision block in the Inbox with a deadline; what the humans keep: every refund, every exception and the tone of the replies; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the enquiry log, the billing system export, the knowledge base), a customer-experience specialist joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
A 10-year programme inspects about 64 km of large-diameter pipe with a wireless acoustic device that works inside live pressurised pipes, AI-assisted on a cloud platform, to find hidden leaks
Where
BR, Latin America
Segment
enterprise (state water utility)
Problem
the utility loses nearly 29% of its water supply to leaks
Technology
in-pipe acoustic inspection (Nautilus), AI analysis on a cloud platform
Who uses it
Sabesp (Brazil), Aganova (Spain), Effico (Brazil), Amazon Web Services as sponsor
Outcome
none stated as a result; estimated 210 million litres a year saved over about 64 km of pipe (programme estimate); Sabesp loses nearly 29% of supply to leaks
Effort
high
Risks
Savings are an estimate; vendor and sponsor interest; repair capacity may lag detection.
How to adopt
Pick the large-diameter mains with the highest loss, inspect them in service, rank findings by water lost, and tie each finding to a repair work order.
How a room runs it
the ROOM for a water utility's loss-control team: the network manager, the field crew lead, the programme sponsor's liaison and a leak-analysis teammate (an agent definition with the utility's pipe register); the agent teammate reads the inspection results and the pipe register the utility already holds and GENERATES the crew's repair and re-inspection tasks (excavate at node 214 and confirm by Monday) as plan steps with owners and due dates, and each repair-versus-replace choice for the network manager (repair now / schedule, with the leak evidence attached) as a decision block in the Inbox with a deadline; what the humans keep: the excavation, the repair and the capital spend; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the pipe asset register, the inspection result files, the work-order system), a water-network engineer joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
AI and satellite ground-movement data predict which water mains are likely to break; validated on 250 miles and then planned for 1,200 miles
Where
US, North America
Segment
public (non-profit county utility)
Problem
reactive repair of main breaks at about $7,000 each
Technology
machine-learning failure model, satellite ground-motion and vegetation data (Rezatec Pipeline Risk)
Who uses it
WaterOne, Johnson County Kansas (United States), Rezatec
Outcome
model 78 % accurate in predicting likelihood of failure on 250 miles (utility's GIS lead, via the vendor) the company’s own figure
Effort
medium
Risks
Vendor-published; validation was by the utility and the vendor; ranking errors misdirect capital.
How to adopt
Assemble break history, pipe material, diameter and age; validate a risk model on a withheld year; use it to rank replacements, not to skip inspection; expect GIS and data-quality work.
How a room runs it
the ROOM for a distribution-engineering team: the engineering lead, the capital-planning manager and a main-risk teammate; the agent teammate reads the data the operator already has and GENERATES the planning tasks (inspect the 40 highest-risk mains in zone 3 by month-end) as plan steps with owners and due dates, and the replacement-priority proposal for the capital-planning manager with the risk ranking and cost as a decision block in the Inbox with a deadline; what the humans keep: the capital budget and the field decision; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the pipe asset register, the break history, the GIS layers), a water-asset-management engineer joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Source: vendor · 2023 · vendor only; to corroborate
Advanced metering with 39 district metering points, enhanced leak detection, real-time consumption data and text alerts for network losses
Where
ID, Southeast Asia
Segment
public
Problem
non-revenue water
Technology
battery-powered electromagnetic meters, data logging, pressure detection, SMS alerts
Who uses it
PDAM Kota Malang, East Java (Indonesia), with Itron
Outcome
none stated
Effort
high
Risks
Old source from 2012 and pre-AI; maintenance of battery meters; cost.
How to adopt
Prerequisite: a mapped network with district metering points. First step: install battery-powered meters with data loggers and SMS alerts on a few districts. Running: field crews respond to leak alerts.
How a room runs it
the same utility room: the technical manager and the crew lead with a teammate that reads the alerts the metering already raises and turns each into a field TASK with an owner and a due date, plus a pressure-zone decision for the director; the crew keeps the field work, the director the shutdown call; the receipt records the alert, the visit and the finding
The engagement on top
the metering vendor's data export wired through a connector loaded on this pull; the water-network engineer as the paid professional on the match
Source: vendor · 2012 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 5 adoptions here, 0 were low effort, 3 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: a registration list with attendee interests. First step: switch on AI matchmaking and meeting booking in an event-networking platform. Running: organiser promotes it and reviews analytics.
AI matchmaking, messaging and meeting scheduling at conferences, so attendees meet the right people instead of drifting through the crowd
Where
multi, global
Segment
mid and enterprise
Problem
attendees overwhelmed by crowds and unable to find people with shared interests or industries
Technology
event-networking platform with AI matchmaking, in-app messaging, meeting booking, pre/post-event analytics
Who uses it
Forbes (30 Under 30 Summit 2025 and other Forbes events), USA, with the platform Brella
Outcome
none stated
Effort
low
Risks
Vendor page with no usage figures; attendee data privacy; the country is not stated on the page.
How to adopt
Prerequisite: a registration list with attendee interests. First step: switch on AI matchmaking and meeting booking in an event-networking platform. Running: organiser promotes it and reviews analytics.
How a room runs it
the ROOM for one event: the organiser, the programme lead, the sponsor manager and an events teammate (an agent definition with the event skills); from the attendee list and agenda the organiser already holds, the agent GENERATES the run-up as plan steps with owners and due dates (invite the top-requested speakers, confirm sponsor meeting slots, brief the floor team) and puts to the organiser, as a decision in the Inbox with a deadline, which sessions to move or merge; the humans keep the guest relationships, the venue calls and any spend above the event budget; the receipt records every confirmed meeting slot and every change; the budget degrades before it blocks
The engagement on top
a custom engagement: the organiser's own registration and ticketing system and attendee list wired to the room through connectors loaded on this pull, the event producer as the paid professional on the match, the events pack with run-sheet and sponsor-fulfilment task templates; attendee and speaker updates over WhatsApp (the channel is built)
Source: vendor · 2025 · vendor only; to corroborate
A flagship sporting event gives fans a generative AI assistant that answers questions during live matches from match data, player statistics and expert opinion
Where
GB, Europe
Segment
enterprise
Problem
fans want instant, trustworthy answers during play; editorial staff cannot answer every question
Technology
generative AI on a vendor platform with orchestrated AI agents and large language models trained on the event's editorial style
Who uses it
All England Lawn Tennis Club (Wimbledon), UK, with IBM
Outcome
none stated
Effort
high
Risks
Wrong answers during live play; no engagement figures published; vendor lock-in.
How to adopt
Prerequisite: match data, statistics and editorial content in clean form. First step: build a generative assistant with a cloud-AI partner on that content. Running: AI agent monitoring, content updates and a cloud contract.
How a room runs it
the ROOM for a large event's digital team: the communications lead, the editorial lead, the technology lead and a fan-experience teammate; the agent reads the content and data feeds the club already owns and GENERATES the pre-event checklist as plan steps (review the assistant's answer templates, sign off the tone, test the fallback) with owners and due dates, and files a decision in the editorial lead's Inbox whenever an answer type needs a human line; the humans keep the editorial voice and every public statement; the receipt records which answer templates were signed off and by whom; a public room with receipts can carry the event's post-match notes, never a feed
The engagement on top
a custom engagement: the organiser's content management system, data feeds and ticketing wired to the room through connectors loaded on this pull, an editor or communications professional as the paid professional on the match, the events pack's sign-off task templates
A fair organiser gives visitors and exhibitors a 24/7 multilingual chatbot for floor plans, exhibitors, programme and services, plus a matchmaking tool that suggests business partners from visitor interests and exhibitor product categories, with chat and appointment booking
Where
DE, Europe
Segment
enterprise (trade fair)
Problem
visitors and exhibitors at a large fair cannot find each other or the programme
Technology
chatbot, recommendation algorithm, event app
Who uses it
Messe Düsseldorf (K 2025 trade fair), Germany
Outcome
none stated
Effort
medium
Risks
Low exhibitor take-up leaves suggestions thin; wrong chatbot answers on practical details; data-protection duties on visitor profiles.
How to adopt
Prerequisite: clean exhibitor and product-category data and an event app. First step: switch on profile-based suggestions for one hall or one visitor group. Running: content upkeep and exhibitor onboarding each edition.
How a room runs it
the ROOM for a trade fair's digital team: the project director, the exhibitor-services lead, the app owner and a visitor-experience teammate; the agent reads the exhibitor list, programme and category data the organiser already holds and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the exhibitor relationship and every public statement; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the organiser's registration, exhibitor database and app wired to the room through connectors loaded on this pull, an exhibition-operations professional as the paid professional on the match, the events pack's sign-off task templates
Event-technology platforms add AI matchmaking, session recommendation, chatbots, translation and copy drafting, which organisers switch on for their events
Where
multi, global
Segment
SMB, mid and enterprise
Problem
organisers lack time to match attendees, answer repeat questions, translate and write promotion
Technology
AI matchmaking, recommendation engines, chatbots, neural translation, language-model copy tools
Who uses it
event organisers using Cvent, Stova, Bizzabo, Swapcard, Brella, Grip, Wordly, Interprefy and others (a trade survey of the vendors, 61% offering at least one AI feature)
Outcome
61 % of event-tech vendors offer at least one AI feature; about 40 % use AI for connection suggestions; under 15 % offer AI chatbots (Skift Meetings 2023 vendor survey)
Effort
low
Risks
The survey names vendors, not organiser adopters; the vendors' own claims; attendee privacy.
How to adopt
Prerequisite: an event-technology platform. First step: switch on matchmaking, session recommendation and chatbot features for one event. Running: organiser reviews AI translation and drafted copy.
How a room runs it
the ROOM for a small organiser: the owner, the ops coordinator and an events teammate; the agent reads the organiser's attendee export and run-sheet and GENERATES the week's tasks (send the speaker brief, chase the unpaid sponsors, translate the schedule) with owners and due dates, and the owner's choices (extend registration, add a session) arrive as proposals with a deadline; the humans keep the guests, the venue and the money; the receipt records each task done; the organiser keeps their own platform and the agent reads from it
The engagement on top
a custom engagement: the organiser's own event platform and spreadsheet exports wired in through connectors loaded on this pull, a freelance event producer as the paid professional, the events pack's task templates; attendee notices over WhatsApp
A world fair launched an AI-powered digital assistant (Amal) on its website and app, built with the city's smart-government office and a consultancy, which served visitors during the fair (Oct 2021 to Mar 2022); the event has ended, so this is a completed past deployment, not a current service
Where
AE, Middle East
Segment
public (world fair)
Problem
millions of international visitors need information in their own language
Technology
AI assistant, website and mobile app
Who uses it
Expo 2020 Dubai with Smart Dubai and Accenture, United Arab Emirates
Outcome
none stated
Effort
medium
Risks
Written before opening, results unknown; wrong answers in many languages; accessibility and privacy duties.
How to adopt
Prerequisite: programme and venue content in clean form and a translation process. First step: launch the assistant for the top visitor questions in two languages. Running: content updates and human escalation.
How a room runs it
the ROOM for a large venue's visitor-services team: the visitor-services lead, the languages coordinator, the app owner and a visitor teammate; the agent reads the venue's programme, map and service content and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the escalation to staff and every public statement; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the venue's content system, ticketing and app wired to the room through connectors loaded on this pull, a visitor-experience professional as the paid professional on the match, the events pack's answer-template sign-off tasks
The national trade body gives every registered participant of an international halal fair a dedicated AI agent for business matching, multilingual support, navigation, meeting management and post-event follow-up
Where
MY, Southeast Asia
Segment
public (trade promotion agency)
Problem
thousands of buyers and exporters must find the right counterpart within four days
Technology
AI agents, business-matching platform, event app
Who uses it
Malaysia External Trade Development Corporation (Matrade) for MIHAS 2026, Malaysia
Outcome
none stated (the organiser gives expected figures only: 50,000 visitors, about 4,000 business-matching meetings, a RM5 billion sales target)
Effort
medium
Risks
Announced, not yet shown to work; matching quality unproven; data-protection duties on participant profiles.
How to adopt
Prerequisite: structured buyer and exporter profiles and a meeting scheduler. First step: give agents to one buyer programme and compare meetings made with the previous edition. Running: profile upkeep and multilingual content.
How a room runs it
the ROOM for an export-fair delegation: the exporter, the sales lead, the buyer-programme coordinator and a matching teammate; the agent reads the buyer and exporter profiles the programme already holds and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the buyer relationship, the quote and the contract; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the agency's buyer registry and meeting scheduler wired to the room through connectors loaded on this pull, an export or trade-fair professional as the paid professional on the match, the events pack's meeting and follow-up task templates
Your own team builds it. Across the 6 adoptions here, 2 were low effort, 3 medium and 1 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: cloud property-management and loyalty platforms with unified guest data. First step: Copilot for staff and a beta at a few properties. Running: shared AI orchestration and an 18-month rollout.
A hotel group launches a conversational AI concierge on its website, app and messaging channels, plus a separate AI tool for hotel staff
Where
FR, Europe
Segment
enterprise
Problem
guests need help from trip planning to post-stay; hotel teams need quick access to information
Technology
conversational AI on web, mobile app, WhatsApp and iMessage; staff-side AI tool (The Butler)
Who uses it
Accor, France (ALL Concierge)
Outcome
none stated
Effort
medium
Risks
Scale of rollout not stated; wrong answers to guests; guest data privacy.
How to adopt
Prerequisite: property and guest data and messaging channels. First step: launch a conversational concierge on web, app and messaging and a separate staff tool. Running: content upkeep and escalation to staff.
How a room runs it
the ROOM for a hotel team: the GM, the duty manager, the concierge and a guest teammate; the guest's WhatsApp request becomes a TASK for the named person with a due time, and anything outside the envelope (a comp night, a transfer) lands in the duty manager's Inbox as a proposal with a deadline; the staff keep the hospitality; the receipt records each request and its fulfilment
The engagement on top
a custom engagement: the hotel's own PMS, housekeeping and loyalty tools wired to the room through connectors loaded on this pull (WhatsApp is a built channel), a hospitality consultant as the paid professional, the hospitality pack's task templates
The national hotel and restaurant association launches a multilingual AI travel assistant tied to its booking platform
Where
ID, Southeast Asia
Segment
public
Problem
foreign and domestic travellers cannot find hotel information in their language and the path to payment is long
Technology
multilingual AI assistant, recommendation learning from collected data, link to the BookingINA booking site
Who uses it
Indonesian Hotel and Restaurant Association (PHRI) with PT AI Indonesia (Travina Jelajah Indonesia), Indonesia, Vina AI Travel Assistant
Outcome
none stated
Effort
medium
Risks
No usage data; hotel information may be wrong or out of date; payment-step security.
How to adopt
Prerequisite: a booking platform and hotel information. First step: connect a multilingual AI assistant to the booking platform. Running: member hotels keep data current.
How a room runs it
the ROOM for a member hotel: the owner or GM, the reservations clerk and a hotel teammate; enquiries arriving through the association platform become TASKS (confirm the room for 12 March, reply in Mandarin, send the payment link) for the clerk with due times, and a rate or group-discount choice is a proposal in the GM's Inbox with a deadline; the humans keep the host role; the receipt records each enquiry answered and booked
The engagement on top
a custom engagement: the hotel's own PMS or booking sheet and WhatsApp number wired to the room through connectors loaded on this pull, a hotel revenue consultant as the paid professional, the hospitality pack's task templates
A hotel franchisor deploys 230 AI agents that know each of its 8,300 hotels, covering direct bookings, guest questions, travel recommendations, voice and messaging on WhatsApp and Apple Messages
Where
US, North America
Segment
enterprise (franchise)
Problem
franchised hotels need direct bookings and shorter handling times without more staff
Technology
AI agents, voice assistants, messaging channels
Who uses it
Wyndham Hotels & Resorts, USA
Outcome
Over 500,000 customer interactions handled; about 25 % lower average handle time; 300 basis points better direct contribution at fully adopted hotels; about 7 % of the portfolio adopted (Wyndham, Q3 2025 earnings call, as reported)
Effort
medium
Risks
Self-reported figures; low adoption so far; wrong answers on rates and policy.
How to adopt
Prerequisite: property content kept current and a booking integration. First step: enable the agent on one hotel's direct-booking and FAQ channels. Running: franchisee onboarding and content upkeep.
How a room runs it
the ROOM for a franchised hotel: the owner, the front-desk lead, the reservations lead and a guest teammate; the agent reads the property's rates, availability and policy content and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the guest relationship and every exception to a rate or policy; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the property management system, booking engine and messaging channels wired to the room through connectors loaded on this pull, a hotel operations professional as the paid professional on the match, the hospitality pack's task templates
A hotel group builds cloud platforms and an 'agentic mesh' to reuse AI across functions, with Copilot for associates and back-of-house automation
Where
US, North America
Segment
enterprise
Problem
manual back-of-house processes and fragmented guest data
Technology
cloud reservation, property-management and loyalty platforms, shared AI orchestration, Microsoft 365 Copilot, generative AI for contact-centre coaching
Who uses it
Marriott International, USA
Outcome
none stated
Effort
high
Risks
The CIO says no agent pilots yet; guest data privacy; cost and change management.
How to adopt
Prerequisite: cloud property-management and loyalty platforms with unified guest data. First step: Copilot for staff and a beta at a few properties. Running: shared AI orchestration and an 18-month rollout.
How a room runs it
the ROOM for a hotel's general manager: the GM, the front-office, housekeeping and F&B heads and a hotel-operations teammate (an agent definition with the hospitality skills); from the reservation and staffing data the hotel already has, the agent GENERATES the week's tasks (hold two rooms for the group arriving Friday, move the linen order, brief the night shift) with owners and due dates, and the GM's calls (overbook, upgrade, close the pool) as proposals with a deadline; the humans keep the guests and the sign-off; the receipt records each task and decision
The engagement on top
a custom engagement: the property's own PMS, channel manager and rota wired to the room through connectors loaded on this pull, a hotel operations manager as the paid professional, the hospitality pack's task templates; guest and staff notices over WhatsApp
A hotel group launches three AI agents: one answers guests by email and WhatsApp in more than 50 languages with a route to staff, one answers staff on operating procedures, one analyses booking and demand data to recommend pricing, timing and marketing
Where
HK, East Asia
Segment
enterprise
Problem
guests and staff need answers fast and revenue teams need clearer demand signals
Technology
AI agents, messaging channels, demand analytics
Who uses it
Langham Hospitality Group, Hong Kong (31 properties on four continents)
Outcome
none stated
Effort
medium
Risks
Wrong or off-brand replies; pricing advice acted on without review; no results published.
How to adopt
Prerequisite: current procedures, rate data and a guest-messaging channel. First step: start with the staff-facing knowledge agent on one property, then guest messaging. Running: content upkeep and human review of pricing advice.
How a room runs it
the ROOM for a hotel group's regional team: the general manager, the revenue manager, the guest-services lead and a guest teammate; the agent reads the booking data, house rules and procedures the group already holds and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the guest relationship, the rate decision and every public statement; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the property management system, channel manager and messaging inbox wired to the room through connectors loaded on this pull, a hotel revenue or operations professional as the paid professional on the match, the hospitality pack's task templates
A hotel group embeds a generative AI travel planner in its loyalty app, built on a cloud AI platform, for recommendations near a stay
Where
GB, Europe
Segment
enterprise
Problem
guests want local, specific recommendations during planning
Technology
generative AI (Gemini models on Vertex AI) inside the loyalty app; AI tools in contact centres
Who uses it
IHG Hotels & Resorts, UK, with Google Cloud
Outcome
none stated
Effort
medium
Risks
Launch was planned, not confirmed live; wrong or outdated recommendations; vendor lock-in.
How to adopt
Prerequisite: a loyalty app and a cloud AI platform. First step: embed a generative travel planner for local recommendations. Running: model monitoring and content checks.
How a room runs it
the ROOM for a hotel's guest-experience desk: the concierge lead, the GM and a guest teammate; the agent reads the guest's booking and requests the hotel already holds and GENERATES the arrival tasks (book the jazz table, arrange the early check-in) for the concierge with due dates, and a request above the hotel's envelope reaches the GM's Inbox as a proposal with a deadline; the humans keep the welcome; the receipt records what was promised and delivered
The engagement on top
a custom engagement: the hotel's own PMS, loyalty or guest-request tool and WhatsApp wired to the room through connectors loaded on this pull, a concierge or guest-relations professional as the paid professional, the hospitality pack's task templates
Your own team builds it. Across the 6 adoptions here, 0 were low effort, 5 medium and 1 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: guest messaging channels. First step: switch on an AI guest-communication assistant for routine queries. Running: front desk reviews upsell and escalations.
ChatGPT Enterprise across the corporate and hotel workforce after two years of building the data layer; natural-language search on Hyatt.com and agents in group sales
Where
US, North America
Segment
enterprise
Problem
manual corporate and sales work and rigid website search
Technology
ChatGPT Enterprise (GPT-5.4 and Codex); Hyatt's private-cloud AI and data layer
Who uses it
Hyatt Hotels Corporation, United States
Outcome
Rolled out April 2026 across finance, marketing and brand, business development and real estate, product and engineering and customer experience; earlier projects (natural-language Hyatt.com search, group-sales agents) credited by Hyatt with productivity and revenue gains, no figure in the visible text
Effort
high
Risks
Data leakage; unproven return; reliance on one model vendor.
Next: Hyatt calls ChatGPT Enterprise a core component of how the business runs day to day and names finance, marketing, development, product and customer experience as the first areas source
How to adopt
Build the data layer first; roll out with live training per department; keep multiple model providers behind one gate.
How a room runs it
the sales and property room: the group-sales lead, the hotel GM, revenue management and a drafting teammate; the agent drafts proposals and rate options and GENERATES the follow-up tasks per property with deadlines, and each contract term or discount becomes a proposal in the sales lead's Inbox; the receipt records the draft, the reviewer and the final rate
The engagement on top
connectors to the property-management and CRM systems, a hotel-sales teammate, training sessions per department
A home-sharing platform deploys an AI customer-service agent for guests and hosts, cutting contacts to live agents
Where
US, North America
Segment
enterprise
Problem
high-stakes support queries need fast answers
Technology
custom AI agent built on 13 models, tuned on thousands of conversations
Who uses it
Airbnb, USA
Outcome
15 % fewer users needing a human agent (the CEO's figure)
Effort
high
Risks
High-stakes mistakes in support; the figure is the company's own; staff displacement.
How to adopt
Prerequisite: a large set of past support conversations and policy data. First step: build and tune an AI agent on those and introduce it to a share of users. Running: ongoing tuning and human escalation.
How a room runs it
the ROOM for a villa or homestay host: the owner, the caretaker, the cleaner and a host teammate; guest messages from the platform become TASKS (fix the hot water before check-in, send the door code) for the named person with due times, and a refund or compensation above the envelope is a proposal in the owner's Inbox with a deadline; the host keeps the welcome; the receipt records each issue and its resolution
The engagement on top
a custom engagement: the host's channel manager, calendar and WhatsApp wired to the room through connectors loaded on this pull, a property manager as the paid professional, the hospitality pack's turnover and guest-issue task templates
A Jakarta hotel adopts an AI guest-communication tool to handle routine guest queries, upsell and analyse guests, freeing the front desk
Where
ID, Southeast Asia
Segment
mid
Problem
front desk overwhelmed by routine queries
Technology
AI guest-messaging assistant, automated upselling, guest analytics
Who uses it
Hotel Jakarta, Indonesia, with Runnr.ai
Outcome
none stated
Effort
low
Risks
No figures; guest data privacy; wrong answers to guests.
How to adopt
Prerequisite: guest messaging channels. First step: switch on an AI guest-communication assistant for routine queries. Running: front desk reviews upsell and escalations.
How a room runs it
the ROOM for a hotel front office: the front-office manager, the duty staff and a guest-desk teammate; the agent reads the guest messages and reservations the hotel already has and GENERATES the tasks (prepare the extra bed for room 412, confirm the airport pickup) for the named staff with due times, and an offer or complaint remedy outside the envelope reaches the manager's Inbox as a proposal with a deadline; staff keep the personal service; the receipt records each request and its fulfilment
The engagement on top
a custom engagement: the hotel's own PMS and WhatsApp number wired to the room through connectors loaded on this pull (the channel is built), a front-office trainer as the paid professional, the hospitality pack's guest-request task templates
Source: vendor · 2023 · vendor only; to corroborate
A hotel brand builds in-house a virtual concierge on a ChatGPT-class model that tours the hotel, guides guests across the owner's other properties and recommends city attractions and events
Where
AE, Middle East
Segment
enterprise
Problem
guests ask the same questions about the hotel and the city at all hours
Technology
generative AI concierge
Who uses it
Address Hotels + Resorts (Emaar Hospitality), United Arab Emirates
Outcome
none stated
Effort
low
Risks
Wrong or stale recommendations; vendor model dependence; no results published.
How to adopt
Prerequisite: service and city content in clean form. First step: launch on one hotel's lobby QR code or website. Running: content upkeep and review of recommendations.
How a room runs it
the ROOM for a hotel's guest-services team: the duty manager, the concierge lead, the marketing lead and a guest teammate; the agent reads the hotel's service content and the city guide the brand already owns and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the guest relationship and every recommendation that carries a commitment; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the property management system, spa and dining booking tools and the content store wired to the room through connectors loaded on this pull, a concierge or guest-experience professional as the paid professional on the match, the hospitality pack's task templates
Source: press · 2023 · vendor only; to corroborate
A hotel group launches a beta generative AI concierge on its website that compares properties and answers open-ended planning questions
Where
US, North America
Segment
enterprise
Problem
guests cannot easily compare brands and properties
Technology
generative AI concierge on the brand website
Who uses it
Hilton, USA, Hilton AI Planner
Outcome
none stated
Effort
medium
Risks
Limited beta; wrong comparisons or availability; brand-trust risk.
How to adopt
Prerequisite: structured property data on the brand website. First step: beta a generative concierge that compares properties. Running: monitoring and content updates.
How a room runs it
the ROOM for a hotel property team: the GM, the reservations lead, the sales manager and a guest teammate; enquiries become TASKS for the named person (send the group rate, hold the ballroom) with due dates, and a rate exception is a proposal in the GM's Inbox with a deadline; the sales team keeps the relationship; the receipt records the quote and the outcome
The engagement on top
a custom engagement: the property's PMS, sales tool and WhatsApp wired to the room through connectors loaded on this pull, a hotel sales professional as the paid professional, the hospitality pack's task templates
A hotel operator runs robots and automation at the front desk and in the lobby; one hotel went from about 40 employees to 7, with nearly 300 robots of 30 types across the chain
Where
JP, East Asia
Segment
enterprise (chain)
Problem
labour cost at hotels
Technology
front-desk robots, automated check-in, IoT
Who uses it
H.I.S. Hotel Holdings (Henn na Hotel), Japan
Outcome
By January 2019 the hotel had retired about half of its 243 robots and rehired humans for reception, concierge and luggage after malfunctions (Wall Street Journal report, via Matador Network and Ubergizmo, 15 Jan 2019); earlier, the 144-room Sasebo hotel ran on 7 staff, down from about 40 (operator PR office via Nippon.com, 2018)
Effort
high
Risks
Guest complaints with robots; maintenance burden; an older source and some robots later withdrawn.
How to adopt
Prerequisite: capital for hardware and a site willing to redesign check-in. First step: automate check-in at one property. Running: robot maintenance and human fallback staff.
How a room runs it
the ROOM for a hotel's operations team: the general manager, the front-office lead, the maintenance lead and an operations teammate; the agent reads the room status, check-in and equipment records the hotel already holds and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the guest welcome, every complaint and the sign-off on any automation left running; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the property management system and the robot or kiosk vendor's service log wired to the room through connectors loaded on this pull, a hotel operations professional as the paid professional on the match, the hospitality pack's maintenance and escalation task templates
Your own team builds it. Across the 6 adoptions here, 2 were low effort, 1 medium and 3 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: drive-thru audio hardware and POS integration. First step: pilot voice AI in some restaurants. Running: order-accuracy monitoring and staff intervention.
A quick-service chain rolls out voice AI to take drive-thru orders across hundreds of restaurants
Where
US, North America
Segment
enterprise
Problem
drive-thru order accuracy, wait times and staff workload
Technology
voice AI ordering at the drive-thru
Who uses it
Taco Bell (Yum! Brands), USA; KFC Australia testing
Outcome
More than 100 US Taco Bell restaurants use voice AI (CNBC reporting); Yum Brands says order accuracy and waits improved, with no figures
Effort
high
Risks
Order errors; customer acceptance; staff displacement.
How to adopt
Prerequisite: drive-thru audio hardware and POS integration. First step: pilot voice AI in some restaurants. Running: order-accuracy monitoring and staff intervention.
How a room runs it
the ROOM for a restaurant group: the owner, the area manager, the store managers and a restaurant-ops teammate (an agent definition with the F&B skills); from the sales, stock and rota the group already records, the agent GENERATES the week's tasks (reorder cups for outlet 3, cover Saturday lunch, check the freezer log) with owners and due dates, and a menu or price change as a proposal in the owner's Inbox with a deadline; the humans keep the kitchen and the service; the receipt records each task done
The engagement on top
a custom engagement: the group's own POS, supplier ordering and staff rota wired to the room through connectors loaded on this pull, a restaurant operations manager as the paid professional, the F&B pack's task templates; supplier and staff messages over WhatsApp
Restaurants across Southeast Asia deploy robot food runners, QR ordering and kiosks to cope with labour and volume
Where
multi, Southeast Asia
Segment
mid and enterprise
Problem
staff shortages and service speed in high-volume outlets
Technology
service robots, QR ordering, self-service kiosks, intelligent kitchen systems
Who uses it
MK Restaurant (Thailand, 530 robots in 355 branches), Haidilao and Swee Choon (Singapore), others
Outcome
MK runs 530 robots across 355 branches (as cited by the article from a third-party source); no time or cost savings stated
Effort
high
Risks
The article names no Indonesian operator; capital cost; staff displacement.
How to adopt
Prerequisite: floor layout suitable for robots and a POS integrated with QR ordering. First step: add QR ordering or kiosks, then robot runners in busy outlets. Running: maintenance and staff retraining.
How a room runs it
the ROOM for a multi-outlet owner: the owner, outlet managers and a restaurant teammate; the agent GENERATES shift cover and prep tasks per outlet with owners and due dates, and a purchase above the envelope arrives as a proposal; the humans keep the service floor; the receipt records each task done; the agent never touches the robots, it reads what the outlets already record
The engagement on top
a custom engagement: the group's POS, kitchen display and rota wired to the room through connectors loaded on this pull, a restaurant operations professional as the paid professional, the F&B pack's task templates
A franchised fried-chicken chain deploys drive-thru voice ordering across its restaurants
Where
US, North America
Segment
enterprise (franchise)
Problem
drive-thru labour and speed
Technology
voice ordering AI
Who uses it
Church's Texas Chicken, USA (about 1,500 stores worldwide), with SoundHound
Outcome
About 90 % of orders without human intervention (the vendor's figure)
Effort
high
Risks
Accuracy with accents and noise; the figure is the vendor's; franchisee buy-in.
How to adopt
Prerequisite: drive-thru audio hardware and POS integration. First step: deploy voice ordering with franchisees at a few restaurants. Running: monitoring and fallback to staff.
How a room runs it
the ROOM for a franchisee: the franchisee, the store managers and an F&B teammate; the agent GENERATES the store tasks (order chicken for the weekend, schedule the fryer service) with owners and due dates, and a franchisor-driven change is a proposal in the franchisee's Inbox with a deadline; the receipt records each store's tasks done
The engagement on top
a custom engagement: the franchisee's POS and supplier orders wired to the room through connectors loaded on this pull, an operations consultant as the paid professional, the F&B pack's task templates
A food-ordering platform rolls out an AI voice assistant across its European markets that recommends restaurants and dishes by mood, budget, diet or occasion and can place a usual order
Where
NL, Europe
Segment
enterprise (platform)
Problem
finding what to order takes long on a long menu list
Technology
generative AI voice assistant
Who uses it
Just Eat Takeaway.com, Netherlands (UK, Germany and other European markets)
Outcome
Customers find what they want about 40 % faster than with the text assistant and are 30 % more likely to choose a recommended restaurant or dish (the company, early UK launch data) the company’s own figure
Effort
low
Risks
Self-reported figures; wrong recommendations on allergies; dependence on the platform.
How to adopt
Prerequisite: structured menus and a platform with order history. First step: test voice search on one market. Running: menu data upkeep and quality review.
How a room runs it
the ROOM for a restaurant group's digital team: the owner, the menu manager, the delivery lead and an ordering teammate; the agent reads the menu, stock and order data the group already holds and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the menu, the price and every refund; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the point-of-sale, the delivery-platform menu feed wired to the room through connectors loaded on this pull, a restaurant digital-ordering professional as the paid professional on the match, the hospitality pack's menu and order task templates
A pizza chain trials a humanoid service robot that takes orders and processes payments through the app and a QR code per table
Where
SG, Southeast Asia
Segment
enterprise
Problem
queues at the counter
Technology
service robot, app wallet payment, table QR code
Who uses it
Pizza Hut Singapore (Safra Punggol), with Mastercard Masterpass
Outcome
none stated
Effort
high
Risks
Trial only, not confirmed as lasting; novelty value; payment security.
How to adopt
Prerequisite: app wallet payments and table QR codes. First step: trial a service robot at one outlet. Running: robot upkeep and staff supervision.
How a room runs it
the ROOM for an outlet team: the outlet manager, the shift leads and a teammate; the agent GENERATES the shift tasks from the orders and rota already recorded, and any change to the table-order process is a proposal with a deadline; the receipt records each task done
The engagement on top
a custom engagement: the outlet's POS and rota wired to the room through connectors loaded on this pull, a restaurant professional as the paid professional, the F&B pack's task templates
A hotpot chain opens a restaurant with robot kitchen arms that prepare vegetables and meat from tablet orders, with guided vehicles carrying dishes to tables and a human waiter assisting; built in a joint venture with a robotics maker
Heavy hardware cost; no savings published; an older source.
How to adopt
Prerequisite: a joint hardware venture and a kitchen redesigned around robots. First step: pilot one restaurant. Running: maintenance and human service staff.
How a room runs it
the ROOM for a restaurant group's kitchen operations team: the restaurant manager, the kitchen lead, the maintenance lead and an operations teammate; the agent reads the order, stock and equipment logs the restaurant already keeps and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the guest service, food safety and the decision to scale a pilot; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the point-of-sale and kitchen equipment logs wired to the room through connectors loaded on this pull, a restaurant operations professional as the paid professional on the match, the hospitality pack's kitchen and maintenance task templates
Your own team builds it. Across the 6 adoptions here, 1 were low effort, 0 medium and 5 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Prerequisite: live booking and payment data behind secure APIs. First step: an agentic in-app assistant for status, changes and refunds. Running: guardrails and human fall-back.
An online travel agency's generative AI assistant covers trip planning, search, booking and post-sales help in several Indian languages, and its voice and chat bots resolve about half of flight and hotel queries without a human
Where
IN, South Asia
Segment
enterprise
Problem
travellers in smaller cities need booking help in their own language
Technology
generative AI assistant, voice bot, chatbot
Who uses it
MakeMyTrip (Myra), India
Outcome
Over 50,000 conversations a day (up from about 25,000 a quarter earlier); over 72 % rated good; about half of flight and hotel queries resolved without a human (the CEO, Q3 FY26 earnings call)
Effort
high
Risks
Self-reported figures; wrong answers on fares and refunds; language coverage.
How to adopt
Prerequisite: bookings and policies reachable through APIs and a multilingual content base. First step: launch for pre-sales questions, then booking and after-sales. Running: quality review of conversations.
How a room runs it
the ROOM for a travel agency's service desk: the owner, the sales lead, the after-sales lead and a travel teammate; the agent reads the bookings, fares and itineraries the agency already holds and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the quote, the booking and every refund or change; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the agency's booking engine and inbox wired to the room through connectors loaded on this pull, a travel-agency professional as the paid professional on the match, the hospitality pack's itinerary and after-sales task templates
An online travel agency rebuilds customer service as an agentic in-app assistant that checks bookings, adds extras and files refund requests against live booking and payment data
Where
ID, Southeast Asia
Segment
enterprise
Problem
support queues for status, changes and refunds
Technology
agentic assistant on cloud-hosted LLM services, live booking and payment data
Who uses it
tiket.com, Indonesia, with Microsoft
Outcome
75,000 interactions a month (as reported, not verified against a primary source)
Effort
high
Risks
Wrong refunds; the figure is unverified; payment-data security.
How to adopt
Prerequisite: live booking and payment data behind secure APIs. First step: an agentic in-app assistant for status, changes and refunds. Running: guardrails and human fall-back.
How a room runs it
the ROOM for a small agency or tour operator facing the same queue: the owner, the booking agent and a travel teammate (an agent definition with the booking and refund skills); the teammate reads the bookings the agency already holds and GENERATES the follow-ups as tasks (reconfirm the hotel, chase the supplier refund) with owners and due dates, and puts any refund or change above a threshold to the owner as a decision with a deadline; the humans keep the traveller relationship and the money; the receipt records each change and refund and who approved it
The engagement on top
a custom engagement: the agency's own booking system, supplier portals and WhatsApp inbox wired to the room through connectors loaded on this pull, a travel consultant as the paid professional, the tourism pack's booking and refund task templates
A travel retailer deploys AI agents that help advisors assemble complex multi-supplier holidays and answer customer questions instantly
Where
AU, Oceania
Segment
enterprise
Problem
assembling bespoke holidays from many supplier sources is slow for advisors
Technology
AI agents on a CRM platform, natural-language search
Who uses it
Flight Centre Global Leisure (Flight Centre Travel Group), Australia
Outcome
none stated
Effort
medium
Risks
Results stated without numbers; wrong supplier data; advisor over-reliance.
How to adopt
Prerequisite: a CRM with supplier data. First step: AI agents that help advisors assemble holidays. Running: advisors verify results.
How a room runs it
the ROOM for a travel advisor team: the advisors, the team lead and a trip-building teammate; the teammate reads the supplier options the agency already uses and GENERATES the quote work as plan steps (hold the flight, request the hotel, send the quote by Friday) with owners and due dates, and the advisor's sell decision as a proposal with a deadline; the advisor keeps the customer and the sale; the receipt records what was quoted, held and booked
The engagement on top
a custom engagement: the agency's own booking tools, supplier feeds and CRM wired to the room through connectors loaded on this pull, a senior travel consultant as the paid professional, the tourism pack's quoting task templates
A tourism ministry puts an AI travel assistant on the national tourism site that builds itineraries, recommends destinations and summarises attractions in several languages
Where
ID, Southeast Asia
Segment
public
Problem
visitors cluster at a few destinations and lack tailored guidance
Technology
AI assistant embedded in the national tourism website, interactive maps, multilingual summaries
Who uses it
Ministry of Tourism (Kemenparekraf), Indonesia, MAiA, launched by the tourism minister
Outcome
none stated
Effort
medium
Risks
No usage data; unstated data sources; outdated or wrong recommendations.
How to adopt
Prerequisite: a national tourism website with attraction content. First step: embed an AI itinerary assistant with multilingual summaries. Running: content upkeep.
How a room runs it
the ROOM for a regional tourism office or a destination's operators: the office head, the local guides and a destination teammate; the agent takes the itineraries visitors request and GENERATES the local tasks (guide availability, a homestay slot, a craft workshop) for the named operator with a due date, and the office's choices on promotion arrive as proposals with a deadline; the operators keep the hosting; the receipt records each booked experience and what the operator confirmed
The engagement on top
a custom engagement: the office's own listings, bookings and WhatsApp contact points wired to the room through connectors loaded on this pull, a tourism planner as the paid professional, the tourism pack's task templates
An online travel group launches an AI travel agent that asks three questions, builds a priced itinerary and lets the traveller edit and book every part in one tool
Where
CN, East Asia
Segment
enterprise
Problem
trip planning across flights, rail, lodging, restaurants and attractions is fragmented
Technology
AI itinerary agent with real-time prices, availability and verified listings
Who uses it
Trip.com Group, China, Trip.Planner
Outcome
none stated
Effort
high
Risks
No usage figures; stale prices; vendor concentration.
How to adopt
Prerequisite: real-time prices and availability feeds. First step: an itinerary agent that asks a few questions and builds a priced plan. Running: data feeds and agent quality monitoring.
How a room runs it
the ROOM for a group tour: the organiser, the group leader and a trip teammate; the agent GENERATES the plan as steps with owners and due dates (collect passports by the 10th, pay the deposit, book the transfers), the organiser's choices between options arrive as proposals with a deadline, and each traveller's share appears on the shared bill; the humans keep the choices and the payments; the receipt records what was booked and by whom
The engagement on top
a custom engagement: the operator's own supplier and payment systems wired to the room through connectors loaded on this pull, a tour planner as the paid professional, the tourism pack's group-trip task templates; travellers reached over WhatsApp
In-app AI chat for guests across the fleet, tested on one ship and extended to the newest ship, plus AI-driven personalised recommendations and pricing
Where
US, North America
Segment
enterprise
Problem
long onboard customer-service queues and generic offers
Technology
Royal Caribbean in-app chat and AI personalisation
Who uses it
Royal Caribbean Group, United States
Outcome
CEO: 35 percent increase in guest adoption of the new in-app chat, which led to a 20 percent reduction in the customer service line onboard (earnings call, Jan 2025)
Effort
medium
Risks
Wrong answers on safety or billing; guest data; over-aggressive dynamic pricing.
Next: the CEO named the development of industry-leading digital and AI capabilities as a key part of the future source
How to adopt
Test on one ship; measure adoption and queue length; extend to the fleet; keep a human line for complaints.
How a room runs it
the voyage room: the guest, the purser, the shore-excursion team and a concierge teammate; the agent answers and GENERATES the service tasks (a table, a repair, a refund) with due times, and each refund or upgrade becomes a proposal in the purser's Inbox; the receipt records the request, the offer and who approved
The engagement on top
connectors to the guest-services and booking systems, a guest-experience teammate, an escalation playbook
First airline collaboration with OpenAI: an upgraded customer assistant, a staff assistant and AI inside crew-scheduling and planning tools
Where
SG, Southeast Asia
Segment
enterprise
Problem
routine customer queries and complex crew and resource planning
Technology
OpenAI models; Singapore Airlines' virtual assistant and flight recommender
Who uses it
Singapore Airlines, Singapore
Outcome
Announced April 2025: a customer assistant upgrade first, then a staff assistant and OpenAI models in crew-scheduling tools; no numbers published
Effort
high
Risks
Safety-critical scheduling errors; regulatory compliance of crew rules; customer data handling.
Next: the staff assistant and the crew-planning integration are the announced later phases source
How to adopt
Upgrade the existing customer assistant first; add the staff tool next; put planning recommendations behind a human duty manager.
How a room runs it
the operations room: the duty manager, crew planners, the customer-care lead and a planning teammate; the agent drafts the schedule options within regulations and GENERATES the resolution tasks per team with due times, and each customer-affecting change becomes a proposal in the duty manager's Inbox; the receipt records the options, the constraints applied and the approver
The engagement on top
connectors to the reservation and crew-planning systems, an aviation-ops teammate, a rostering rule set
Company-wide ChatGPT Enterprise rollout with AI literacy programmes, an internal champion network and a centre of excellence across the airline and the wider group
Where
AE, Middle East
Segment
enterprise
Problem
scaling AI skills and use cases across a very large workforce
Technology
ChatGPT Enterprise (OpenAI)
Who uses it
Emirates Group, United Arab Emirates
Outcome
November 2025 agreement for a company-wide ChatGPT Enterprise rollout with literacy training, a champion network and a centre of excellence; no usage figures published
Effort
high
Risks
Shadow use; data residency; uneven adoption.
Next: joint sandboxes, an AI Centre of Excellence and early access to OpenAI research are the announced next steps source
How to adopt
Name champions per department; run sandboxes before production; track use cases in one register.
How a room runs it
the champions room: the department head, the AI champion, the compliance lead and a drafting teammate; the agent drafts the use-case brief and GENERATES the pilot tasks per department with dates, and each tool or data-access request becomes a proposal in the head's Inbox; the receipt records the use case, the data touched and the reviewer
The engagement on top
connectors to the group's identity and document systems, an AI-champion training track, a use-case template set
Conversational AI travel assistant that plans trips, books, and handles disruptions, including inside group chats
Where
US, North America
Segment
enterprise
Problem
fragmented trip planning and slow rebooking when plans change
Technology
Romie (Expedia's AI assistant, built on OpenAI models)
Who uses it
Expedia Group, United States
Outcome
Alpha release (May 2024) that plans, books and re-plans around weather and disruption; no performance figures published
Effort
medium
Risks
Wrong bookings; hallucinated availability; personal data in group chats.
Next: Expedia describes Romie as built to get more personal as travellers use it, with the alpha the first step source
How to adopt
Start with an alpha for a limited audience; keep booking and payment behind explicit approval; add group-chat planning once accuracy is proven.
How a room runs it
the trip room: the traveller, the companions, the travel agent and a planning teammate; the agent drafts the itinerary and GENERATES the booking and rebooking tasks per person with deadlines, and each change or paid add-on becomes a proposal in the organiser's Inbox before money moves; the receipt records the plan, the choices and who approved the spend
The engagement on top
connectors to the booking and loyalty systems, a travel-ops teammate, a disruption-playbook template set
A tour operator trials a ChatGPT-based assistant in its mobile app to help customers find holidays, live to half of UK app users in a phased rollout
Where
GB, Europe
Segment
enterprise
Problem
most package customers use the app before departure and bookings through the app are low
Technology
generative AI chatbot in a mobile app
Who uses it
TUI Group, United Kingdom
Outcome
More than 10,000 customers had used the ChatGPT function (TUI, as reported); app bookings in the UK doubled to 9 % but this is not attributed to the chatbot (the CEO)
Effort
medium
Risks
A trial, effect on bookings unproven; wrong suggestions; customer-data handling.
How to adopt
Prerequisite: a bookable catalogue and a mobile app. First step: trial the assistant on half of app users and compare conversion. Running: prompt and content review.
How a room runs it
the ROOM for a tour operator's digital team: the product lead, the app owner, the customer-care lead and a travel teammate; the agent reads the holiday catalogue and booking data the operator already holds and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the booking, the price and every customer-care promise; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the booking system and the app's content feed wired to the room through connectors loaded on this pull, a tour-operator professional as the paid professional on the match, the hospitality pack's itinerary task templates
Your own team builds it. Across the 10 adoptions here, 0 were low effort, 5 medium and 5 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Page undated; no GPS or IoT mention; driver monitoring raises privacy concerns.
How to adopt
Prerequisite: a fleet with vehicle monitoring. First step: adopt a fleet management system with a command centre. Running: staff in the command centre.
How a room runs it
the ROOM for a regional rental branch: the branch manager, the fleet coordinator, the drivers' supervisor and a fleet teammate; the agent reads the maintenance dates and trip records the company's fleet system already holds and GENERATES the week's tasks (service due on three units, replace a driver on Tuesday's airport run) with owners and due dates, and an out-of-envelope repair is a proposal in the manager's Inbox with a deadline; the humans keep the driving, the workshop and the spend above the envelope; the receipt records each service done
The engagement on top
a custom engagement: the company's fleet system, rental app and WhatsApp customer line wired to the room through connectors loaded on this pull, a fleet manager as the paid professional, the rental pack's task templates
Source: vendor · 2026 · vendor only; to corroborate
Airport rental counters scan every car at the gate with AI drive-through scanners to compare pickup and return condition and flag new damage
Where
US, North America
Segment
enterprise
Problem
manual damage checks are slow and miss small damage
Hertz, USA (airports starting with Atlanta), with UVeye; Sixt also reported using AI inspection
Outcome
none stated
Effort
high
Risks
Wrongful damage charges; scale and savings not stated; hardware cost.
How to adopt
Prerequisite: gate space and cameras. First step: install AI drive-through scanners at one airport. Running: human sign-off on damage claims.
How a room runs it
the ROOM for a rental fleet: the owner or fleet manager, the counter staff, the workshop and a fleet teammate (an agent definition with the rental skills); from the photos and condition notes the company already records, the agent GENERATES the turnaround tasks (inspect car B 1432, book the panel repair, hold the unit off the rota until Friday) with owners and due dates, and a damage charge a customer disputes goes to the manager's Inbox as a proposal with the evidence attached and a deadline; no charge is issued without a human sign-off; the receipt records the evidence, the decision and who made it
The engagement on top
a custom engagement: the company's own booking and fleet system and the inspection photo store wired to the room through connectors loaded on this pull, a fleet or workshop manager as the paid professional, the rental pack's handover and damage task templates
A homestay and short-stay rental operator centralises WhatsApp, Instagram, Messenger and Telegram guest chats in one shared inbox and automates routing, with booking details shown in the chat
Where
MY, Southeast Asia
Segment
SMB
Problem
guest messages arrive on many channels and assigning each by hand is slow
Technology
shared inbox, workflow automation, API link to the company's intranet
Who uses it
Nai Hospitality (homestay management), Malaysia, with respond.io
Outcome
89.7 % of contact assignments automated, 11,000+ conversations a month (vendor's figures) the company’s own figure
Effort
low
Risks
Vendor's figures; customer data across channels; platform dependence.
How to adopt
Prerequisite: accounts on the main messaging channels. First step: put them in one shared inbox and set routing rules. Running: staff monitor the queues.
How a room runs it
the ROOM for a villa or homestay manager: the owner, the cleaner and the guest-support lead and a rental teammate; the agent reads the booking calendar and guest messages the manager already has and GENERATES the turnover tasks (clean villa 3 after Tuesday's checkout, restock, meet the 14:00 arrival) for the named person with a due date, and the owner's choices (late checkout, a refund) arrive as proposals with a deadline; the humans keep the guest welcome and the keys; the receipt records each turnover done
The engagement on top
a custom engagement: the operator's booking calendar or channel manager and the WhatsApp number wired to the room through connectors loaded on this pull (the WhatsApp channel is built), a property manager as the paid professional, the rental pack's turnover task templates
Source: vendor · 2026 · vendor only; to corroborate
A rental group tests an AI vehicle scanner at its sites to digitise and automate check-in and check-out inspections, run by a start-up
Where
FR, Europe
Segment
enterprise
Problem
manual inspections are slow and disputed
Technology
autonomous AI vehicle scanner
Who uses it
Europcar Mobility Group with Tchek, France
Outcome
none stated
Effort
high
Risks
A test with no results; wrongful damage charges; hardware cost.
How to adopt
Prerequisite: site space and a scanner vendor. First step: test at one site against manual inspections. Running: human review of flagged damage.
How a room runs it
the ROOM for a rental branch: the branch manager, the counter staff, the workshop lead and a fleet teammate; the agent reads the scan images and condition notes the branch already records and GENERATES the checklist as plan steps with owners and due dates, and files a decision in the responsible lead's Inbox as a proposal with a deadline whenever an answer or exception needs a human line; the humans keep the final damage decision and any charge; the receipt records what was signed off, by whom and when
The engagement on top
a custom engagement: the booking and fleet system and the scan photo store wired to the room through connectors loaded on this pull, a fleet or workshop manager as the paid professional on the match, the rental pack's handover and damage task templates
Your own team builds it. Across the 4 adoptions here, 1 were low effort, 0 medium and 3 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs a data pool on borrowers from the group and its partners and a scoring model trained on it. Start with one micro-lending product and compare model scores with loan officers' decisions. Running it takes data engineers, model monitoring and credit-policy sign-off.
AI underwrites MSME loans of up to Rs 5 crore and large-language-model reading of low-value cheques with straight-through processing, plus AI early-warning signals; a control unit samples the AI-processed cheques
Where
IN, South Asia
Segment
enterprise (public-sector bank)
Problem
slow MSME credit decisions and manual cheque checking at national scale
Technology
AI credit underwriting, LLM document reading, straight-through processing, early-warning models
Who uses it
State Bank of India (SBI), India
Outcome
About Rs 1 lakh crore of MSME loans underwritten with AI in FY26 (stated by SBI's Managing Director; cost-to-income impact not yet quantified)
Effort
high
Risks
Credit-model bias and drift; the figure is the amount underwritten, not a performance or default measure; automation errors in cheque reading.
How to adopt
Needs digitised customer and transaction data, a model validated against past credit outcomes, a sampling review by a control unit, and regulator engagement; start with one loan product and one cheque band.
How a room runs it
the ROOM for a lending desk: the relationship manager, the credit officer, the control-risk reviewer and a credit teammate; the agent assembles the file from what the bank already holds and GENERATES the officer's tasks (verify the turnover figures, request the missing statement by Friday) as plan steps, and the lending recommendation as a decision in the credit officer's Inbox with a deadline; the humans keep the credit decision, the sampling review and the sanction; the receipt records inputs, model output and signer. Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
an engagement for a lender: the loan origination and document systems wired through connectors loaded on this pull, the sampling-review routine set up with the control-risk team, a credit-risk professional as the paid match
Group-wide generative AI on an enterprise platform with a multi-model approach: most staff use it, 87 use cases approved, AI-enabled recommendations feed client interactions, and coding assistants are used by engineers
Where
ZA, Africa
Segment
enterprise (pan-African bank group)
Problem
a large bank must raise productivity and personalise service across many markets without losing control of risk
Technology
enterprise generative AI platform on a cloud model service (multi-model), AI coding tools, recommendation models
Who uses it
Standard Bank Group (South Africa), CEO Sim Tshabalala as quoted
Outcome
72% of employees are active generative-AI users; 87 approved use cases; AI-enabled recommendations supported 10 million+ personalised client interactions (bank CEO, reported by ITWeb); no productivity or cost figure on the page
Effort
high
Risks
Usage is not value: the page gives no hours or cost saved; model and data-protection risk across markets; dependence on one cloud model service.
How to adopt
Stand up one governed enterprise AI platform, publish a use-case register with a risk review per entry, train staff in waves, and measure active use per business line before adding customer-facing cases.
How a room runs it
the ROOM for a bank's AI programme office: the head of digital, a use-case owner per business line, the risk lead and an AI-programme teammate (an agent definition with the bank's approved use-case register); the agent teammate reads the register and usage reports the bank already holds and GENERATES each business line's next-step tasks (finish the risk review for use case 41 by the 15th) as plan steps with owners and due dates, and each new use case for the risk lead (approve / hold, with the evidence attached) as a decision block in the Inbox with a deadline; what the humans keep: the risk sign-off, model approval and every customer-facing decision; the receipt records every finished task and every decision, and the client receives the reports and gives direction there Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the use-case register, the usage analytics export, the model-risk tracker), an AI-governance specialist joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Banks put AI into credit analysis and staff assistants: Bradesco's BIA assistants and AI-supported credit analysis, Banco do Brasil with 2,000+ AI solutions in production including machine learning, generative AI and autonomous agents
Where
BR, Latin America
Segment
enterprise (two of Brazil's largest banks)
Problem
credit decisions and internal work at scale need faster, more consistent analysis
Bradesco and Banco do Brasil (Brazil); ranking by Evident covering 20 Latin American institutions
Outcome
Bradesco: AI-supported credit analysis reportedly added over R$1 billion in revenue; Banco do Brasil: 2,000+ AI solutions in production by Q1 (figures as reported from the Evident ranking; measurer not stated)
Effort
high
Risks
Revenue attribution is unaudited on the page; credit-model bias and explainability duties; counting 'solutions' says little about value.
How to adopt
Start with one credit product, keep the model beside the existing scorecard, compare decisions for a quarter, and give analysts an override with a reason.
How a room runs it
the ROOM for a bank's credit-analytics team: the head of credit, two analysts, the model-risk lead and a credit-analysis teammate (an agent definition with the bank's credit policy); the agent teammate reads the applications and model outputs the bank already holds and GENERATES the analysts' review tasks (re-check these 30 flagged files by Friday) as plan steps with owners and due dates, and each policy exception for the head of credit (grant / decline, with the evidence attached) as a decision block in the Inbox with a deadline; what the humans keep: the credit decision, the policy and the model approval; the receipt records every finished task and every decision, and the client receives the reports and gives direction there Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the credit decision log, the model-monitoring export, the policy documents), a credit-risk specialist joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
A bank-wide ChatGPT Enterprise rollout, extended to the whole workforce after an earlier pilot of a few thousand accounts, as part of a multi-year partnership with the model supplier
Where
ES, Europe
Segment
enterprise (global bank)
Problem
routine tasks consume staff time across the organisation
Technology
enterprise large-language-model assistant licence, single sign-on
Who uses it
BBVA, Spain
Outcome
Employees reportedly saved close to three hours a week on routine tasks and over 80% use the tools daily (as reported by Fintech Global; measurer not named)
Effort
medium
Risks
The time-saving figure has no named measurer; uneven use across departments; data-protection and model-risk review for customer data.
How to adopt
Needs an enterprise licence with data-protection terms, single sign-on and a champion in each department; start with a pilot group, measure time on a few named tasks, then widen.
How a room runs it
the ROOM for a branch-support or back-office team: the team lead, the analysts, the risk officer and an operations teammate; the agent takes the team's routine requests and GENERATES tasks for named analysts (reconcile the exceptions list, draft the reply to the customer query) with due dates, and the lead's sign-off on policy-sensitive replies as a decision in the lead's Inbox; the humans keep the judgement and every customer-facing send; the receipt records request, draft, reviewer and date.
The engagement on top
a rollout engagement: the bank's own systems wired in through connectors loaded on this pull, the use cases chosen with each department head, an operations professional as the paid match
GPT-4-based assistant and meeting-notes tool for wealth advisors over the firm's research library
Where
US, North America
Segment
enterprise (global bank)
Problem
advisors spend time searching research and writing notes
Technology
AI @ Morgan Stanley Assistant and Debrief (OpenAI)
Who uses it
Morgan Stanley Wealth Management, United States
Outcome
Advisor use of the assistant reached 98% of advisor teams, about nine months after the September 2023 rollout (Morgan Stanley figure reported June 2024)
Effort
high
Risks
Hallucinated research; client confidentiality; advisor suitability duties stay with the human.
Next: Debrief joined the assistant as the firm widened its suite for advisors source
How to adopt
Curate the document set first, test answers with human experts, add note-taking only with client consent.
How a room runs it
the client-meeting room: the advisor, the assistant and the client's file; the agent drafts the meeting note, action items and follow-up email, saves to the CRM only with the client's approval, and any recommendation becomes a proposal the advisor reviews; the receipt records the sources used and the approval
The engagement on top
connectors to the research library and the CRM, a consent rule for recording, a note template
Virtual financial assistant for retail clients, scaled over seven years, with an employee version for internal help
Where
US, North America
Segment
enterprise (global bank)
Problem
clients wait for simple banking answers and actions
Technology
Erica (Bank of America in-house assistant)
Who uses it
Bank of America, United States
Outcome
More than 3 billion client interactions since 2018, over 58 million a month, nearly 50 million users and 18.7 million hours of conversation (Bank of America, Aug 2025) the company’s own figure
Effort
high
Risks
Account data exposure; regulatory conduct rules; over-reliance on scripted flows.
Next: the bank plans to expand Erica's abilities over the next year and widen Erica for Employees source
How to adopt
Begin with read-only questions, add transactions behind a gate, then widen to staff-facing help.
How a room runs it
the client room: the client, the assistant and the banker; the assistant answers balance, spending and bill questions within clearance, and anything consequential (a limit change, a dispute) becomes a proposal for the banker's Inbox; masked fields show as the literal mask and the receipt records what was shown
The engagement on top
connectors to the core banking and card systems, a clearance and masking policy pack, a handover rule to the banker
AI assistant handling customer-service chats in 35+ languages, with a live agent always one tap away
Where
SE, Europe
Segment
enterprise (fintech)
Problem
long waits and repeat contacts in payments and disputes
Technology
OpenAI-based assistant built by the company
Who uses it
Klarna (Sweden), 23 markets
Outcome
In its first month the assistant handled 2.3 million conversations (two thirds of chats), did the work of about 700 full-time agents, cut repeat inquiries by 25% and resolution time from 11 minutes to under 2 (Klarna, Feb 2024) the company’s own figure
Effort
medium
Risks
Quality drift on edge cases; customers who want a person; money-moving answers need a gate.
Next: Klarna said new features were already in the pipeline source
How to adopt
Start with one market and one intent family; keep a one-tap human handover; measure repeat contacts, not only speed.
How a room runs it
the support room: customer, assistant and a human agent; the agent answers refund, payment and dispute questions from policy, and anything that moves money (a refund above the limit, a dispute ruling) becomes a proposal in the agent lead's Inbox with a deadline; the receipt records the answer, the policy cited and who approved
The engagement on top
connectors to the helpdesk and the payments ledger, a policy pack for refund limits, a human-handover rule set
Government platform for village cooperatives: a registry dashboard, a per-cooperative profile (SIMKOPDES) and a member mobile app for savings, loans and financial information, with planned links to OSS and SIKP
Where
ID, Southeast Asia
Segment
public (cooperatives)
Problem
81,000+ new village cooperatives need registration, profiles and member services with little accounting or digital experience
Technology
dashboard, microsite, mobile app, planned government-system interoperability
Who uses it
Ministry of Cooperatives (Kemenkop), Kopdes Merah Putih, Indonesia
Outcome
Ministry of Cooperatives, as of 2025-09-15: 81,485 Kopdes Merah Putih cooperatives are legal entities, 42,349 hold a SIMKOPDES account, 9,317 updated their profile (programme counts, not an AI result)
Effort
high
Risks
Uneven data quality across thousands of small cooperatives; connectivity gaps; planned links to other systems not yet live. Not an AI deployment.
How to adopt
Needs a national or regional registry of cooperatives and a member app with identity checks. Start with registering cooperatives and a basic profile, then savings and loan views for members. Running it takes programme staff, connectivity and training of cooperative managers.
How a room runs it
the ROOM for a cooperative: the chair, the treasurer, the secretary and a cooperative teammate (an agent definition with cooperative bookkeeping and reporting skills); the agent reads the member ledger the cooperative already keeps and GENERATES the treasurer's monthly tasks (post the savings, reconcile loans, prepare the annual-meeting figures) with due dates, and each loan request as a decision in the chair's Inbox with a deadline; the members' board keeps the vote and the signature on every loan; the receipt records the ledger lines behind each decision Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the cooperative's existing savings-and-loan app or spreadsheets loaded on this pull; WhatsApp reminders to members; an accountant as the paid match; the pack's cooperative templates
AI across back and front office: credit scoring, anti-fraud and risk analysis on structured and unstructured data, framed as support for staff
Where
ID, Southeast Asia
Segment
enterprise
Problem
manual risk review and fraud pattern-spotting do not scale with volume
Technology
machine learning on big data, credit scoring, fraud and pattern detection
Who uses it
Bank Rakyat Indonesia (BRI), director of digital and IT, Indonesia
Outcome
none stated
Effort
high
Risks
Model error on credit or fraud calls; staff concern about displacement; regulator expectations on explainability.
How to adopt
Needs digitised credit and transaction records and a clear policy that AI supports staff decisions. Start with fraud screening or credit scoring on one portfolio, with staff reviewing flagged cases. Running it needs model governance and retraining.
How a room runs it
the same kind of room for a risk and compliance team: the head of risk, two analysts and a fraud-watch teammate; the agent reads the cases the bank already flags and GENERATES review tasks for the analysts with due dates and the evidence attached, and a decision for the head of risk (hold, release, escalate) with a deadline; the analysts keep the judgement, the head keeps the block or release; the receipt records the evidence read, the proposal and the signer Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the bank's case-management and transaction-monitoring systems loaded on this pull; a risk or compliance professional as the paid match; the pack's review task templates
An in-house generative AI assistant, LLM Suite, given to staff in stages for writing, idea generation, Excel problem-solving and summarising documents; staff may not use third-party consumer chatbots for work
Where
US, North America
Segment
enterprise (global bank)
Problem
staff time lost on drafting and summarising; client data must not leave the bank's servers
Staff may paste confidential data into the tool if controls are weak; output errors in client documents; the access count is a press report, not a usage or benefit measure.
How to adopt
Needs a private model deployment so client data stays inside the bank, an acceptable-use policy and a staged rollout by division; a first step is one team with document-heavy work. Takes security review and training.
How a room runs it
the ROOM for a banking team: the relationship manager, the team head, the compliance officer and a banking-assistant teammate; the agent reads the documents the team already holds and GENERATES the week's tasks (summarise the credit file by Tuesday, draft the client note for review) as plan steps with owners and due dates, and the head's approval of anything client-facing as a decision with a deadline; what the humans keep: judgement, sign-off, anything sent to a client; the receipt records the sources used and who approved. Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
an engagement for a bank or lender: the client's own document and case systems wired through connectors loaded on this pull, the usage policy written with the client's compliance lead, a banking professional as the paid match
Generative-AI model that scores card transactions in real time using relationships among the parties to a payment
Where
US, North America
Segment
enterprise (payment network)
Problem
card fraud and false declines
Technology
Decision Intelligence Pro (Mastercard in-house)
Who uses it
Mastercard, United States, serving issuing banks worldwide
Outcome
Mastercard said detection rates are expected to improve by 20% on average and by as much as 300% in some cases, with scoring in about 50 milliseconds (CNBC, Feb 2024; early modelling)
Effort
high
Risks
Model drift; false declines hurt customers; the headline gain is a best case.
Next: Mastercard expects banks to cut the cost of assessing illegitimate transactions by as much as 20% source
How to adopt
Run the model in shadow beside the current scorer, compare on held-out cases, then switch thresholds with human sign-off.
How a room runs it
the fraud-ops room: analysts, the risk lead and the issuer's case owner; the agent explains why a transaction scored high and drafts the case file, and any block or customer-contact decision becomes a proposal with a deadline in the analyst's Inbox; the receipt records the score, the features and the reviewer
The engagement on top
connectors to the case-management tool and the transaction feed, a review-threshold policy pack, a case-note template
In-house AI platform (BRIBrain) pulling data from BRI Group and partner ecosystems to score micro-borrowers and speed loan approvals
Where
ID, Southeast Asia
Segment
enterprise
Problem
slow, thin-file credit decisions for micro, small and medium enterprises
Technology
proprietary machine learning, group and partner data streams
Who uses it
Bank Rakyat Indonesia (BRI), Indonesia
Outcome
none stated
Effort
high
Risks
Biased or opaque scoring that wrongly excludes thin-file borrowers; privacy and consent over partner data.
How to adopt
Needs a data pool on borrowers from the group and its partners and a scoring model trained on it. Start with one micro-lending product and compare model scores with loan officers' decisions. Running it takes data engineers, model monitoring and credit-policy sign-off.
How a room runs it
the ROOM for a branch's lending desk: the branch head, the credit officers, the risk officer and a credit-desk teammate (an agent definition with the bank's lending skills); the agent reads the application file and the data the bank already holds and GENERATES the week's tasks for the credit officer (collect the missing document, visit the borrower by Friday) as plan steps with owners and due dates, and lays each recommendation as a decision block in the approver's Inbox with a deadline; the officer keeps the borrower visit, the approver keeps the sign-off on the loan and every amount above the envelope; the receipt records what the agent read, what it proposed and who signed Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
a custom engagement: the bank's own loan-origination and core-banking systems stay theirs, wired to the room through connectors loaded on this pull; a credit-risk professional joins as the paid match; the sector pack carries the lending task templates and the document checklists; borrower follow-ups go over WhatsApp
DBS Bank, Singapore, with Taiwan and Hong Kong next
Outcome
Rolled out to 500 customer service officers in Singapore handling queries from over 250,000 customers a month; pilot accuracy nearly 100% and about 90% of pilot officers reported a positive effect; call handling time expected to fall by up to 20% (DBS, Jul 2024) the company’s own figure
Effort
medium
Risks
Transcription errors in names and amounts; customer data in prompts; projected savings not yet proven.
Next: progressive rollout to Taiwan and Hong Kong over the following 12 months and over 20 generative AI use cases in build source
How to adopt
Pilot with a small officer group, keep the officer in the loop, measure handling time after the pilot.
How a room runs it
the service room: the officer, the customer's case and a drafting teammate; the agent transcribes, finds the answer in the knowledge base and pre-fills the service request, and any request that changes the account becomes a proposal the officer approves; the receipt records the summary, the sources and the officer
The engagement on top
connectors to the telephony transcript feed and the knowledge base, a summary template, an officer-approval rule
Gradual AI adoption through proof of concept: document validation and admin, data gathering for credit decisions, planned verification and chatbot
Where
ID, Southeast Asia
Segment
mid (multifinance)
Problem
routine validation and paperwork slow financing approvals
Technology
AI document validation, data analysis, chatbot or virtual assistant (planned or underway)
Who uses it
Mandala Finance, Indonesia
Outcome
none stated
Effort
medium
Risks
Source does not separate live from planned, so benefit is unproven; document-extraction errors; customer data privacy.
How to adopt
Needs scanned or digital loan documents and a defined set of checks. Start with a proof of concept on document validation and data gathering, keeping credit decisions with staff. Running it needs an integration to the loan system and an owner for exceptions.
How a room runs it
the ROOM for a multifinance branch: the branch manager, the survey and collection staff and a financing teammate; the agent checks each application's documents against the checklist and GENERATES tasks for the surveyor (verify the address, collect the missing paper by Thursday) and a decision for the manager (approve, decline, ask more) with a deadline; the managing director's rule that AI does not replace workers is the room's rule: people do the survey and sign the financing; the receipt records documents checked and the signer Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the multifinance company's origination and receivables systems loaded on this pull; a credit professional as the paid match; the pack's multifinance templates
In-house text-generating AI (Wiz Chat) open to all employees in Japan, with search for administrative procedures and drafting of credit approval documents in proof of concept
Where
JP, East Asia
Segment
enterprise (global bank)
Problem
drafting and procedure lookups consume time; credit approval documents take 1-2 hours
Technology
in-house generative AI chat, retrieval over internal procedures, document drafting
Who uses it
Mizuho Financial Group (Mizuho Bank), Japan
Outcome
none stated (a target of about 10 minutes for credit approval drafting, from 1 to 2 hours, is a goal not a result)
Effort
high
Risks
Draft errors reaching an approval; the figure is a projection; employees may over-trust drafts.
How to adopt
Needs an in-house or private model, a library of the bank's own procedures and approval templates, and a proof of concept with one credit team before general release.
How a room runs it
the ROOM for a credit team: the relationship manager, the credit analyst, the approver and a credit-drafting teammate; the agent drafts from the file and GENERATES the analyst's tasks (check the figures in section 3, attach the guarantee by Wednesday) as plan steps, and the draft approval as a decision in the approver's Inbox with a deadline; the humans keep the credit judgement and the approval; the receipt records the draft, the edits and the approver. Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
an engagement for a bank: the client's own credit and procedure systems wired through connectors loaded on this pull, templates built from the client's approval documents, a credit professional as the paid match
BPR industry body urging rural banks to move from digitalisation to AI-assisted credit scoring that combines their own data with demographic and regional data, measured and prudent
Where
ID, Southeast Asia
Segment
SMB and mid (rural banks, BPR)
Problem
BPRs serve MSMEs with thin data and are still at the digitalisation stage; AI adds cyber risk to credit, operational and liquidity risk
Technology
credit scoring on own plus secondary demographic and regional data
Who uses it
BPR sector, via Perbarindo (association chair Tedy Alamsyah), Indonesia
Outcome
none stated
Effort
medium
Risks
Thin data at small lenders yields weak models; compliance with the credit-scoring regulation (POJK 19/2025 per the article); cost for small institutions.
How to adopt
Needs the lender's own loan history in a usable digital form plus regional and demographic data. Start with a pilot score alongside the existing credit committee on a small loan segment. Running it needs a shared or outsourced scoring service and staff trained to read scores.
How a room runs it
the ROOM for a small BPR: the director, the account officers and the compliance head, with a lending teammate; the agent reads the loan files the BPR already keeps and GENERATES the account officers' field-visit and document tasks with due dates, and a decision per file in the director's Inbox (approve, decline, ask for more) with a deadline; the director signs every loan, the officers do the visits; budgets degrade before they block; the receipt records each file's reasoning and signer Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the BPR's core-banking and loan system loaded on this pull; WhatsApp follow-up with borrowers; a credit professional as the paid match; the pack's BPR lending templates
Your own team builds it. Across the 16 adoptions here, 0 were low effort, 5 medium and 11 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs an audit platform holding engagement data in structured form and auditor training. Start with narrow tasks such as expense vouching or searches for unrecorded liabilities, with auditors reviewing every output. Running it needs agent monitoring and quality review.
The UK audit regulator publishes its first guidance on generative and agentic AI in audit, placing accountability on the auditor, while large firms add AI tools to their audit platforms and test agents that direct other models on routine audit tasks
Where
GB, Europe
Segment
regulator and Big Four firms
Problem
audit teams adopting AI need a rule for who answers for the output
Technology
audit platforms with generative AI, orchestration agents, regulator guidance
Who uses it
Financial Reporting Council (FRC), EY and KPMG, United Kingdom
Outcome
none stated
Effort
medium
Risks
Responsibility stays with the auditor; the trade summary may differ from the guidance text; over-reliance on drafted output.
How to adopt
Read the regulator's guidance, define which audit steps may use AI, require human review and record keeping of AI output, pilot on one engagement type, and train staff.
How a room runs it
the ROOM for an audit engagement: the engagement partner, the manager, the senior and an audit teammate; the agent drafts working-paper entries and GENERATES the senior's tasks (re-perform the sample, document the evidence by Thursday) as plan steps, and the manager's acceptance of any AI-drafted conclusion as a decision with a deadline; the humans keep the judgement and the opinion; the receipt records what the AI did, who reviewed it and when, which is what the regulator asks of the auditor. Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
an engagement for a firm: the firm's own audit platform and working-paper system wired through connectors loaded on this pull, an AI-use protocol written to the regulator's guidance, an audit-methodology professional as the paid match
A suite of AI agents for auditors worldwide: media-scanning for emerging threats, first drafts of staff memos, autonomous agents on manual data work and coordinating agents for disclosure requirements, on an upgraded audit platform
Where
multi, global
Segment
enterprise (global audit firm)
Problem
manual data work and slow client response in audits
Technology
AI agents, upgraded global audit platform
Who uses it
Deloitte (global audit), multiple countries
Outcome
none stated (the $3 billion, five-year commitment is an investment, not a result)
Effort
high
Risks
Accountability stays with the auditor; draft quality; a paywalled news source.
How to adopt
Needs an audit platform that can host agents, review rules for AI-drafted work papers, and training; start with one agent on one audit step.
How a room runs it
the ROOM for an audit engagement: the partner, the manager, the audit staff and an audit teammate; the agent sifts transactions and documents and GENERATES staff tasks (review the flagged items, finish the memo draft by Friday) as plan steps, and the manager's acceptance of AI-drafted documentation as a decision before it enters the work papers; the humans keep the opinion and the right to set the AI aside; the receipt records agent output, reviewer and date. Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
an engagement for a firm: the firm's platform and client data wired through connectors loaded on this pull, review rules set with the quality leader, an audit-methodology professional as the paid match
AI document processing for accounting firms on Xero: invoices, receipts and statements extracted, GL coded, bank-feed reconciled, with approval routing and multi-client operation
Where
ID, Southeast Asia
Segment
SMB and mid (accounting firms)
Problem
client paper and email drive manual data entry across many entities
Technology
AI data extraction, automated GL tagging, bank feeds, approval routing, Xero sync
Who uses it
fileAI (fileLedger) for accounting firms on the Xero marketplace (Indonesia listing)
Outcome
Marketplace listing shows 37 Xero user reviews rated 4.8 out of 5 (the supplier's listing; no time-saved or accuracy figure) the company’s own figure
Effort
low
Risks
Mis-coded entries reaching client books; vendor lock-in to one ledger; client data handled by a third party. Year is the read date (page undated).
How to adopt
Needs a cloud ledger (the pattern is built on Xero) and a stream of invoices, receipts and statements. Switch on the extraction add-on for one client, check coding against the accountant's own, then add approval routing. Running it needs a reviewer for exceptions.
How a room runs it
the ROOM for a bookkeeping firm: the firm owner, the accountants and a client-books teammate; the agent reads each client's incoming documents and GENERATES the accountants' weekly tasks per client (code these 60 items, chase two missing receipts by Thursday), and the month-end close as a decision for the owner with a deadline; the accountants keep review and filing, the client keeps the approvals; the receipt records every posting and who approved Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the firm's ledger software and client mailboxes loaded on this pull; WhatsApp document intake; an accountant as the paid match; the pack's month-end templates
Source: vendor · 2026 · vendor only; to corroborate
AI agents deployed in the firm's audit platform to its auditors for expense vouching and searches for unrecorded liabilities, with a disclosure-checklist engine
Where
US, North America
Segment
enterprise (global audit firm)
Problem
manual substantive testing consumes audit hours
Technology
AI agents inside an audit platform, financial-report analyser
Who uses it
KPMG (KPMG Clara), 95,000+ auditors, global
Outcome
KPMG's own figures: more than 95,000 auditors to use the agents; more than 2.19 million Audit Chat conversations so far; no audit-quality or hours-saved figure the company’s own figure
Effort
high
Risks
Audit quality and professional liability if an agent misses an exception; client confidentiality; the source is the firm's own announcement.
How to adopt
Needs an audit platform holding engagement data in structured form and auditor training. Start with narrow tasks such as expense vouching or searches for unrecorded liabilities, with auditors reviewing every output. Running it needs agent monitoring and quality review.
How a room runs it
the ROOM for an engagement: the partner, the manager, the seniors and an audit-support teammate (an agent definition with the firm's audit programme); the agent reads the client files the firm already holds and GENERATES the seniors' testing tasks (vouch these 40 expenses by Wednesday) as plan steps with owners and due dates, and each exception as a decision for the manager with a deadline; the auditors keep the judgement and the opinion, the partner signs; the receipt records the sample, the evidence and who concluded Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the firm's audit workpaper tool and the client's ledger export loaded on this pull; a senior auditor as the paid match; the pack's engagement templates and PBC lists
Source: vendor · 2025 · vendor only; to corroborate
Uneven AI adoption: Big 4 in a developmental phase, smaller firms limited by funds, skills and regulatory uncertainty
Where
ID, Southeast Asia
Segment
SMB and mid (Indonesian audit firms)
Problem
small audit firms lack accessible tools, rules and training
Technology
AI in financial-statement audit (tools not named in the study)
Who uses it
eight audit firms interviewed, Indonesia
Outcome
none stated
Effort
medium
Risks
Cost and skills gap for small firms; regulatory uncertainty about reliance on AI output; small interview sample, so adoption levels are unknown.
How to adopt
Needs audit working papers in digital form and staff with data skills. Start with one repetitive audit step (sampling, reconciliation) using an off-the-shelf tool. Smaller firms should budget for training and a supervised first season.
How a room runs it
the ROOM for a small firm's engagement: the partner, one or two staff and an audit-support teammate; the agent works from the client's export and GENERATES the staff's tasks (tick and tie the bank confirmations, request the missing schedules by Friday) with owners and due dates, and the open items as decisions for the partner with a deadline; the firm does not need to buy a platform: the room is the tool; the partner signs the opinion, the staff do the client contact Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the client's accounting software and the firm's document folder loaded on this pull; WhatsApp requests to the client's finance staff; an auditor as the paid match; the pack's small-firm templates
Your own team builds it. Across the 5 adoptions here, 1 were low effort, 2 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs historical claims data with known fraud outcomes and a claims system to feed scores into. Start with auto claims, routing flagged claims to investigators. Running it needs investigators and regular recalibration.
AI service representatives and an AI assistant (Express Service) handle most customer service, with quick-claim settlement and smart fraud detection in claims
Where
CN, East Asia
Segment
enterprise (insurance group)
Problem
service volume and claims cost at group scale
Technology
AI service agents, quick-claim automation, fraud detection models
Who uses it
Ping An Insurance Group, China
Outcome
1H 2026, Ping An's own figures: AI handled 81% of customer service volume; 59% of life insurance claims settled via quick claim; claims savings from smart fraud detection up 10.4% year on year to RMB 7.11 billion the company’s own figure
Effort
high
Risks
Group-scale figures are the company's own; exceptions routed badly; customer data protection.
How to adopt
Needs years of service and claims data, a large in-house AI team and strong escalation to humans; most insurers would start with one service channel and one claim type.
How a room runs it
the ROOM for a service and claims team: the service lead, the claims handler, the fraud analyst and a service teammate; the agent answers routine cases and GENERATES the handler's tasks for the exceptions (call back the customer, request the missing document by Monday) as plan steps, and the settlement of non-routine claims as a decision in the claims head's Inbox with a deadline; the humans keep judgement on disputed claims and every payout sign-off; the receipt records the case, the evidence and the signer. Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
an engagement for an insurer: the client's service and claims systems wired through connectors loaded on this pull, escalation rules set with the service lead, a claims professional as the paid match
Source: vendor · 2026 · vendor only; to corroborate
AI across motor claims routing and liability assessment, extended in 2025-26 to AI summaries of GP reports for life and critical-illness underwriting
Where
GB, Europe
Segment
enterprise (insurer)
Problem
slow claims and underwriting reviews of long documents
Technology
Aviva in-house AI and generative AI models
Who uses it
Aviva, United Kingdom
Outcome
AI in motor claims saved more than GBP 60m in 2024; the average time to assess liability on complex motor claims fell by 23 days; the underwriting summary tool halved review time per application (Aviva, as reported Mar 2026)
Effort
high
Risks
Missed clinical detail in a summary; unfair underwriting outcomes; regulator expectations on explainability.
Next: Aviva is moving the same summaries to income protection policies source
How to adopt
Start with summarising for the human decision-maker, test heavily, then widen the document types.
How a room runs it
the claims room: the handler, the claimant's file and a drafting teammate; the agent summarises long medical or loss documents and proposes a routing, each decision becomes a proposal the underwriter or handler approves with a deadline; the receipt records the summary, the source pages and the person who decided
The engagement on top
connectors to the claims system and document store, a summary template per document type, an underwriter-approval rule
AI guides policyholders to photograph the vehicle; counts scratches and dents and determines the coverage type for the survey step
Where
ID, Southeast Asia
Segment
mid
Problem
vehicle survey needs a visit to an agent or the insurer, slowing claims
Technology
AI image analysis of customer photos
Who uses it
Zurich Asuransi Indonesia, Indonesia
Outcome
none stated
Effort
medium
Risks
Wrong damage or coverage classification; photo fraud; customer image privacy.
How to adopt
Needs a customer app with a photo flow and a damage-detection model trained on vehicle images. Start with motor claims, with a surveyor reviewing the output. Running it needs integration to the claims system.
How a room runs it
the ROOM for a motor claim: the claimant (invited), the adjuster, the claims head and a claims teammate; the claimant sends photos over WhatsApp; the agent reads what the insurer's tools extract and GENERATES the adjuster's tasks (confirm the damage list, order the repair estimate by Tuesday), and the settlement as a decision in the claims head's Inbox with a deadline; the adjuster keeps judgement of the damage, the head signs the payout; the receipt records photos, extraction, estimate and signer Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the insurer's claims system and its photo-assessment tool loaded on this pull; WhatsApp is the claimant's door; a claims adjuster as the paid match
AI vehicle-damage assessment from images for underwriting and claims, with a bilingual guide and real-time claim visibility
Where
ID, Southeast Asia
Segment
SMB and mid (digital insurer)
Problem
slow underwriting and claims evaluation, fraud, operating cost
Technology
AI image analysis, interactive customer guide
Who uses it
Oona Insurance with CamCom, Indonesia
Outcome
none stated
Effort
medium
Risks
Assessment errors leading to disputes; dependence on one technology partner; image data privacy.
How to adopt
Needs an image-based damage assessment service connected to underwriting and claims flows. Start with vehicle inspection at policy purchase, then claims. Running it needs a partner or model provider and a review path for disputes.
How a room runs it
the ROOM for a small insurer or broker's motor book: the underwriter, the claims officer, the broker and an underwriting-and-claims teammate; the agent reads the photos and policy data on file and GENERATES the underwriter's referral tasks and the claims officer's follow-ups with due dates, and each quote or payout above the envelope as a decision for a named signer with a deadline; the customer can see the status in the room; the receipt records the assessment and the signer Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the insurer or broker's policy and claims systems loaded on this pull; WhatsApp for the customer; an underwriter or adjuster as the paid match; the pack's motor templates
A generative AI platform handles incoming customer emails and texts for a digital insurer, over 30% with no human intervention
Where
US, North America
Segment
SMB and mid (digital insurer)
Problem
service workload as the book grows
Technology
generative AI customer-service platform, AI claims agent
Who uses it
Lemonade, United States
Outcome
More than 30% of incoming email and text interactions handled with no human intervention (Lemonade's co-CEO on the Q2 2024 earnings call, reported by PYMNTS)
Effort
medium
Risks
Wrong answers on coverage; disputes where escalation fails; the figure is the company's own.
How to adopt
Needs a digital channel for customer messages, a knowledge base of policy rules and a handover path to staff; start with the most repetitive message types and measure the share closed unaided.
How a room runs it
the ROOM for a service desk: the service lead, the claims handler and a service teammate; the agent answers routine messages and GENERATES the handler's tasks for what it cannot close (call the policyholder back, collect the missing photo by Wednesday) as plan steps, and a goodwill payment or disputed claim as a decision in the claims head's Inbox with a deadline; the humans keep disputes and every payout; the receipt records message, answer, escalation and signer. Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
an engagement for an insurer: the customer-message channels and claims system wired through connectors loaded on this pull, an escalation policy written with the service lead, a claims or service professional as the paid match
Fraud-detection software applied to auto claims, meant also to speed legitimate claims
Where
ID, Southeast Asia
Segment
mid
Problem
auto-insurance fraud and slow claims handling
Technology
AI fraud detection (Shift Technology Force)
Who uses it
PT Asuransi Tokio Marine Indonesia, Indonesia
Outcome
none stated
Effort
high
Risks
False positives delaying honest claimants; model drift as fraud changes; vendor lock-in. Source is the vendor's release.
How to adopt
Needs historical claims data with known fraud outcomes and a claims system to feed scores into. Start with auto claims, routing flagged claims to investigators. Running it needs investigators and regular recalibration.
How a room runs it
the ROOM for a claims desk: the claims head, the adjusters, a special-investigations officer and a claims teammate (an agent definition with the insurer's claims rules); the agent reads each claim file the insurer already has and GENERATES the adjusters' tasks (request the workshop quote, call the claimant by Friday) with due dates, and routes a flagged claim as a decision to the investigations officer with a deadline; adjusters keep the claimant contact, the claims head signs every settlement above the envelope; the receipt records the file read, the flag and the signer Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the insurer's policy-admin and claims system loaded on this pull; a claims professional as the paid match; the pack's motor-claims templates
Source: vendor · 2020 · vendor only; to corroborate
Project Nemo: seven task-specific AI agents (coverage check, weather verification, fraud screening, payout calculation, audit summary) handle simple food-spoilage claims after storms; a human makes the final payout decision
Where
AU, Oceania
Segment
enterprise (global insurer, local pilot)
Problem
slow settlement of simple, high-volume catastrophe claims
Technology
agentic AI with seven task agents, weather data, fraud screening
Who uses it
Allianz (Allianz Services, Allianz Australia), Australia
Outcome
80% reduction in claim processing and settlement time (Allianz Services' Chief Transformation Officer; Allianz's own figure) the company’s own figure
Effort
high
Risks
Wrong coverage or fraud calls on edge cases; the figure is the insurer's own; claim-value caps limit scope.
How to adopt
Pick one simple, high-frequency claim type with a clear policy rule and data feeds; build task agents with a human payout step; run on a small value cap first.
How a room runs it
the ROOM for a catastrophe-claims desk: the claims handler, the claims head, the fraud analyst and a claims teammate; the agent runs the checks and GENERATES the handler's tasks (confirm the weather evidence, approve the calculation by end of day) as plan steps, and the payout as a decision in the claims head's Inbox with a deadline; the humans keep the payout decision, by design; the receipt records each agent step, the evidence and the signer. Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
an engagement for an insurer: the claims system, policy data and weather feed wired through connectors loaded on this pull, claim types chosen with the claims head, a claims professional as the paid match
Source: vendor · 2025 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 7 adoptions here, 0 were low effort, 3 medium and 4 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs bank statements and invoices and an SMB willing to move its books to the cloud. Switch on automatic categorisation and bank import. Running it needs someone to review categories monthly. Year is the read date (page undated).
Accounting software for SMBs with automatic transaction categorisation, journal entries, bank-statement import, recurring invoices and AI-based report analysis
Where
ID, Southeast Asia
Segment
SMB
Problem
manual bookkeeping and reconciliation consume the owner's week
Technology
rule-based automation, AI report analysis (Airene), POS and marketplace sync
Who uses it
Mekari Jurnal customers, Indonesia and Southeast Asia
Outcome
none stated
Effort
low
Risks
Mis-categorised entries; vendor lock-in; customer counts are unverified.
How to adopt
Needs bank statements and invoices and an SMB willing to move its books to the cloud. Switch on automatic categorisation and bank import. Running it needs someone to review categories monthly. Year is the read date (page undated).
How a room runs it
the ROOM for a small shop's books: the owner, the part-time bookkeeper and a bookkeeping teammate; the agent reads the ledger the owner already keeps and GENERATES the bookkeeper's weekly tasks (categorise 30 transactions, reconcile the bank by Friday) with due dates, and each invoice to send or bill to pay as a proposal in the owner's Inbox; nothing is paid or issued until the owner signs; the receipt records every posting Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connector to the owner's accounting package and bank statements loaded on this pull; WhatsApp receipt capture; an accountant as the paid match; the pack's SMB bookkeeping templates
Source: vendor · 2026 · vendor only; to corroborate
A digital bank builds its own transformer model on customer transaction data for credit risk (Nuformer) and puts over 5% of its workforce into AI-related roles
Where
BR, Latin America
Segment
enterprise (digital bank)
Problem
credit risk models built on hand-made features plateau
Technology
in-house transformer model on transaction sequences, large AI team
Who uses it
Nubank (Brazil, also Mexico and Colombia); CTO Eric Young as quoted
Outcome
Nuformer reportedly cut credit risk by 70% against earlier models; over 5% of the workforce in AI roles, about three times the regional bank average (as reported from the Evident ranking; measure undefined)
Effort
high
Risks
'70% less credit risk' is undefined on the page; model drift in downturns; fair-lending review; heavy in-house talent need.
How to adopt
Back-test a sequence model against the current scorecard on held-out cohorts, define the risk measure before comparing, and run both in parallel before switching decisions.
How a room runs it
the ROOM for a consumer-lending team: the credit lead, a data scientist, the compliance officer and a credit-model teammate (an agent definition with the lender's policy); the agent teammate reads model monitoring and portfolio reports the lender already holds and GENERATES the data scientist's checks (compare the new model on the March cohort by Thursday) as plan steps with owners and due dates, and each model change for the credit lead (ship / hold, with the back-test attached) as a decision block in the Inbox with a deadline; what the humans keep: the model release, the policy and the regulator reporting; the receipt records every finished task and every decision, and the client receives the reports and gives direction there Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the model registry, the portfolio performance export, the policy repository), a credit-modelling specialist joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Cloud accounting for the monthly close: a brand's accounting and finance manager reports monthly accounting finishing in days
Where
ID, Southeast Asia
Segment
SMB and mid
Problem
monthly close takes weeks of manual work
Technology
cloud accounting SaaS (Mekari Jurnal)
Who uses it
Brodo (accounting and finance manager testimonial), Indonesia
Outcome
monthly accounting closed in days (one customer's testimonial) the company’s own figure
Effort
low
Risks
Testimonial not tied to any AI feature; vendor lock-in; migration errors.
How to adopt
Needs transactions in a cloud accounting tool. Move one entity's ledger over and close one month in parallel. Running it needs an accountant.
How a room runs it
the ROOM for the month-end close: the finance manager, the owner and a close teammate; the agent reads the ledgers and bank lines already there and GENERATES the close checklist as tasks with owners and due dates (reconcile, accrue, chase missing invoices) and the close sign-off as a decision with a deadline; the finance manager keeps the sign-off and any payment; the receipt records each step and who closed Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connector to the brand's accounting software, POS and bank statements loaded on this pull; WhatsApp chasers to staff for missing documents; an accountant as the paid match; the pack's month-end templates
Source: vendor · 2026 · vendor only; to corroborate
AI financial assistant (JAX) coordinating agents to automate bank reconciliation, data entry and getting paid, with web research and data-issue spotting
Where
multi, global
Segment
SMB
Problem
small businesses lose hours to reconciliation and chasing payment
Technology
multi-agent platform, LLM assistant, OpenAI web research
Who uses it
Xero, global (small businesses and advisors)
Outcome
Xero's own figure: 73% of customers have used AI across the platform since March 2025 (release of 2025-09-03); no time-saved figure the company’s own figure
Effort
low
Risks
Mis-posted entries; data handled by a vendor; autonomy of agents over financial records.
How to adopt
Needs a cloud ledger connected to bank feeds. Switch on the assistant for reconciliation and data entry, with approval on postings. Running it needs a reviewer.
How a room runs it
the ROOM for an owner and their bookkeeper: the owner, the external accountant and a bookkeeping teammate; the agent GENERATES the weekly tasks (chase these three overdue invoices, review the unmatched bank lines by Thursday) and drafts payment reminders and bills to pay as proposals in the owner's Inbox with a deadline; the owner signs every payment and every invoice issued, the accountant keeps the filing; the receipt records each reconciliation and who approved Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
connectors to the owner's ledger and bank loaded on this pull; WhatsApp reminders to customers; an accountant as the paid match; the pack's receivables templates
Source: vendor · 2025 · vendor only; to corroborate
A virtual team of AI agents (accounting, customer, finance, project management, VAT in beta) inside a small-business accounting platform, with an AI-powered bank feed
Where
GB, Europe
Segment
SMB and accountants
Problem
bookkeeping and admin time for small businesses
Technology
AI agents, AI-powered bank feed, categorisation
Who uses it
Intuit (QuickBooks), United Kingdom
Outcome
Intuit says businesses can save up to 12 hours a month; the footnote's basis is an Intuit-commissioned survey of US customers using the AI bank feed in April 2025, not a study of the agents the company’s own figure
Effort
low
Risks
Miscategorised transactions reaching the books; the saving figure is a vendor survey; VAT agent still beta.
How to adopt
Needs an accounting platform subscription with connected bank feeds; switch on the agents, review categorisations for a month, then widen to invoicing and reporting.
How a room runs it
the ROOM for a small business and its accountant: the owner, the bookkeeper or accountant and a bookkeeping teammate; the agent reads the bank feed and GENERATES the bookkeeper's tasks (confirm the 14 uncategorised items, chase the missing invoice by Friday) as plan steps, and the month-end close and any VAT return as a decision in the owner's Inbox with a deadline; the humans keep the categorisation judgement and the filing; the receipt records each entry, the reviewer and the filing. Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
an engagement for a business or practice: the client's accounting platform and bank feeds wired through connectors loaded on this pull, a month-end routine set up, a bookkeeper or accountant as the paid match
Source: vendor · 2025 · vendor only; to corroborate
The same AI agents offered to Canadian small businesses, with a sales-tax agent announced as coming
Where
CA, North America
Segment
SMB and accountants
Problem
bookkeeping, lead follow-up and reporting load on small firms
Technology
AI agents for accounting, customers, finance and projects
Who uses it
Intuit (QuickBooks), Canada
Outcome
Intuit says up to 12 hours a month saved by businesses already using it (no method or sample given) the company’s own figure
Effort
low
Risks
Vendor-reported saving with no method; sales-tax agent not yet released; drafts may be wrong.
How to adopt
Needs an accounting platform subscription; enable one agent, compare its output with the bookkeeper's for a month, then widen.
How a room runs it
the ROOM for a small firm: the owner, the bookkeeper and a finance teammate; the agent drafts the profit-and-loss and cash-flow summaries and GENERATES the bookkeeper's tasks (reconcile the account, send the reminder for the overdue invoice by Monday) as plan steps, and the owner's choice on spending or pricing as a decision with a deadline; the humans keep the books' sign-off and the sales-tax filing; the receipt records each draft, the reviewer and the filing. Anything that moves money, issues an invoice or binds a contract is a PROPOSAL a named human signs, never automatic.
The engagement on top
an engagement for a small business: the accounting platform and inbox wired through connectors loaded on this pull, a reporting rhythm set up, a bookkeeper as the paid match
Source: vendor · 2025 · vendor only; to corroborate
Four ways to adopt it
Build it in-house
Your own team builds it. Across the 6 adoptions here, 5 were low effort, 0 medium and 1 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs a single member database across clubs and a payment link. Start with class booking and check-in in one app, then add waitlist and no-show rules. Running it needs front-desk training and an owner for member complaints. Year is the read date (page undated).
An unstaffed, 24-hour convenience-gym chain replaces the personal trainer with digital support: an app, a free scale and body-composition data, and a five-minute daily habit programme, at a flat monthly fee
Where
JP, East Asia
Segment
enterprise
Problem
gyms priced and staffed beyond the reach of people who want a light, habitual start
Technology
unstaffed 24-hour clubs, member app, connected scale and health data, DX-delivered coaching
Who uses it
chocoZAP (RIZAP Group), Japan; RIZAP director Takayuki Kamaya explains the launch
Outcome
572 locations and about 550,000 members by May 2023, one year after the July 2022 launch (RIZAP earnings briefing, reported by Toyo Keizai)
Effort
high
Risks
No staff on site (safety, cleanliness); fast growth strains support; the article does not show retention.
How to adopt
Needs low-cost sites, remote access control and a member app with billing. Open one site, add a health-data habit programme, then replicate. Running it needs a cleaning and equipment-fault rota and a member-support line.
How a room runs it
the room for an unstaffed-club operator: the area manager, a community-care lead and a member-support teammate; the agent reads the member, visit and equipment-fault records the chain already holds and GENERATES the area team's tasks as plan steps with owners and due dates (clean-and-check rota for a site, a call to members who have not visited in 30 days), and a DECISION in the area manager's Inbox (open another site / hold) with a deadline; the members are never contacted without the community-care lead's approval; the receipt records the visit data used and each outreach
The engagement on top
a custom engagement: the operator's own access-control, billing and app data wired in through connectors loaded on this pull, a fitness-operations professional on the match, the sector pack's site-rota and win-back templates
AI health companion inside a wearable app, rolled out to all members
Where
FI, Europe
Segment
consumer
Problem
members get numbers but not guidance
Technology
Oura Advisor (LLM plus Oura algorithms)
Who uses it
Oura (ring maker, Finland), members worldwide
Outcome
87% of members said it recalled their goals; 56% of users said chats led to tangible health benefits (Oura surveys, Mar 2025) the company’s own figure
Effort
medium
Risks
Not a medical device; survey results are the vendor's own; sensitive health data.
Next: discussing women's health and heart health with personal data is planned source
How to adopt
Test with a labs group, publish survey results, then roll out to all with editable memory.
How a room runs it
the personal room: the member, an optional coach and the advisor agent; the agent reads sleep and activity data and GENERATES a weekly goal with check-in tasks, each a proposal the member accepts; memory items are visible and deletable, and the receipt shows which data fed each tip
The engagement on top
connector to the wearable data, a coach on the match, a goal template
A multi-club chain moves bookings, check-in, membership and class waitlists into one free member app: 24/7 class booking six days ahead, automatic waitlist promotion with a notification, no-show suspension after four misses in 30 days
Where
ID, Southeast Asia
Segment
enterprise
Problem
empty class slots and no-shows; front-desk load; members who cannot see or change their own booking
Technology
member mobile app (iOS/Android), class booking engine, waitlist automation, QR/app check-in at reception
Who uses it
Fitness First Indonesia (Evolution Wellness), Indonesia
Outcome
none stated
Effort
medium
Risks
Strict no-show suspension may annoy members; app outages block access; member data privacy.
How to adopt
Needs a single member database across clubs and a payment link. Start with class booking and check-in in one app, then add waitlist and no-show rules. Running it needs front-desk training and an owner for member complaints. Year is the read date (page undated).
How a room runs it
the ROOM for a club: the club manager, the front-of-house lead, the head instructor and a membership teammate (an agent definition with the gym's skills); the agent reads the booking list and waitlist the club already has and GENERATES the week's tasks as plan steps with owners and due dates (instructor cover for a class with a long waitlist, a nudge to members after repeated no-shows), and a DECISION for the manager in the Inbox (open an extra class slot / hold) with a deadline; the humans keep the front desk, the class, the sign-off and any spend above the envelope; the receipt records each change made and who approved it
The engagement on top
a custom engagement: the club's own booking and access system wired to the room through connectors loaded on this pull, the member list reached over WhatsApp (the channel is built), a fitness-operations professional joins as the paid professional on the match, and the sector pack carries the class and membership task templates
Source: vendor · 2026 · vendor only; to corroborate
A listed franchisor deploys AI-enabled CRM software and churn-prediction models to improve member retention, and pilots AI personalised coaching inside its mobile app
Where
US, North America
Segment
enterprise
Problem
member cancellations; one-size-fits-all coaching at scale
Technology
AI-enabled CRM, predictive churn models, in-app AI coaching (pilot)
Who uses it
Planet Fitness (USA), CEO Colleen Keating on the Q4 2025 earnings call
Outcome
none stated
Effort
high
Risks
Health and body data raises consent and privacy duties; coaching advice quality; churn model wrongly targets members.
How to adopt
Needs member visit and billing history in a CRM. Start with a churn-prediction model feeding retention outreach, then test AI coaching in the app on a small group. Running it needs data science and a consent flow for coaching data.
How a room runs it
the same club room at chain scale: the regional manager, the club manager and a retention teammate; the agent reads the member and visit records the chain already holds and GENERATES the manager's follow-up list as tasks with owners and due dates (call these members this week), and a DECISION in the manager's Inbox (offer / hold) with a deadline; the member is never contacted by an agent without the manager's approval; budgets for offers degrade to a request before they block; the receipt records the offer and its outcome
The engagement on top
a custom engagement: the chain's own CRM and membership system wired in through connectors loaded on this pull, a retention professional on the match, the sector pack's follow-up templates
Small gyms, studios and sports venues move off WhatsApp and a notebook onto a shared booking, membership, package and payment platform, with a regional programme to digitalise 1,000 operators by 2027
Where
MY, Southeast Asia
Segment
SMB
Problem
operators run bookings and memberships by hand; the company says most operators in the industry run on WhatsApp and a notebook
Technology
booking and membership platform, payments, consumer marketplace app
Who uses it
Katch International (Malaysia) with operators in Selangor; backed by the Raja Muda Selangor; masterclasses by pitchIN and Anytime Fitness
Outcome
Katch serves more than 34,000 registered users; 50 business owners at the launch gathering; target of 1,000 businesses digitised by 2027 (company figures, reported by The Star)
Effort
low
Risks
Platform dependence; operators' low digital skills; no AI claimed and no results published.
How to adopt
Needs the operator's member and class lists, which usually sit in chat apps or a notebook. Move bookings, packages and payments onto one shared platform for a single venue first. Running it needs staff time to import members and train instructors.
How a room runs it
the ROOM for an owner-run studio: the owner, the front-desk person and one or two coaches with a studio teammate; the agent works from the WhatsApp chats and the notebook the owner already keeps and GENERATES the week's tasks (confirm Saturday's class cover, chase two lapsed packages) with owners and due dates, and a DECISION for the owner (extend a promotion / hold) with a deadline; the owner keeps every payment and every promise to a member; the receipt records what was agreed with whom
The engagement on top
a custom engagement: the studio's booking sheet or payment app wired in through connectors loaded on this pull, WhatsApp as the door (built), a studio-operations professional on the match, the pack's task templates
A low-cost gym chain launches five AI 'Fitbuddy' personas inside its member app to re-engage members four to six weeks after joining, when many lose momentum, built from behavioural and app-interaction data of 4.4 million members
Where
NL, Europe
Segment
enterprise
Problem
early drop-off of new members; one-size-fits-all motivation at a scale of millions
Technology
generative AI personas in the member app, behavioural data from app interactions
Who uses it
Basic-Fit, Netherlands (chief commercial officer Erica van Vonderen-Hahn quoted)
Outcome
none stated
Effort
medium
Risks
Unproven retention effect; members may find AI personas intrusive; claims about data handling are the company's own.
How to adopt
Needs app-interaction data per member and a view of when new members stop coming. Start with one persona and one cohort at the usual drop-off week, and compare with a holdout group. Running it needs content review and a way for members to switch it off.
How a room runs it
the club-chain room: the retention lead, a club manager and a member-engagement teammate; the agent reads the app-engagement data the chain already holds and GENERATES the retention team's tasks as plan steps with owners and due dates (review the week-4 drop-off cohort, rewrite the nudge for one persona), and the wording of each new nudge lands in the retention lead's Inbox as a proposal with a deadline; the humans keep the tone of voice and what members are told; the receipt records which nudge went to which cohort and who approved it
The engagement on top
a custom engagement: the chain's own member-app and engagement data wired in through connectors loaded on this pull, a retention professional on the match, the sector pack's nudge-review templates
AI insights on every uploaded workout for paying members, opt-out at any time
Where
US, North America
Segment
consumer
Problem
raw metrics are hard to interpret
Technology
Strava Athlete Intelligence
Who uses it
Strava, United States
Outcome
Built for a base of over 100 million users and more than 10 billion activity uploads; 14 languages (Canadian Running, Oct 2024)
Effort
low
Risks
Generic or wrong training advice; not medical advice; privacy of health data.
Next: more insights and training features were planned source
How to adopt
Start with a summary insight, keep it a paid and opt-out feature, add planning later.
How a room runs it
the club room: the athlete, the coach and a training agent; the agent summarises each session and GENERATES next-week workout tasks with dates, each a proposal the athlete accepts or declines; the coach sees the thread and the receipt shows what data the insight used
The engagement on top
connector to the workout platform, a coach on the match, a training-plan template
Your own team builds it. Across the 7 adoptions here, 2 were low effort, 3 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs a membership list and a payment account. Switch on membership management, renewals and accounting sync. Running it needs a volunteer admin.
A UK athletics club replaces spreadsheets and cash/cheque payments with online subscriptions, training fees, clothing orders and group messages to coaches and committees, with permissions shared across volunteers
Where
GB, Europe
Segment
SMB (volunteer-run)
Problem
payments by cash and cheque; communications depended on key volunteers
Technology
club-management software, online subscription payments, group communications
Who uses it
Harmeny Athletic Club, Edinburgh (UK)
Outcome
none stated
Effort
low
Risks
Vendor lock-in; volunteer turnover; handling of minors' data.
How to adopt
Needs a member list and an online payment service. Move subscriptions, fees and messages to one platform with permission levels for volunteers. Running it needs a volunteer who administers it. Year is the read date (page undated).
How a room runs it
the same club room: the secretary, the coaches and the treasurer share the work so no club depends on one volunteer's availability; the agent GENERATES the fees-due and kit-order tasks with owners and due dates and puts the committee's choices (fee change, kit order) in the Inbox as proposals with a deadline; the humans keep the coaching and the money; the receipt records each step so a new volunteer inherits the history
The engagement on top
a custom engagement: the club's subscription and payment tool wired in through connectors loaded on this pull, a club-administration professional on the match, the pack's season-start templates
Source: vendor · 2026 · vendor only; to corroborate
Indonesian neighbourhoods and community associations (RT/RW, paguyuban) manage dues, bills, payment proof, polls and reminders in a mobile app, with WhatsApp broadcast and WhatsApp login
Where
ID, Southeast Asia
Segment
SMB (community)
Problem
volunteer treasurers chasing dues and keeping books by hand
Technology
mobile app, bill and payment history, notifications and reminders, WhatsApp broadcast
Who uses it
RTPINTAR (PT Arsitek Perangkat Lunak Indonesia) for pengurus and warga, Indonesia
Outcome
none stated
Effort
low
Risks
Resident data privacy; low uptake by older residents; app vendor continuity.
How to adopt
Needs a member list per neighbourhood and a treasurer willing to run dues digitally. Start with dues and bill reminders in one neighbourhood. Running it needs an admin and resident onboarding. Year is the read date (page undated).
How a room runs it
the ROOM for a community or hobby group: the organiser, the treasurer and a few members, with a community teammate; the agent reads the dues list the group keeps and GENERATES reminder and follow-up tasks for the treasurer with due dates, and a DECISION for the organiser (waive / pursue an overdue member) with a deadline; the group keeps the money and the relationships; budgets degrade before they block; the receipt records dues collected and decisions made
The engagement on top
a custom engagement: the group's dues sheet or app wired in through connectors loaded on this pull, WhatsApp as the door (built), the pack's dues and event templates
Source: vendor · 2026 · vendor only; to corroborate
Conversational match assistant for tennis fans at a Grand Slam
Where
US, North America
Segment
consumer
Problem
fans want live answers during matches
Technology
IBM Granite models and watsonx Orchestrate
Who uses it
US Open (USTA), United States, with IBM
Outcome
Match chat, live win predictions and play-by-play insights shipped for the 2025 tournament; no usage figure published
Effort
medium
Risks
Wrong predictions; latency on match days; no published outcome figure.
Next: the same stack adds a tennis-coach style experience for novices and experts source
How to adopt
Start with post-match summaries, then add live chat grounded in the scoring feed.
How a room runs it
the event room: the tournament's digital team and a match agent; the agent answers fan questions from live data and GENERATES post-match summaries with publish deadlines, each a proposal an editor can check; the receipt records the data used
The engagement on top
connectors to the scoring feed, an editor on the match
Generative AI hole-by-hole insights and bilingual narration on a major golf tournament's site and app
Where
US, North America
Segment
consumer
Problem
fans want live context for every shot
Technology
IBM watsonx
Who uses it
Augusta National Golf Club (Masters Tournament), United States, with IBM Consulting
Outcome
Eight years of data including more than 170,000 shots; about 20,000 highlight videos narrated over the tournament (IBM, Apr 2024)
Effort
high
Risks
Wrong narration on a global stage; language quality; vendor dependence.
Next: Spanish narration extends the English AI narration of 2023, with watsonx used to monitor the model source
How to adopt
Train on owned data, add a second language with expert tuning, keep an editor in the loop.
How a room runs it
the event room: the tournament's digital editors, a data agent and a narration agent; the agent GENERATES insight cards and narration drafts per hole and per highlight with publish deadlines, each a proposal an editor can edit before release; the receipt records data and reviewer
The engagement on top
connectors to the shot data and the video library, an editorial review template
Community organisers run announcements and many sub-groups inside one WhatsApp Community (admin-only announcement channel, up to 50 groups, up to 5,000 members, polls)
Where
multi, global
Segment
SMB (community)
Problem
coordination across many chat groups for volunteer-run communities
Technology
WhatsApp Communities
Who uses it
WhatsApp (Meta) feature rollout to Indonesian users, reported by CNBC Indonesia
Outcome
none stated
Effort
low
Risks
Platform dependence; spam and moderation load; members' phone numbers are visible within groups.
How to adopt
Needs a group admin and a clear structure of sub-groups. Create a community with an announcement channel and attach sub-groups. Running it needs moderation.
How a room runs it
the room starts where the group already is: the organiser adds the community teammate to the WhatsApp group; the agent turns the chatter into the week's tasks with owners and due dates (book the field, collect dues, confirm the match) and puts the organiser's choices in the Inbox with a deadline; it posts to the announcement channel only after the organiser approves; the receipt records what was announced and agreed
The engagement on top
WhatsApp as the door (the channel is built); the group's own calendar and dues sheet wired in through connectors loaded on this pull; the pack's event templates
A volunteer-run sports club (1,000+ members, no employees) replaces paper applications and manual payments with membership management, auto-renewal, event duplication, polls and accounting sync
no change measured; the club reports 1,000+ members and over $1 million cash flow run by volunteers with no employees (club webmaster, quoted in the vendor's case study) the company’s own figure
Effort
low
Risks
Vendor lock-in; dues-stability claim is the vendor's; member data privacy.
How to adopt
Needs a membership list and a payment account. Switch on membership management, renewals and accounting sync. Running it needs a volunteer admin.
How a room runs it
the ROOM for a club run by volunteers: the chair, the treasurer, the event lead and a club teammate (an agent definition with the volunteer-club skills); the agent reads the member list and payments the club already has and GENERATES the volunteers' tasks with owners and due dates (renewal reminders, trip sign-up chase, the event checklist) and a DECISION for the committee (confirm the trip / cancel) with a deadline; members' money and the vote stay with people; the receipt is the committee's record of who decided what
The engagement on top
a custom engagement: the club's membership and accounting system wired in through connectors loaded on this pull, WhatsApp as the door, a club-administration professional on the match, the pack's event and renewal templates
Source: vendor · 2026 · vendor only; to corroborate
AI companion for fans inside the league's official app and website
Where
GB, Europe
Segment
consumer
Problem
fans cannot find answers across decades of stats, articles and video
Technology
Microsoft Copilot on Azure OpenAI
Who uses it
Premier League (United Kingdom), with Microsoft
Outcome
Archive of about 30 seasons of stats, 300,000 articles and 9,000 videos; league reaches 900 million homes in 189 countries (Microsoft release, Jul 2025) the company’s own figure
Effort
medium
Risks
Wrong or stale statistics; brand risk from off-topic chat; data use and consent.
Next: open-text questions in fans' native languages, then an AI assistant manager in Fantasy Premier League source
How to adopt
Ground the assistant in the owned archive, label it as AI, add languages and features by season.
How a room runs it
the fan room: the fan, a club's community team and a stats agent; the agent answers from the league's own archive and GENERATES a personal match-day digest and reminders, and a question it cannot ground becomes a request to the editorial team rather than an invented answer; the receipt shows the sources used
The engagement on top
connector to the club's stats and content archive, a moderation template, a community manager
A development bank digitises the books of village self-help groups: member savings, loans and repayments are recorded on a shared system that gives banks a management information view for credit decisions
Where
IN, South Asia
Segment
public (self-help groups)
Problem
paper book-keeping weakens groups' records and banks' credit decisions
Technology
digital SHG records, bank management information system
Who uses it
NABARD (National Bank for Agriculture and Rural Development), India; figures given in Parliament by the Minister of State for Finance
Outcome
130,176 groups digitised in 25 districts by 31 March 2017, about 69,696 of them credit-linked by 31 July 2017 (Minister of State for Finance, as reported by DSIJ)
Effort
high
Risks
Data entry errors at scale; groups' data used against them in credit decisions; pilot figures from 2017.
How to adopt
Needs group books in a consistent format and a trained digitiser or volunteer per cluster. Digitise one district's groups, then connect the records to the lending bank's view. Running it needs training and funding for the field workers.
How a room runs it
the room for a federation of self-help groups: the federation lead, a field officer, a bank liaison and a records teammate; the agent reads the group books the federation already holds and GENERATES the field officer's tasks with owners and due dates (fix three groups' missing entries, prepare the loan-application pack for the bank), and the choice to put a group forward for credit lands in the federation lead's Inbox as a proposal with a deadline; the members and the bank keep every credit decision; the receipt records what was submitted and who approved it
The engagement on top
a custom engagement: the federation's records system or sheets wired in through connectors loaded on this pull, WhatsApp as the door, a rural-finance professional on the match, the pack's loan-pack and audit templates
Rural savings groups record each meeting's savings, loans, repayments, fines and welfare contributions in an Android e-Recording app instead of paper books, with automatic meeting times, checks on wrong entries and a cloud backup
Where
KE, Africa
Segment
SMB (community)
Problem
paper records invite errors and tampering; share-out calculations are slow
Technology
Android recording app, internal data checks, cloud backup, basic group and member reports
Who uses it
Financial Sector Deepening (FSD) Kenya with Care Kenya and CRS Kenya, western Kenya
Outcome
none stated (groups report shorter meetings, easier share-outs and more accurate records; no figures)
Effort
low
Risks
Phone loss or failure; low digital literacy; qualitative evidence from one visit; adoption varied by region.
How to adopt
Needs one phone per group, a trained record-keeper and a facilitating organisation. Start with groups that already keep paper books and run both in parallel for a cycle. Running it needs facilitator support for phone and data problems.
How a room runs it
the room for a savings group: the chair, the treasurer, the secretary and a group-records teammate; the agent reads the meeting records the group already keeps and GENERATES the officers' tasks with owners and due dates (reconcile this month's contributions, prepare the share-out sheet, chase two overdue repayments), and a DECISION for the group (grant or defer a loan) lands in the chair's Inbox as a proposal with a deadline; the members keep every lending and share-out decision and the money; the receipt records each entry and who approved it
The engagement on top
a custom engagement: the group's records app or sheet wired in through connectors loaded on this pull, WhatsApp as the door (built), a community-finance professional on the match, the pack's meeting and share-out templates
Your own team builds it. Across the 9 adoptions here, 5 were low effort, 2 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs a knowledge base of tax rules and procedures and an escalation path to a human officer. Start with the top questions by volume. Running it needs content updates when rules change.
A tax authority pre-fills individual income tax returns from data it already holds (employers, banks, health), so a filer reviews and submits instead of typing; a majority of 2026 returns used it
Where
BR, Latin America
Segment
public
Problem
manual data entry and errors on individual returns; the 'fine mesh' of mismatches
Technology
pre-filled return (declaração pré-preenchida) built from third-party reporting
Who uses it
Receita Federal, Brazil
Outcome
16,093,257 returns received by 26 April 2026, 60.2% of them using the pre-filled option (Receita Federal, reported by ISTOÉ Dinheiro)
Effort
high
Risks
Wrong source data pre-filled and accepted unchecked; privacy of third-party data; fewer filers read what they sign.
How to adopt
Needs third-party reporting of income, expenses and assets in a common format and a secure filer login. Pre-fill the lowest-risk fields first and let the filer confirm each. Running it needs a clean-up route for wrong source data.
How a room runs it
the room for an accountant's individual clients: the accountant, the client and a filing teammate; the agent compares the pre-filled data with the documents the client sends and GENERATES the client's tasks with due dates (explain this deduction, send the receipt for the health expense), and the accountant's decision to accept or correct the pre-filled figures lands in their Inbox with a deadline; the accountant keeps the filing and the signature; the receipt records the differences found and who approved the return
The engagement on top
a custom engagement: the practice's client documents and the authority's pre-filled export wired in through connectors loaded on this pull, WhatsApp as the door, a tax professional on the match, the pack's reconciliation templates
A tax-software company uses LLMs to explain tax calculations to individual filers, with real-time accuracy checks over its tax knowledge engine and human experts assisted in TurboTax Live
Where
US, North America
Segment
enterprise
Problem
filers do not understand why a figure is what it is
Technology
LLMs (Anthropic Claude via Amazon Bedrock) over a rules-based tax knowledge engine, accuracy checks
Who uses it
Intuit TurboTax (USA)
Outcome
millions of individual filers received contextual explanations (Intuit's own blog; no exact count or accuracy figure) the company’s own figure
Effort
high
Risks
Wrong explanations of tax law; privacy of financial data; regulation of tax advice.
How to adopt
Needs a tax knowledge engine and filers' structured data. Start with LLM explanations of calculations backed by accuracy checks, with experts available. Running it needs continuous rule updates and evaluation.
How a room runs it
the same room at the person's scale: the client, the consultant and the filing teammate; the agent explains each figure in plain words from the rules the consultant uses and GENERATES the client's missing-document tasks; the consultant's review is a decision with a deadline; no number is filed without the human sign-off; the receipt records the explanation given and who approved the return
The engagement on top
a custom engagement: the consultancy's tax-engine or filing software wired in through connectors loaded on this pull, a tax consultant on the match
A conversational AI 'digital workforce' for tax teams that drafts step-by-step plans, pulls transactions via API from CRM/ERP/accounting, checks them against tax rules and prepares a filing-ready summary
Where
ID, Southeast Asia
Segment
SMB and mid
Problem
manual data validation before filing; Coretax transition
Technology
conversational AI, API integration to ERP/accounting
Who uses it
GITS.ID, promoted by Mekari Klikpajak (no named adopter), Indonesia
Outcome
none stated
Effort
medium
Risks
Wrong calculations leading to penalties; data security; vendor claims without an adopter or results.
How to adopt
Needs API access to accounting or ERP data and a defined tax rule set. Start with one tax type for one entity, with the tax team approving the summary. Running it needs rule updates and an audit trail.
How a room runs it
the consultant's room with the validation step made visible: the agent pulls what the client's books already hold, GENERATES the reconciliation tasks (explain this mismatch) for the client's bookkeeper with due dates, and the consultant's filing decision lands in their Inbox with a deadline; the consultant keeps the rule interpretation and the signature; the receipt records the data used and each approval
The engagement on top
a custom engagement: the client's accounting package wired through connectors loaded on this pull, a tax consultant on the match, the pack's reconciliation templates
Source: vendor · 2026 · vendor only; to corroborate
Tax consultants and individual taxpayers train on Coretax-based annual SPT filing through an association programme (training of trainers, online briefing, filing simulations)
Where
ID, Southeast Asia
Segment
SMB (consultants)
Problem
automated Coretax data mechanisms are hard for individual filers and practitioners
Technology
Coretax portal, simulation-based online training
Who uses it
IKPI Sidoarjo (Ikatan Konsultan Pajak Indonesia), Indonesia
Outcome
none stated
Effort
low
Risks
No AI involved; reach limited to attendees; system changes outdate the training.
How to adopt
Needs trainers who know the filing system and practice data. Run training of trainers, a briefing and filing simulations. Running it needs a repeat before each filing season.
How a room runs it
the consultant's room for the March rush: the associates, the consultant and each client; the agent walks a client through the filing steps the consultant taught and GENERATES the client's tasks with due dates; questions beyond the script become decisions for the consultant; the consultant keeps the filing and the advice; the receipt records the guidance given
The engagement on top
a custom engagement: the consultancy's client list wired in through connectors loaded on this pull, WhatsApp as the door, a tax consultant on the match, the pack's Coretax filing-season templates
Volunteer-assisted filing plus Coretax improves individual compliance: a study of 123 filers of 2025 returns finds both significant
Where
ID, Southeast Asia
Segment
public and SMB
Problem
late or missed individual filing
Technology
Coretax, tax-volunteer assistance
Who uses it
DJP tax volunteers and individual filers, Indonesia
Outcome
volunteer help and Coretax each show a significant association with individual compliance (survey of 123 filers of 2025 returns; correlational, no causal claim)
Effort
low
Risks
Small self-selected sample; correlation not causation; no AI.
How to adopt
Needs volunteers trained on the filing system and a place to help filers. Run assisted-filing sessions before the deadline. Running it needs volunteer coordination.
How a room runs it
the room is the volunteer made scalable: the consultant or volunteer coordinator, the volunteers and the clients; the agent GENERATES the client prompts and the volunteer's review tasks with due dates before the deadline; the coordinator decides who needs a human session; humans keep the filing; the receipt records who helped with which return
The engagement on top
a custom engagement: the coordinator's schedule and client list wired in through connectors loaded on this pull, WhatsApp as the door, the pack's SPT-season templates
A tax authority launches 'Kar Saathi', an AI chatbot on its e-filing portal that answers individual taxpayers' questions on ITR forms, refunds, notices, verification and deductions in plain language, alongside physical help centres
Where
IN, South Asia
Segment
public
Problem
a new income tax law and form changes confuse filers; long waits for help
Technology
AI chatbot widget on the e-filing portal
Who uses it
Income Tax Department / CBDT, India (PRARAMBH outreach programme)
Outcome
none stated
Effort
medium
Risks
Wrong answers on complex matters (capital gains, foreign assets, disputes); privacy of taxpayer data; users filing on its guidance alone.
How to adopt
Needs a knowledge base of tax rules and forms and a path to a human officer. Start with the most frequent questions and keep the help centres open. Running it needs content updates when rules change.
How a room runs it
the room for a chartered accountant's individual clients: the accountant, an associate, the client and a tax-preparation teammate; the agent reads the documents the client uploads and GENERATES the client's checklist as tasks with due dates (upload Form 16, confirm the refund bank account) and the accountant's review as a decision with a deadline; capital gains, foreign assets and notices stay with the accountant; the receipt records the documents used and who approved the return
The engagement on top
a custom engagement: the practice's client list and document store wired in through connectors loaded on this pull, WhatsApp as the door (built), a chartered accountant as the paid professional on the match, the pack's ITR-season templates
The UK government's GOV.UK Chat, a generative-AI assistant in the GOV.UK app, answers questions from official guidance including Income Tax, Stamp Duty and the State Pension, with tax among the strongest areas of demand
Where
GB, Europe
Segment
public
Problem
people cannot find tax answers on a very large website; helplines are overloaded
Technology
generative AI chat over GOV.UK guidance, One Login sign-in
Who uses it
Government Digital Service (GDS), United Kingdom
Outcome
more than 7,800 people and more than 15,000 questions since the March 2026 soft launch, tax among the strongest areas of demand (GDS figures, as reported by Ross Martin)
Effort
high
Risks
Confident but wrong tax answers (testers found a capital gains error in a related review); users treating it as advice; scope limited to guidance.
How to adopt
Needs an authoritative, current guidance set and a sign-in. Start with a soft launch to a limited audience and review answers on tax topics before widening. Running it needs accuracy testing and a visible path to a human.
How a room runs it
the room for an accountant's individual clients: the accountant, the client and a self-assessment teammate; the agent answers the client's routine questions from the guidance the accountant approves and GENERATES the client's missing-record tasks with due dates before the 31 January deadline; any question beyond basics becomes a decision in the accountant's Inbox with a deadline; the accountant keeps the advice and the filing; the receipt records the answer given and its source page
The engagement on top
a custom engagement: the practice's client records and the approved guidance set wired in through connectors loaded on this pull, WhatsApp as the door, a tax adviser on the match, the pack's self-assessment templates
A tax authority runs an AI virtual assistant for individual taxpayers: automated answers around the clock on NPWP, EFIN and annual SPT reporting, with hand-off to a live officer through Kring Pajak 1500200
Where
ID, Southeast Asia
Segment
public
Problem
queues and questions at tax offices; confusion at annual filing
Technology
NLP chatbot (Fiska/Fisko), live-chat hand-off, call-centre link
Who uses it
Direktorat Jenderal Pajak (DJP), Indonesia
Outcome
none stated
Effort
medium
Risks
Wrong tax advice from the assistant; privacy of taxpayer identifiers; users trusting it over an officer.
How to adopt
Needs a knowledge base of tax rules and procedures and an escalation path to a human officer. Start with the top questions by volume. Running it needs content updates when rules change.
How a room runs it
the ROOM for a tax consultant's individual clients: the consultant, an associate, the client and a tax-preparation teammate (an agent definition with the filing-season skills); the agent reads the documents the client uploads and GENERATES the checklist as tasks for the client with due dates (send the employer slip, confirm the NIK-NPWP match) and the consultant's review as a decision with a deadline before the 31 March filing date; the consultant signs and files; the receipt records every document received and the filing reference
The engagement on top
a custom engagement: the consultancy's own client list and document store wired in through connectors loaded on this pull, WhatsApp as the door (built), a licensed tax consultant as the paid professional on the match, the pack's SPT-season templates
Your own team builds it. Across the 8 adoptions here, 2 were low effort, 3 medium and 3 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs a warehouse with stable layout, a private wireless network and a warehouse system to take orders. Start with automated guided vehicles on one lane and camera analytics on one zone. Running it needs maintenance staff and network support.
Case-unloading robots in trailers and containers, moving to case picking under a 2025 strategic agreement
Where
DE, Europe
Segment
enterprise (contract logistics)
Problem
the most physically heavy warehouse work and labour gaps
Technology
Stretch (Boston Dynamics)
Who uses it
DHL Group and DHL Supply Chain, Germany; deployments in North America, UK and Europe
Outcome
Unloading rates up to 700 cases an hour; DHL Group uses more than 7,500 robots and over 90% of its warehouses have an automation solution; more than EUR 1 billion invested in contract-logistics automation in three years (DHL, May 2025)
Effort
high
Risks
Capital cost; worker safety and acceptance; the memorandum may not convert in full.
Next: more than 1,000 additional Stretch units under the agreement, with case picking a priority source
How to adopt
Pilot at one site with one robot, prove throughput and safety, then scale by site.
How a room runs it
the dock room: the site manager, the shift lead and the robot vendor; the agent plans trailer arrivals against robot and crew availability and generates the shift tasks with dates, and any change to the safety zone becomes a proposal for the safety officer; the receipt records the plan, the throughput and the approval
The engagement on top
connectors to the WMS and the dock scheduler, a shift-plan template, a safety-sign-off rule
Over one million warehouse robots coordinated by a generative AI foundation model for fleet movement
Where
US, North America
Segment
enterprise (e-commerce fulfilment)
Problem
robot congestion and slow order processing
Technology
DeepFleet (Amazon Robotics, AWS)
Who uses it
Amazon, United States; 300+ facilities
Outcome
One millionth robot deployed (July 2025); robots assist in roughly 75% of orders; DeepFleet is estimated to improve robot travel efficiency by about 10% (Amazon, Jul 2025)
Effort
high
Risks
Capital and energy cost; workforce transition; estimated gains not independently measured.
Next: next-generation robotics systems across existing facilities, and a Shreveport centre designed to cut processing times by up to 25% source
How to adopt
Start with simulated routing on historical data, then a single-site live trial.
How a room runs it
the fulfilment room: the site manager, the planner and the reliability team; the agent proposes robot routes and shift plans from congestion data and generates the maintenance tasks with dates, and a layout change becomes a proposal the planner approves; the receipt records the plan, the travel data and the approver
The engagement on top
connectors to the robot fleet manager and the WMS, a congestion report template, a layout-change approval rule
Intelligent logistics park: an automatic sorting centre with more than 80 sorting lines and about 10,000 sorting robots, sorting 4.5 million parcels a day
Where
CN, East Asia
Segment
enterprise
Problem
sorting volume of a very large e-commerce network and the cost and error rate of manual sorting
JD Logistics (JD.com group), Kunshan Asia No.1 Intelligent Logistics Park, Jiangsu province, China
Outcome
sorting capacity 4.5 million parcels a day, five-fold efficiency against traditional methods and 99.99 % sorting accuracy (company figures reported by Post&Parcel, 2023)
Effort
high
Risks
Capital cost; single-site dependency; the figures are the operator's own; labour displacement.
How to adopt
Needs a parcel network large enough to keep sorting lines busy, a site with fixed layout and a control system that talks to tracking. Start with one sorting line and compare throughput and accuracy with the manual line. Running it needs robot maintenance and exception handling.
How a room runs it
the ROOM for a sorting park: the park manager, the shift leads and a sorting-operations teammate (an agent definition with lines, chutes and exception vocabulary); the agent reads the exception and misroute queue the robots already produce and GENERATES the leads' tasks (clear these 120 mis-sorted parcels by 14:00, inspect the stalled line 17) as plan steps with owners and due times, and a DECISION for the park manager (open an extra line, reroute a surge to a second hub) with a deadline, landing in the manager's Inbox as a proposal; the humans keep the safety, labour and carrier decisions; the receipt records each exception cleared and who approved each reroute
The engagement on top
a custom engagement: the sorting control system and parcel tracking wired to the room through connectors loaded on this pull, a parcel-operations professional on the match, the courier pack's templates
Case-level warehouse automation: robots and machine vision break down inbound pallets, store goods at case level and build store replenishment orders, connected to the retailer's existing ordering system; first site live 2017, rollout to 25 distribution centres announced
Where
US, North America
Segment
enterprise
Problem
slow, error-prone manual pallet handling and order building for store replenishment
Technology
robotics, machine vision, storage and order-building software (Symbotic)
Who uses it
Walmart with Symbotic, United States (first system Brooksville, Florida, 2017; Symbotic acquired Walmart's Advanced Systems and Robotics unit, announced Jan 2025)
Outcome
none stated for the distribution-centre rollout; Jan 2025 deal: Walmart plans about 400 Symbotic systems for in-store accelerated pickup and delivery centres, if performance targets are met
Effort
high
Risks
Very high capital cost and long retrofit; supplier dependence; staff displacement; results are not published in this source.
How to adopt
Needs a distribution centre with stable pallet flow and a warehouse system that can pass orders to the automation. Start with one site and compare cases per hour and order accuracy with the manual baseline. Running it needs maintenance staff and a plan for the staff whose tasks change.
How a room runs it
the ROOM for a distribution centre: the site manager, the inbound and outbound supervisors and a warehouse-operations teammate (an agent definition with receiving, putaway and replenishment vocabulary); the agent reads the exceptions the automation already raises (a damaged pallet, a short case count, a store order that cannot be built) and GENERATES the supervisors' tasks (recount lane 4, re-slot these SKUs, call the carrier about the late trailer) as plan steps with owners and due times, and a DECISION for the site manager (re-route a store order, pause an inbound door) with a deadline, landing in the manager's Inbox as a proposal; the humans keep the safety stops, the labour plan and the carrier commitments; the receipt records each exception, the proposal and who approved it
The engagement on top
a custom engagement: the warehouse management and ordering systems wired to the room through connectors loaded on this pull, a warehouse-automation professional as the paid professional on the match, the logistics pack's task templates
AI agents that phone carriers about missed less-than-truckload pickups and decide the next step
Where
US, North America
Segment
enterprise (freight broker)
Problem
missed pickups force return trips and delay freight
Technology
custom-built AI agents (C.H. Robinson)
Who uses it
C.H. Robinson, United States, 11,000+ customers
Outcome
95% of missed-pickup checks automated, more than 350 hours of manual work saved per day, unnecessary return trips down 42%, freight moves about a day sooner (C.H. Robinson, Jan 2026)
Next: carriers can use the agents' data to improve scheduling and network performance source
How to adopt
Pick one exception type, let agents call, keep the decision with a human above a cost limit.
How a room runs it
the dispatch room: the account manager, the carrier contact and the shipper; one agent checks each missed pickup and a second decides the next step, and any reroute that costs money becomes a proposal with a deadline in the manager's Inbox; the receipt records the calls, the decision and the approver
The engagement on top
connectors to the TMS and the carrier call system, an exception playbook, a cost-approval rule
Fully automated container terminal with remotely operated automated guided vehicles and AI support for vessel positioning and traffic
Where
SG, Southeast Asia
Segment
public and enterprise (port)
Problem
port congestion and capacity limits
Technology
automated terminal and vessel traffic AI (PSA Singapore, MPA)
Who uses it
PSA Singapore and the Maritime and Port Authority (MPA), Singapore
Outcome
Singapore handled a record 41.12 million TEUs in 2024 (+5%); Tuas is planned to lift capacity to 65 million TEUs; automated vehicles run 6 to 8 hours after a 20-minute charge (Datamar News, citing MPA, Feb 2025)
Effort
high
Risks
Cyber exposure; single-point failures; volume figure is national, not Tuas-only.
Next: Tuas completes in the 2040s and MPA plans a next-generation AI-and-satellite vessel traffic system source
How to adopt
Automate one terminal phase at a time, keep manual fallback, add AI planning on top of stable operations.
How a room runs it
the port-operations room: the terminal planner, the vessel agent and the customs liaison; the agent proposes berth and yard plans from vessel data and generates the tasks per shift, and any change to a safety-critical sequence becomes a proposal for the duty officer; the receipt records the plan and the approver
The engagement on top
connectors to the terminal operating system and the vessel schedule feed, a berth-plan template, a duty-officer approval rule
5G smart warehouse: automated guided vehicles and AI-powered CCTV on a private 5G network to shorten processing time and cut handling errors
Where
ID, Southeast Asia
Segment
enterprise
Problem
manual warehouse handling is slow and error-prone
Technology
5G, AGVs, AI-powered CCTV, warehouse automation
Who uses it
Telkomsel Enterprise with Huawei (the operator of the warehouse is not named), Indonesia
Outcome
none stated
Effort
high
Risks
Capital cost and cost overrun; owner and results not named, so benefit is unproven; vendor lock-in to a network supplier.
How to adopt
Needs a warehouse with stable layout, a private wireless network and a warehouse system to take orders. Start with automated guided vehicles on one lane and camera analytics on one zone. Running it needs maintenance staff and network support.
How a room runs it
the ROOM for a warehouse shift: the warehouse manager, the shift leads and a warehouse-operations teammate (an agent definition with the inbound, putaway and picking vocabulary); the agent reads the exceptions the WMS and the camera system already raise and GENERATES the shift's tasks for the leads (recount bay 12, clear the blocked aisle, re-pick order 884) as plan steps with owners and due times, and a DECISION for the manager (overtime, re-route inbound, hold the dock) with a deadline in the Inbox; the humans keep the floor work, the safety sign-off and the spend above the envelope; the receipt records each exception, who closed it and when
The engagement on top
a custom engagement: the operator's own WMS and exception feeds wired to the room through connectors loaded on this pull, a warehouse-operations lead as the paid professional on the match, and the logistics pack's task templates; shift leads reached over WhatsApp
Source: vendor · 2024 · vendor only; to corroborate
Automated spare-parts warehouse: mini-load AS/RS with two stacker cranes, goods-to-person picking stations with put-to-light, planned pallet AS/RS
Where
ID, Southeast Asia
Segment
mid
Problem
three-day shipping lead times and picking errors at a parts distributor
PT Mitra Pinasthika Mulia (MPM), Honda motorcycle parts distributor for East Java, Sidoarjo, Indonesia
Outcome
shipping time cut to a third of the previous three days; shipping accuracy 99.9 % (trade-article figures, no named measurer)
Effort
high
Risks
High capital cost; breakdown stops the whole store; the article's lead-time claim is inconsistent. Automation, not AI.
How to adopt
Needs a purpose-built or rebuilt warehouse, barcode-tracked stock and a warehouse system. Start with a mini-load automated store and goods-to-person picking stations. Running it needs maintenance technicians and spare parts for the cranes.
How a room runs it
the ROOM for the parts desk: the logistics head, the dispatch lead and a distribution teammate; the agent reads the order queue and stock position the system already holds and GENERATES the day's tasks (release same-day wave, replenish fast movers, chase the short SKU with the supplier) as plan steps with owners and due times, and a DECISION for the head when the backlog will miss same-day shipping (add a shift, split the wave) with a deadline; the humans keep the floor work and the spend above the envelope; the receipt records each wave, its cut-off and its accuracy count
The engagement on top
a custom engagement: the distributor's WMS and ERP order feed wired in through connectors loaded on this pull, a warehouse-operations professional on the match, the distribution pack's templates
matching algorithm, mobile ordering and tracking, carrier verification
Who uses it
Logisly, Jakarta, for shippers including Unilever, Haier, Grab, Maersk and JD.ID; 40,000 trucks in network
Outcome
network of 40,000 trucks and more than 1,000 business customers, 300 of them corporate shippers (company figures as reported by TechCrunch in 2020; no shipment, revenue or savings figure)
Effort
medium
Risks
Two-sided market may not reach liquidity; carriers bypass the platform; source is dated (2020) and the platform may have changed.
How to adopt
Needs verified carriers with trackers and shippers willing to post loads on a platform. Start with a few lanes where demand is regular. Running it needs carrier onboarding and verification staff.
How a room runs it
the ROOM for a shipper's transport desk or a trucking firm's dispatch: the logistics manager, the dispatcher and a dispatch teammate; the agent reads the loads and trucks the company already has and GENERATES the day's tasks (book a return load for truck B 4421, confirm pickup windows, chase the missing proof of delivery) as plan steps with owners and due times, and a DECISION for the manager when a load needs a spot-rate truck above the envelope, with a deadline; the humans keep the carrier choice, the rate above the envelope and the driver calls; the receipt records each load from booking to proof of delivery
The engagement on top
a custom engagement: the company's TMS or spreadsheet of loads wired to the room through connectors loaded on this pull, a logistics operations professional on the match, the trucking pack's templates; drivers reached over WhatsApp
Freight forwarder's AI booking assistant that reads free-form booking requests and turns them into structured booking records
Where
DK, Europe
Segment
mid
Problem
staff retype emailed booking requests; slow and error-prone
Technology
LLM-based booking assistant (extraction from free text)
Who uses it
LEMAN, freight forwarder (international)
Outcome
more than 10,000 bookings processed by the assistant (company's own figure) the company’s own figure
Effort
medium
Risks
Misread bookings cause shipping errors; customer data in a third-party model; country is inferred (Denmark is likeliest, page does not state it).
How to adopt
Needs booking requests arriving by email in free form and a transport management system to receive structured bookings. Start with one customer group's emails, with staff checking records before confirmation. Running it needs a feedback loop for misread requests.
How a room runs it
the ROOM for a forwarding desk: the operations manager, the booking clerks and a forwarding teammate (an agent definition with booking, documents and customs vocabulary); the agent reads the inbox the desk already has and GENERATES the clerk's tasks (confirm the vessel cut-off, chase the missing packing list) as plan steps with owners and due dates, and a DECISION for the manager when a booking breaks a rate or a cut-off, with a deadline; the humans keep the carrier commitment and the customs sign-off; the receipt records the request, the draft booking and who approved it
The engagement on top
a custom engagement: the forwarder's TMS and mailbox wired to the room through connectors loaded on this pull, a customs broker as the paid professional on the match, the forwarding pack's task templates; customers reached over WhatsApp
Source: vendor · 2026 · vendor only; to corroborate
More than 30 AI agents automating freight procurement, execution, tracking, payment and analytics inside the transport management system
Where
US, North America
Segment
enterprise (freight marketplace)
Problem
manual freight bidding, tracking calls and invoice checks
Technology
Insights AI and TMS agents (Uber Freight)
Who uses it
Uber Freight, United States; early pilots with five enterprise customers including Colgate-Palmolive
Outcome
More than 30 agents covering procurement to analytics; early pilots with five enterprise customers (Uber Freight, May 2025); no savings figure published in the article
Effort
medium
Risks
Pricing errors; carrier trust; no published savings yet.
Next: the company plans to expand AI capabilities through 2025 and turn its TMS into a real-time command center source
How to adopt
Pilot on one lane with one customer, compare with human buyers, then widen.
How a room runs it
the freight room: the shipper's logistics lead, the carrier and the agent; the agent drafts the tender, tracks the load and flags delays, and an award or a payment above the limit becomes a proposal with a deadline in the lead's Inbox; the receipt records the bids, the choice and the approver
The engagement on top
connectors to the TMS and the carrier network, a tender template, a payment-approval rule
Pelindo (Pelindo I and IV, since merged), ports of Balikpapan and Belawan, Indonesia
Outcome
none stated
Effort
high
Risks
'AI' is the supplier's wording; wrong allocations affect port safety; staff displacement; supplier lock-in.
How to adopt
Needs digital records of vessel calls and a marine operating system that plans tugs, pilots and berths. Start with allocation planning, then digitise pilot job records into billing. Running it needs port staff trained and a supplier contract.
How a room runs it
the ROOM for a port's marine desk: the harbour master, the duty planners and a port-planning teammate; each arriving vessel is a deliverable the agent opens, and it GENERATES the planner's tasks (confirm tug and pilot for ETA 14:00, check berth clash) as plan steps with owners and due times, and a DECISION in the harbour master's Inbox when two vessels contend for one berth, with a deadline; the humans keep the berthing call and every safety sign-off; the receipt records what was proposed, what was decided and by whom, and flows to billing as the record
The engagement on top
a custom engagement: the port's marine operating system and vessel feeds wired to the room through connectors loaded on this pull, a port-operations professional on the match, the port pack's task templates
Your own team builds it. Across the 12 adoptions here, 0 were low effort, 4 medium and 8 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs a hub with stable parcel flow, a sorting system and a plan for redeploying sorters. Start with robots in one hub and measure sorting accuracy against manual. Running it needs maintenance staff and retraining.
Government-contracted autonomous drones deliver vaccines, blood and medicines on demand from distribution centres to health facilities
Where
GH, Africa
Segment
enterprise (national health supply chain)
Problem
stockouts of essential medicines at remote health facilities
Technology
autonomous delivery drones, ordering by phone or message, distribution centres
Who uses it
Zipline with the Government of Ghana; evaluation by IDinsight
Outcome
Zipline facilities had 5 fewer stockout days than control facilities' 45 over 3 months across 18 products (11%, p<0.01); vaccine stockout duration down 60% (3.2 days from 5.3, p=0.01); measured by IDinsight
Effort
high
Risks
Effects on patient referrals were not significant; weather and airspace limits; cost per delivery not in the summary; dependence on one operator.
How to adopt
Start with a distribution centre and a defined list of high-value products, evaluate against comparable facilities, and integrate ordering into facility routines.
How a room runs it
the ROOM for a health-supply operation: the supply chain manager, a regional pharmacist, the dispatch lead and a stock-monitoring teammate (an agent definition with the essential-medicines list); the agent teammate reads the facility stock reports the programme already holds and GENERATES the pharmacist's replenishment checks (confirm anti-snake serum stock at these 12 facilities by Tuesday) as plan steps with owners and due dates, and each stockout pattern for the supply chain manager (rebalance / reorder, with the evidence attached) as a decision block in the Inbox with a deadline; what the humans keep: the dispatch, the clinical use and the procurement; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the facility stock reports, the order and dispatch log, the procurement system), a public-health supply-chain specialist joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Route optimisation for delivery drivers: ORION builds the day's route for each driver, evaluating a very large number of options against service commitments
Where
US, North America
Segment
enterprise
Problem
delivery routes planned by hand cost miles, fuel and driver time
Technology
route optimisation algorithm (ORION), telematics and map data
Who uses it
UPS, United States (deployment completed 2016)
Outcome
100 million miles driven saved a year, about 10 million gallons of fuel and 100,000 metric tons of emissions, and USD 300 to 400 million a year (UPS's own submission, 2016) the company’s own figure
Effort
high
Risks
Long build time (over a decade at UPS); driver trust; self-reported savings; a plan that ignores local knowledge.
How to adopt
Needs stop-level data, a fleet with telematics and driver acceptance of a computed route. Start with one depot and compare miles per route with the drivers' own plans. Running it needs a team to tune business rules and handle driver feedback.
How a room runs it
the ROOM for a delivery hub: the hub manager, the dispatchers and a dispatch teammate (an agent definition with routes, stops and service-commitment vocabulary); the agent reads the day's planned routes and the exceptions that come back (a blocked street, a missed pickup window) and GENERATES the dispatchers' tasks (re-sequence route 12, call the customer for the 15:00 window) as plan steps with owners and due times, and a DECISION for the hub manager (add a van, defer low-priority stops) with a deadline, landing in the manager's Inbox as a proposal; the humans keep the driver relations, the service promise and the overtime call; the receipt records the plan, each change and who approved it
The engagement on top
a custom engagement: the courier's dispatch and tracking systems wired to the room through connectors loaded on this pull, a delivery-operations professional on the match, the courier pack's templates; we do not offer our own route optimiser, the agent works from the company's own routes
1,000 AI sorting robots in four city hubs; accuracy above 99 %; idle hours leased to other couriers; sorting staff at one site cut from 200 to 40 with reassignment
Where
ID, Southeast Asia
Segment
public (state-owned enterprise)
Problem
manual parcel sorting cost and errors
Technology
AI sorting robots, address-recognition with a safety line for unclear addresses
Who uses it
Pos Indonesia (state postal company), hubs in Surabaya, East Jakarta, Bandung, Yogyakarta
Outcome
sorting accuracy above 99 %; sorting staff at one site 200 to 40 with reassignment; labour saving up to 36 % (press report, unattributed; another outlet reports 30 %)
Effort
high
Risks
Staff displacement and labour relations; capital cost; the labour saving figure differs between outlets.
How to adopt
Needs a hub with stable parcel flow, a sorting system and a plan for redeploying sorters. Start with robots in one hub and measure sorting accuracy against manual. Running it needs maintenance staff and retraining.
How a room runs it
the ROOM for a sorting hub: the hub head, the shift supervisors and a hub-operations teammate; the agent reads the exception queue the robots already produce (unclear addresses on the safety line) and GENERATES the supervisors' tasks (resolve these 40 parcels by 16:00, call the sender, relabel) as plan steps with owners and due times, and a DECISION for the hub head (lease idle hours to a partner courier, move staff to another division) with a deadline; the humans keep the manual handling, the staffing decisions and the sign-off; the receipt records each resolved exception
The engagement on top
a custom engagement: the hub's sorting and tracking systems wired to the room through connectors loaded on this pull, a postal-operations professional on the match, the courier pack's templates
QR labels, apps for sellers, couriers and truckers, e-wallet, push tracking
Who uses it
Exaq (Indonesia) and J&T Express (real-time tracking and integration, no AI named)
Outcome
none stated
Effort
medium
Risks
Partner quality varies; cash-on-delivery fraud; dated (2018) and no AI.
How to adopt
Needs partner-run sorting stations, scan-capable apps for couriers and a payment wallet for cash on delivery. Start with QR-coded labels in one city. Running it needs partner management and cash reconciliation.
How a room runs it
the ROOM for a courier partner or an online seller's dispatch: the owner, the couriers and a dispatch teammate; the agent reads the day's pickups and cash-on-delivery balances the owner already has and GENERATES the couriers' tasks (collect from sellers A and B, top up the COD wallet by noon) as plan steps with owners and due times, and a DECISION for the owner on a late top-up or an address-change request, with a deadline; the humans keep the pickups, the cash and the sign-off; the receipt records each handover and each settlement
The engagement on top
a custom engagement: the courier's order and wallet systems wired in through connectors loaded on this pull, a courier-operations professional on the match; couriers and sellers reached over WhatsApp
National parcel-sorting investment: a new parcel processing machine at the Strathfield, Sydney facility handling 10,500 parcels an hour, with new sorters in Brisbane and Melbourne taking the network above 40,000 parcels an hour
Where
AU, Oceania
Segment
public (state-owned enterprise)
Problem
parcel growth from online shopping and manual handling injury risk
Technology
automatic parcel sorting machines
Who uses it
Australia Post, Strathfield (Sydney), Brisbane and Melbourne, Australia
Outcome
Strathfield machine 10,500 parcels an hour; network above 40,000 parcels an hour with the Brisbane and Melbourne sorters (Australia Post's figures, reported by Post&Parcel)
Effort
high
Risks
Capital cost; dated source; capacity is a rating, not measured throughput; staff displacement.
How to adopt
Needs a facility with growing parcel volume, floor space and a labelling standard the sorter can read. Start with one sorter and compare parcels per hour and injuries with the manual line. Running it needs maintenance staff.
How a room runs it
the ROOM for a parcel facility: the facility manager, the shift supervisors and a hub-operations teammate; the agent reads the reject and unreadable-label queue the sorter already produces and GENERATES the supervisors' tasks (re-label these 60 parcels, clear the jam alert on chute 9) as plan steps with owners and due times, and a DECISION for the facility manager (move volume to another site, add a shift) with a deadline, landing in the manager's Inbox as a proposal; the humans keep the safety, the roster and the service promise; the receipt records each exception and who approved each move
The engagement on top
a custom engagement: the facility's sorter control and tracking systems wired to the room through connectors loaded on this pull, a postal-operations professional on the match, the courier pack's templates
Robot depalletiser with AI vision that unloads randomly mixed pallets of parcels onto the sorting conveyor, first at the Rotterdam sorting centre, planned for other centres
Where
NL, Europe
Segment
enterprise
Problem
heavy, repetitive manual unloading of mixed parcel pallets
Technology
AI-based vision and gripper robot (integrator AWL)
Who uses it
DHL Parcel (Benelux) with AWL, Rotterdam, the Netherlands
Outcome
800 parcels an hour, up to 31.5 kg per parcel (integrator's figures); no savings figure the company’s own figure
Effort
medium
Risks
Vendor-reported; throughput of 800 an hour may not cover peak volume; fallback needed when vision misreads.
How to adopt
Needs a centre that unloads mixed pallets by hand today and room for a robot cell. Start with one cell and compare parcels per hour and injuries with manual unloading. Running it needs maintenance and a fallback to hand unloading.
How a room runs it
the ROOM for a sorting centre: the centre manager, the inbound supervisors and a sorting-operations teammate; the agent reads the robot's fault and mis-pick log and GENERATES the supervisors' tasks (check the gripper on cell 2, hand-unload these 3 pallets) as plan steps with owners and due times, and a DECISION for the centre manager (roll the robot to the next shift pattern, move people to the heavier lanes) with a deadline, landing in the manager's Inbox as a proposal; the humans keep the safety sign-off and the staffing choices; the receipt records each fault and who approved each change
The engagement on top
a custom engagement: the centre's sorting and tracking systems wired to the room through connectors loaded on this pull, a parcel-operations professional on the match, the courier pack's templates
Source: vendor · 2026 · vendor only; to corroborate
Announced next step: AI route optimisation across 642,000 active delivery routes, and geofencing that blocks a delivered status when the courier is over 100 m from the address
Where
ID, Southeast Asia
Segment
public (state-owned enterprise)
Problem
routes planned by hand; couriers marking parcels delivered from elsewhere
Technology
AI route optimisation, geofencing on the courier app
Who uses it
Pos Indonesia, nationwide (Aceh to Lombok)
Outcome
none stated
Effort
medium
Risks
Planned only; geofencing may block legitimate deliveries in poor GPS areas; courier monitoring concerns.
How to adopt
Needs clean delivery route and address data and courier devices that report location. Start with route optimisation on one hub's routes, then add geofencing on delivery status. Running it needs data upkeep and courier training.
How a room runs it
the ROOM for a delivery branch: the branch head, the route supervisors, the couriers over WhatsApp and a delivery teammate; the agent reads the day's parcels and routes the branch already has and GENERATES each courier's task list as plan steps with due times, and a DECISION for the branch head when a route cannot finish by cut-off (borrow a courier, defer a bundle), with a deadline; couriers keep the delivery and a failed-delivery reason; the receipt records each drop with its proof
The engagement on top
a custom engagement: the postal or courier firm's route and tracking systems wired in through connectors loaded on this pull, a delivery-operations professional on the match, the courier pack's templates
Ninja Van Group (Singapore-based courier), automation work in Singapore, the Philippines, Indonesia and Thailand
Outcome
expected 50 % increase in operational productivity across the region (company's own projection)
Effort
high
Risks
Capital cost and delay; dated 2022 plan; automation, not AI; staff displacement.
How to adopt
Needs high parcel volume at hubs and capital. Start with measurement and weighing scanners and a sorting belt at the busiest hub. Running it needs maintenance and integration with billing by weight.
How a room runs it
the ROOM for a courier's hub and customer-care desk: the hub manager, the care lead and a courier-operations teammate; the agent reads the hub scan events and weight discrepancies the system already holds and GENERATES the care lead's tasks (re-bill the 38 mismatched weights, answer these three seller disputes) as plan steps with owners and due dates, and a DECISION for the hub manager when a discrepancy pattern repeats (recalibrate, claim from a seller) with a deadline; the humans keep the claims and the sign-off; the receipt records the original weight, the corrected one and who approved
The engagement on top
a custom engagement: the courier's hub system wired to the room through connectors loaded on this pull, a courier-operations professional on the match, the courier pack's templates
Your own team builds it. Across the 8 adoptions here, 0 were low effort, 3 medium and 5 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs machine sensors or logs and history of failures. Start with a proof of concept on a few critical machines. Running it needs data engineers and maintenance staff to act on alerts.
Factory digital twins on NVIDIA Omniverse, built before the plant opened
Where
TW, East Asia
Segment
enterprise
Problem
slow factory start-up
Technology
NVIDIA Omniverse, Metropolis
Who uses it
Wistron (Taiwan), with NVIDIA; Foxconn also builds twins
Outcome
Simulating first brought the factory online in 2.5 months instead of 5 and raised worker efficiency by over 50% (Wistron, via NVIDIA, Jun 2024)
Effort
high
Risks
Twin drift from reality; vendor-reported results; high build cost.
Next: Wistron is extending its twins to data centres for testing new server systems source
How to adopt
Model one line, compare against actuals, expand to the whole plant.
How a room runs it
the plant room: the plant manager, the layout engineer and a simulation agent; the agent runs the twin and GENERATES layout and commissioning tasks with due dates, each a proposal the manager approves; the receipt records the scenario and the decision
The engagement on top
connectors to the 3D plant data, a simulation engineer
Industrial internet platform linking consumers, suppliers and factory sensors for mass customisation; six Haier factories are in the World Economic Forum Lighthouse network, the Tianjin washing-machine plant using 5G, IoT, automation and analysis
Where
CN, East Asia
Segment
enterprise (appliances)
Problem
slow customisation and high cost and energy in appliance plants
Technology
industrial internet platform (COSMOPlat), 5G, IoT sensors, automation, advanced analytics
Who uses it
Haier (China), COSMOPlat
Outcome
Tianjin washing-machine factory productivity up about 30 % and energy use down 35 % (China Daily report; no measurer named)
Effort
high
Risks
Figures lack a named measurer; large capital; platform lock-in.
How to adopt
Connect the line to a data platform first; digitise orders and machine data on one line; add automation step by step; expect long programme and platform integration.
How a room runs it
the ROOM for a plant operations team: the plant manager, the production planner and a production teammate; the agent teammate reads the data the operator already has and GENERATES the line tasks (rebalance line 3 for the custom orders due Friday) as plan steps with owners and due dates, and the schedule-change proposal for the plant manager with the order and energy data as a decision block in the Inbox with a deadline; what the humans keep: the line changes and the customer commitments; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the MES, the line sensors, the order system), an industrial-engineering consultant joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Used on a battery-inspection machine at thyssenkrupp and on soldering machines at Siemens Erlangen; global rollout planned from 2025 (Siemens, Nov 2024) the company’s own figure
Effort
medium
Risks
Wrong code or fix suggestions; on-premises data needs; no published metrics.
Next: thyssenkrupp to roll the engineering copilot out across its machines at global locations from 2025 source
How to adopt
Pilot on one machine type, keep an engineer approving, widen by site.
How a room runs it
the plant room: the controls engineer, the shift operator and a copilot agent; the agent drafts control code and GENERATES troubleshooting steps from manuals as tasks with due dates, each a proposal a human approves; the receipt records the error code, the suggestion and the approver
The engagement on top
connectors to the engineering portal and spare-parts lists, a controls engineer
National Lighthouse Industry programme: 29 companies across chemicals, food, cement, automotive, textiles and pharma use AI for process optimisation, traceability, demand forecasting and precision assembly
Where
ID, Southeast Asia
Segment
mid and enterprise (29 designated companies)
Problem
wider AI adoption; most firms stay at the 'AI Aware' stage
Technology
AI process optimisation, traceability, demand forecasting, robotics
Who uses it
Ministry of Industry (Kemenperin) and 29 designated companies (Indonesia)
Outcome
none stated
Effort
high
Risks
SMEs cannot afford the entry cost; legacy integration; results not published for the 29 firms.
How to adopt
Needs digital production data and a selected process to optimise. Start with traceability or demand forecasting in one plant. Running it needs data skills and integration with legacy systems; the source names skills, SME cost and legacy integration as barriers.
How a room runs it
the room for a mid-size factory that is still at the 'AI Aware' stage: the owner, the plant manager, the planner and a planning teammate; the agent reads the order book and stock the company already has and GENERATES the weekly production plan as steps with owners and due dates, and the buy-or-wait question as a proposal in the owner's Inbox; the humans keep the line, the purchase above the envelope and the customer promise; the receipt records the plan and what was done
The engagement on top
the company's ERP and spreadsheets wired through connectors loaded on this pull; a production-planning professional on the match; budgets degrade before they block, which answers the cost barrier the source names
AI cameras and a correction laser check that robot-welded studs on vehicle frames are placed correctly at the Spartanburg body shop; humans fix flagged faults
Where
US, North America
Segment
enterprise (automotive)
Problem
manual inspection of about 500,000 studs a day
Technology
computer vision cameras, AI inspection, correction laser
Who uses it
BMW Group, Plant Spartanburg (United States)
Outcome
stud-correction saves more than USD 1 million a year (BMW manager, via BMWBlog); correction laser about five times more efficient than expected
Effort
medium
Risks
Secondary source; savings are the company's claim; some workers were moved to other jobs.
How to adopt
Start with one inspection point with high manual effort; capture labelled images; run in advisory mode beside inspectors; expect camera installation and retraining of staff.
How a room runs it
the ROOM for a body-shop quality team: the quality lead, the shift supervisor and an inspection teammate; the agent teammate reads the data the operator already has and GENERATES the shift tasks (re-check the flagged frame batch on line 2 before the next shift) as plan steps with owners and due dates, and the stop-or-continue proposal for the quality lead with the flagged images as a decision block in the Inbox with a deadline; what the humans keep: the fix on the flagged vehicle and the release decision; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the vision-system results, the MES, the quality log), a manufacturing-quality engineer joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
Digitalisation of output tracking at every production stage since 2010 and automation since 2016; capacity raised from about 90 to 100 million pieces a year without new factories or workers
Where
ID, Southeast Asia
Segment
enterprise (garment)
Problem
capacity and precision across 23 factories and about 37,000 workers
Technology
production digitalisation, automation of sewing steps
Who uses it
PT Pan Brothers Tbk (Indonesia)
Outcome
capacity from about 90 to 100 million pieces a year without new factories or workers (company's corporate secretary)
Effort
high
Risks
Dated (2019); capital cost; labour displacement. Digitisation and automation, not AI.
How to adopt
Needs output tracking at each production stage and capital for automation. Start with digital tracking of output per stage, then automate the slowest steps. Running it takes engineers and operator retraining.
How a room runs it
the room for a garment factory's week: the factory manager, the line supervisors, the planner and a production-planning teammate; the agent reads the stage-by-stage output the factory already tracks and GENERATES the line-balancing and order-follow-up tasks for supervisors as plan steps with owners and due dates, and a late-order or overtime proposal in the manager's Inbox with a deadline; the humans keep the floor, the buyer commitment and the overtime spend; the receipt records each task
The engagement on top
the factory's production-tracking system and ERP wired through connectors loaded on this pull; WhatsApp for line supervisors (the channel is built); an apparel industrial engineer as the paid professional on the match
Highly automated electronics plant: about 1,000 Simatic controllers and a manufacturing execution system storing around 50 million process records a day, products communicating with assembly machines
Where
DE, Europe
Segment
enterprise (electronics)
Problem
defects and manual handling in electronics assembly
Technology
PLC automation, MES, product-to-machine communication
Who uses it
Siemens, Amberg electronics plant (Germany)
Outcome
defect rate 0.0022 % (stated by the plant head, Siemens)
Effort
high
Risks
Dated (2015) and not AI; figure is the owner's; vendor-owned showcase plant.
How to adopt
Standardise machine-to-product data capture and an execution system on one line; use the records for root-cause analysis; expect heavy capital and a decade-long programme.
How a room runs it
the ROOM for an electronics-plant quality team: the plant manager, the quality engineer and a process-data teammate; the agent teammate reads the data the operator already has and GENERATES the analysis tasks (investigate the defect cluster on line 4 from yesterday's records by noon) as plan steps with owners and due dates, and the process-change proposal for the plant manager with the data attached as a decision block in the Inbox with a deadline; what the humans keep: the line change and the release; the receipt records every finished task and every decision, and the client receives the reports and gives direction there
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the MES records, the PLC data, the test-station logs), an industrial-automation engineer joins as the paid professional on the match, and the sector pack carries the vocabulary and the task templates
OEE monitoring: real-time machine data from an IoT platform shows production results and issues at once; the company reports rising OEE and data accuracy up about 10 %
Where
ID, Southeast Asia
Segment
enterprise (food)
Problem
hidden downtime and waste on the production floor
Technology
IoT platform (Antares, Telkom), OEE monitoring
Who uses it
PT Garudafood Putra Putri Jaya Tbk, Telkom (Indonesia)
Outcome
data accuracy up about 10 % (company's own figure); no OEE figure published the company’s own figure
Effort
medium
Risks
Source is a vendor release; retrofit effort on old machines; data security. No AI named.
How to adopt
Needs machines that can send data through an IoT gateway. Start with real-time OEE monitoring on one line. Running it needs network support and a routine for acting on the numbers.
How a room runs it
the room for a production line's week: the plant manager, the line supervisor, the operations-excellence lead and a production teammate; the agent reads the OEE and downtime data the platform already collects and GENERATES the shift's follow-up tasks for the supervisor (check the filler, change the part) as plan steps with owners and due dates, and a planning proposal in the manager's Inbox; the humans keep the line, the changeover and the spend above the envelope; the receipt records the finished tasks
The engagement on top
the platform's data feed and the ERP wired through connectors loaded on this pull; an industrial engineer joins as the paid professional on the match; the sector pack carries the downtime-to-task templates
Source: vendor · 2023 · vendor only; to corroborate
Machine-learning predictive maintenance and real-time monitoring of production machinery, run as a proof of concept at the state digital-industry centre
Where
ID, Southeast Asia
Segment
enterprise (pharma)
Problem
early detection of production-machine faults
Technology
System Invariant Analytics Technology (SIAT, NEC Indonesia with Bogor Technology), machine learning
Who uses it
PT Kalbe Farma Tbk, Ministry of Industry (BPSDMI, PIDI 4.0) (Indonesia)
Outcome
PoC rated 8 of 10 by Kalbe Farma and met two of three key targets (reported by RRI)
Effort
high
Risks
False alarms or missed failures; sensor retrofit cost; the source names no targets or results.
How to adopt
Needs machine sensors or logs and history of failures. Start with a proof of concept on a few critical machines. Running it needs data engineers and maintenance staff to act on alerts.
How a room runs it
the ROOM for a plant's maintenance week: the plant manager, the maintenance supervisor, the quality head and a maintenance teammate (an agent definition with the plant skills); the agent reads the alerts the monitoring system already raises and GENERATES the inspection tasks for the supervisor as plan steps with owners and due dates, and a stop-the-line-or-continue question as a proposal in the plant manager's Inbox with a deadline; what the humans keep: stopping a line (the named manager signs it), the quality release, the spend above the envelope; the receipt records alert, task, decision and signer
The engagement on top
the plant's monitoring system and its CMMS wired through connectors loaded on this pull; a maintenance engineer joins as the paid professional on the match; the sector pack carries inspection and line-stop templates
AI-enabled humanoid robot pilot and AI quality monitoring at a digital factory campus
Where
DE, Europe
Segment
enterprise
Problem
repetitive and hazardous logistics tasks
Technology
Google DeepMind Gemini Robotics, Apptronik Apollo
Who uses it
Mercedes-Benz, Germany (Berlin-Marienfelde Digital Factory Campus)
Outcome
Apollo humanoid first tested on intra-logistics, with AI also applied to quality control (WardsAuto, Mar 2025)
Effort
high
Risks
Worker safety; trial not yet at scale; union and worker acceptance.
Next: the Apptronik partnership is expected to expand into other manufacturing automation source
How to adopt
Pick repetitive, hazardous tasks, test on one campus, widen after a safety review.
How a room runs it
the plant room: the logistics lead, the safety officer and a robot-task agent; the agent GENERATES task lists for the robots with time windows, each a proposal the lead and the safety officer approve; the receipt records the task, the approver and any stop
The engagement on top
connectors to the plant systems, a safety officer on the match
Generative AI assistant for control-code engineering and shop-floor troubleshooting
Where
DE, Europe
Segment
enterprise
Problem
slow automation engineering and machine fault finding
Technology
Siemens Industrial Copilot
Who uses it
Schaeffler, Germany, with Siemens
Outcome
Engineering time and error risk described as substantially reduced; no figure published (Oct 2024)
Effort
high
Risks
Wrong control code on live equipment; no published metrics; vendor lock-in.
Next: extension to planning, validation and unit-test generation, plus continuous equipment monitoring source
How to adopt
Start with code drafts reviewed by an engineer, then add shop-floor fault finding.
How a room runs it
the plant room: the automation engineer, the maintenance lead and a code and fault agent; the agent drafts control code and GENERATES repair steps as tasks with due dates, each a proposal a qualified engineer approves before it touches a machine; the receipt records prompt, draft and approver
The engagement on top
connectors to the engineering portal and manuals library, a controls engineer
Generative AI synthetic images to train optical-inspection AI, piloted at two German plants
Where
DE, Europe
Segment
enterprise
Problem
slow set-up of AI quality inspection
Technology
generative AI, machine vision
Who uses it
Bosch, Germany (Hildesheim and Stuttgart-Feuerbach plants)
Outcome
Hildesheim cut ramp-up cycle times 15%; Feuerbach cut component-test time from 3.5 to 3 minutes; AI planning time expected to fall from 6-12 months to a few weeks (Bosch, Dec 2023)
Effort
medium
Risks
Synthetic data may not match real defects; false accepts in inspection.
Next: after the pilots, offer the synthetic-data service to all Bosch locations source
How to adopt
Pilot at two plants, measure ramp-up, then offer the service to the network.
How a room runs it
the plant room: the quality engineer, the line supervisor and an inspection agent; the agent GENERATES synthetic defect images and test tasks with dates, each batch a proposal the quality engineer approves; the receipt records the data set and the approver
The engagement on top
connectors to the vision systems, a quality engineer
Your own team builds it. Across the 12 adoptions here, 0 were low effort, 4 medium and 8 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Researched in Media, creative & marketing agencies
Maturitypilot 1live 10
Evidencevendor 4press 6paper 1
Adopt first
Needs licences for generative tools and a written policy on copyright and ethics. Start with draft copy and image variations for one client team, with creative review. Running it takes training and policy review.
State-owned firms adopted an AI-driven customer engagement platform for personalised, real-time messaging across app, email and WhatsApp, with AI content governance
Where
ID, Southeast Asia
Segment
enterprise
Problem
personalising communication at scale to 150 million-plus customers; retention
Technology
Braze customer engagement platform with AI, real-time data architecture
Who uses it
Telkomsel (153.5 million customers), Pertamina Patra Niaga (still in development), Indonesia
Outcome
none stated
Effort
medium
Risks
Over-messaging and consent breaches; vendor lock-in; one adopter still developing it; no results published.
How to adopt
Needs a customer data platform and consent records for messaging. Start with personalised messages on one channel with content rules. Running it needs content governance and data integration.
How a room runs it
the ROOM for the CRM team: the head of CRM, the campaign manager, the legal reviewer and a lifecycle teammate; the agent GENERATES the month's journeys as tasks with owners and due dates (draft the retention message, prepare the segment, queue the legal check) and files in the head's Inbox a proposal for each experiment (launch or not) with the send date as deadline; the humans keep the offer, the content sign-off and the audience rules; the receipt records who approved each message and the audience it reached; the monthly send budget degrades to fewer sends before it blocks
The engagement on top
a custom engagement: the company's own customer data platform, app messaging tool and WhatsApp Business account wired to the room through connectors loaded on this pull; a CRM consultant joins as the paid professional on the match; the marketing pack carries journey and test templates
A multinational agency's Indonesian creative teams use AI broadly in advertising content production, with facilities provided by the agency and a policy stance on copyright and ethics
Where
ID, Southeast Asia
Segment
enterprise
Problem
faster content production under client pressure; copyright, regulatory and ethics exposure
Technology
generative AI for creative production (tools not named in the study)
Who uses it
Dentsu Indonesia, Indonesia
Outcome
none stated
Effort
low
Risks
Copyright and ownership of generated content; creative staff concern; client confidentiality. Qualitative interviews only.
How to adopt
Needs licences for generative tools and a written policy on copyright and ethics. Start with draft copy and image variations for one client team, with creative review. Running it takes training and policy review.
How a room runs it
the ROOM for one client account: the account director, the creative lead, the strategist and a production teammate carrying the brand guide; the agent GENERATES the brief's tasks as plan steps with owners and due dates (three concept routes by Tuesday, resize set for six placements, claim check against the legal list) and files the choice of route in the client lead's Inbox as a proposal with the presentation date as deadline; the humans keep the idea, the client relationship and the copyright and ethics sign-off; the receipt records which assets were machine-assisted, who cleared them and the cost against the account's budget
The engagement on top
a custom engagement: the agency's own project tool, DAM and trafficking system wired to the room through connectors loaded on this pull; an IP/media lawyer joins as the paid professional on the match; the marketing pack carries brief, review and clearance templates
A snack maker built a generative-AI tool with an agency and a consultancy that produces social content and product pages at lower cost, aiming at short TV ads next
Where
US, North America
Segment
enterprise
Problem
Marketing content for many brands and markets is costly to produce
Technology
In-house generative AI tool built with Publicis Groupe and Accenture; more than US$40 million invested
Who uses it
Mondelez International, United States (Chips Ahoy; Milka in Germany)
Outcome
30 to 50 % lower content production cost (Mondelez's own range, unaudited); more than US$40 million invested
Effort
high
Risks
Brand and likeness risk; unaudited savings; significant build cost.
How to adopt
Build or buy a content generator trained on approved brand assets, keep a human approval step per asset, start with social and product pages before video.
How a room runs it
the ROOM for brand content: the brand manager, the agency producer and a content teammate; the teammate GENERATES the week's content calendar as plan steps with owners and due dates (asset, market, language, approver) and files each ad for release as a decision in the brand lead's Inbox with a deadline; the humans keep the brand sign-off and the choice of what is not generated (no human likenesses); the receipt records each asset, its prompt and its approver
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the asset library and the ad platforms), and a brand-content producer joins as the paid professional on the match
A consumer-goods group made digital twins of its products so product imagery is produced by AI instead of photo shoots, across its beauty brands
Where
GB, Europe
Segment
enterprise
Problem
Product shoots are slow and costly for many variants, labels and languages
Technology
NVIDIA Omniverse digital twins, OpenUSD standard, AI image tools; designed by Collective
Who uses it
Unilever, United Kingdom (TRESemmé, Dove, Vaseline, Clear)
Outcome
55 % savings and 65 % faster content turnaround (Unilever's figures; which is cost and which is speed is not stated)
Effort
high
Risks
Up-front 3D modelling; supplier standards; unaudited figures.
How to adopt
Build one digital twin per product with its variants and labels, adopt a shared scene standard with suppliers, pilot on one brand line.
How a room runs it
the ROOM for product imagery: the brand manager, the e-commerce lead and an imagery teammate; the teammate GENERATES the shot list as plan steps with owners and due dates (variant, retailer format, language) and files the release of each batch as a decision in the brand manager's Inbox with a deadline; the humans keep the brand check and the retailer requirements; the receipt records each asset version and its approver
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the product-data and asset systems), and a 3D content producer joins as the paid professional on the match
An agency leader states AI is used to read changing consumer behaviour and give marketers regular updated insight, with people guiding adoption
Where
ID, Southeast Asia
Segment
enterprise
Problem
marketers need current insight into shifting consumer behaviour
Technology
AI-based consumer insight (tools not named)
Who uses it
Dentsu Indonesia (CEO Elvira Jakub), Indonesia
Outcome
none stated
Effort
medium
Risks
Insight quality depends on data; leader statement only, no results; privacy.
How to adopt
Needs consumer data sources and analytics tools. Start with a regular insight report for marketers that people review. Running it needs analysts who guide the tool.
How a room runs it
the ROOM for the marketing lead of a brand: the CMO, the brand manager, the agency planner and an insight teammate; the agent reads the reports the brand already receives and GENERATES the fortnight's tasks as plan steps (which audience to re-brief, which message to test, which report to ask the agency for) with owners and due dates, and a decision for the CMO in the Inbox (shift budget between channels or hold) with the planning date as deadline; the humans keep the strategy and the spend above the envelope; the receipt records the proposal, the choice and the budget it moved
The engagement on top
a custom engagement: the brand's own research reports, ad accounts and CRM wired to the room through connectors loaded on this pull; a media planner joins as the paid professional on the match; the marketing pack carries planning and test templates
Group-wide AI layer (CoreAI) on a shared data and talent platform, available to all 100,000 staff
Where
FR, Europe
Segment
enterprise
Problem
disconnected data and slow production across insight, media and creative
Technology
CoreAI (Publicis Sapient), Marcel platform
Who uses it
Publicis Groupe (France)
Outcome
CoreAI made available to all 100,000 staff; built on 2.3 billion consumer profiles and about a petabyte of assets on the Marcel platform; group organic growth 6.3% in 2023 (Publicis figures via Campaign Asia)
Effort
high
Risks
Consumer-profile privacy; over-automation of media decisions; unproven margin payoff.
Next: CoreAI capabilities roll out iteratively through the first half of 2024 across insight, media, creative and production, software and operations source
How to adopt
Unify data first, release capabilities in stages, train staff, measure per-function productivity.
How a room runs it
the client-account room: the account lead, the data analyst, the creative lead, the client's marketer and a media teammate; the agent drafts audience insight and media plans and GENERATES the review tasks per specialist with due dates, and each budget shift becomes a proposal in the account lead's Inbox with a deadline; the humans keep the strategy; the receipt records the data used and the approver
The engagement on top
connectors to the client's customer data and ad accounts, an account lead on the match, a plan-review task template set
Brands run clipping campaigns on a marketplace: the brand posts source material and a reward per 1,000 views, clippers post short cuts on TikTok / Reels / Shorts, the platform reads the view counts and pays out per verified thousand up to a per-video cap
Where
US, global
Segment
SMB and creators (creator economy)
Problem
distribution without an ad budget: paying strangers per verified view instead of per impression; view fraud and unreviewed clips
Technology
Whop Content Rewards (campaign marketplace: reward rate per 1,000 views, minimum payout before review, per-video maximum); clipping tools (Opus, Vyro)
Who uses it
Whop (marketplace), brands and agencies posting campaigns, clippers (individuals), United States and global
Outcome
median USD 1.00 per 1,000 verified views across 50 campaigns (Foundry, Sep 2026; range 0.05–3.00) the company’s own figure
Effort
low
Risks
View fraud; platform terms; 'quick money' positioning (refused: STRATEGY §5a angle 6); payouts need the processor rail (SE-1).
How to adopt
Needs a platform account with API access for the view read and a funded hold. The brand sets the rate, the cap and the date; clippers arrive through the public call.
How a room runs it
the ROOM for a brand's campaign: the brand, its marketer and a campaign teammate; the campaign is an OPEN CALL — a deliverable kind of the marketing pack whose suppliers are any person with a verified handle; the brand's hold on its own balance; a clipper's clip is a proposal; the check is a PLATFORM READ of the view count through a connection on a schedule, never a screenshot; the draw per verified 1,000 views to the clipper's earnings line; the receipt credits the clipper and carries the campaign's public page; no take (D283); the teammate itself cuts the first clips (transcript → segments → ffmpeg trim and captions in the sandbox) and the clipper finishes; blueprint 42 law 1: the kind is pack data, the open call, the connection-read verifier and the handle method are generic kernel objects every field uses
The engagement on top
a custom engagement: the brand's platform accounts wired as connections for the view read; the marketing pack's kinds (clip, UGC video, KOL post, affiliate link) and their verifier rows; a performance marketer or agency as the paid professional on the match
Advertisers moved budget into Meta's AI-optimised Advantage+ campaigns and used an in-platform Opportunity Score to tune performance
Where
ID, Southeast Asia
Segment
SMB and mid
Problem
lead and sales cost per result during peak periods (Ramadan, mega-sale days)
Technology
Meta Advantage+ Sales and Leads campaigns, Opportunity Score in Ads Manager
Who uses it
Wookey Weight, Masela Adawiyah 99 (Indonesian advertisers named by Meta), Indonesia
Outcome
14 % lower cost per lead (Meta's figure, unaudited); one advertiser: 17 % more leads at 15 % lower cost per lead the company’s own figure
Effort
low
Risks
Platform lock-in; results unaudited and reported by the vendor; opaque targeting.
How to adopt
Needs an ad account and conversion tracking set up. Switch campaigns to the platform's automated format and use its score tool to tune. Running it needs budget and monitoring of results.
How a room runs it
the ROOM for a small brand's campaign week: the owner, the person who runs ads and a campaign teammate; the agent reads the ad account the shop already has and GENERATES the week's tasks as plan steps (new creatives by Wednesday, pause two ad sets, refresh the Ramadan offer) with owners and due dates, and a decision in the owner's Inbox (move budget into the better campaign type or keep) with the day the promotion ends as deadline; the owner keeps the spend above the daily envelope and the offer; the receipt records each budget move and its result; budgets degrade before they block
The engagement on top
a custom engagement: the shop's own Meta ad account, store or POS and WhatsApp catalogue wired to the room through connectors loaded on this pull; a performance marketer joins as the paid professional on the match; the marketing pack carries the campaign-week task templates
Source: vendor · 2025 · vendor only; to corroborate
Generative AI inside the ad platform's creative tools, plus a joint effort to build a Japanese ad-copy model
Where
JP, global
Segment
enterprise
Problem
slow creative production and copy that misses Japanese nuance
Technology
Copilot in Microsoft Advertising; Sarashina LLM (SB Intuitions)
Who uses it
Dentsu (Japan, global agency); Dentsu Digital, SoftBank Corp., SB Intuitions
Outcome
Workflows up to 6x faster; a task of about 30 minutes now takes about 2 minutes (Dentsu via Microsoft, Jul 2025) the company’s own figure
Effort
low
Risks
Off-brand or inaccurate claims; vendor-reported speed-ups; language nuance.
Next: in Sep 2025 Dentsu, Dentsu Digital, SoftBank Corp. and SB Intuitions began joint research on a generative AI specialised in Japanese copywriting, with uses beyond copy in naming and marketing research source
How to adopt
Apply the platform's AI tools to routine variants, time the tasks, then fine-tune on the firm's own copy.
How a room runs it
the creative room: the account lead, the copywriter, the client's marketer and a media teammate; the agent drafts ad variants and GENERATES the review tasks per copywriter with due dates, and each claim or tone change becomes a proposal in the account lead's Inbox with a deadline; the humans keep the brand voice; the receipt records the draft, the model and the approver
The engagement on top
connectors to the ad accounts and the brand guide, a copy lead on the match, a variant-review task template set
Generative content engine on a 3D platform for brand-consistent advertising, later scaled to a global beverage client's 100 markets
Where
GB, Europe
Segment
enterprise
Problem
hand-producing hundreds of thousands of ad assets
Technology
NVIDIA Omniverse Cloud and NIM microservices, Adobe Firefly, Getty Images generative models
Who uses it
WPP (UK), with NVIDIA; client Coca-Cola, Ford
Outcome
By July 2024 thousands of WPP production experts could access the tools and Coca-Cola was among the first clients scaling generative 3D across its 100 markets; WPP is first to use NVIDIA's OpenUSD NIM microservices (WPP, Jul 2024) the company’s own figure
Effort
high
Risks
Copyright of training data; off-brand output; client approval bottlenecks.
Next: WPP Open X will roll out the Prod X production studio so Coca-Cola scales generative 3D across its 100 markets source
How to adopt
Start with one product's 3D configurator, license-cleared image sources, per-market review.
How a room runs it
the campaign room: the account lead, the creative director, the client's brand owner and a production teammate; the agent generates asset variants and GENERATES the review tasks per market with due dates, and each brand-rule exception becomes a proposal in the brand owner's Inbox with a deadline; the humans keep the creative direction; the receipt records the asset, the model source and the approver
The engagement on top
connectors to the asset library and the ad platforms, a creative director on the match, a per-market review task template set
A search company launched an AI marketing platform that suggests media buying and generates ads from search intent; beta advertisers reported higher conversion
Where
CN, East Asia
Segment
enterprise
Problem
Advertisers spend hours building campaigns and targeting
Technology
QingGe AI marketing platform with the QingDuo AIGC toolkit
Who uses it
Baidu, China
Outcome
Up to 20 % higher ad conversion among early beta adopters (Baidu and beta participants; no method given)
Effort
low
Risks
Beta figures; opaque targeting; dependence on one ad platform.
How to adopt
Connect the ad account to a platform that drafts targeting and creative, run it beside current campaigns on a small budget, compare conversion.
How a room runs it
the ROOM for a campaign: the marketer, the media buyer and a campaign teammate; the teammate GENERATES the campaign brief into plan steps with owners and due dates (audience, creative variants, budget split) and files the spend above the envelope as a decision in the marketer's Inbox with a deadline; the humans keep the budget and the brand claims; the receipt records each variant, its spend and the result
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the ad accounts and the analytics), and a performance marketer joins as the paid professional on the match
Your own team builds it. Across the 11 adoptions here, 5 were low effort, 2 medium and 4 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Researched in Media, creative & marketing agencies
Maturitypilot 2live 7
Evidencevendor 1press 7analyst 1
Adopt first
Needs a newsroom policy on AI use and editor review. Start with repackaging content for social and headline suggestions, with editors approving. Running it needs training and a correction process.
A broadcaster-streamer built a studio generative-AI platform (VidioGen) on Google Cloud models and used it to re-produce a 2013 animated series: inpainting and outpainting of frames, shot composition, long-context consistency against a production bible
Where
ID, Southeast Asia
Segment
enterprise
Problem
re-developing an existing animation costs time and money; visual drift between shots
Technology
Google Cloud Gemini Enterprise Agent Platform, Veo, Imagen, Gemini; inpainting/outpainting
Who uses it
Emtek Group (Vidio, Mentari TV), Indonesia
Outcome
30 % cut in re-development time and cost; audience share growth of 74 % (free-to-air prime time) and 90 % (Sahur slot) (trade-press report, source of figures not named)
Effort
high
Risks
Rights and copyright; cost of compute; quality consistency; figures unaudited and mixed with a Ramadan effect.
How to adopt
Needs rights to the original series, a production bible and access to cloud generative models. Start with re-producing one archive title. Running it needs studio artists plus a platform team.
How a room runs it
the ROOM for one production: the producer, the VFX lead, the director and a production teammate that holds the production bible; the agent GENERATES the episode's task list as plan steps (frames to extend, props to repaint, shots to compose) with owners and due dates, and files a decision in the producer's Inbox when a generated shot drifts from the bible (accept, redo, or hand to an artist) with the delivery date as the deadline; the artists keep the creative sign-off and the final polish; the receipt records which shots were machine-assisted and by whom approved, and the compute spend against the episode budget, which degrades to fewer variants before it blocks
The engagement on top
a custom engagement: the studio's own asset store and render queue wired to the room through connectors loaded on this pull; a post-production supervisor joins as the paid professional on the match; the media pack carries the episode task templates; reviews are collected over WhatsApp
Newsrooms and broadcasters use AI for graphics, virtual presenters, repackaging TV content for social, automated short articles and headline suggestions to meet production targets with smaller teams
Where
ID, Southeast Asia
Segment
mid and enterprise
Problem
staff and budget cuts against the same output targets; falling organic traffic
Technology
generative image tools, AI virtual presenter, automation tools (Zapier) for short articles, chatbots for story ideas, LLM headline suggestions
Who uses it
iNews, Tech in Asia, Republika (Aiman & Aisha chatbot), Jakarta Globe, Tempo, Indonesia
Outcome
At least a dozen news organisations in Indonesia had adopted AI as of 2024 (Alliance of Independent Journalists, per Project Multatuli; no method)
Effort
low
Risks
Errors and fabricated content; job displacement; trust; critical source with no results.
How to adopt
Needs a newsroom policy on AI use and editor review. Start with repackaging content for social and headline suggestions, with editors approving. Running it needs training and a correction process.
How a room runs it
the ROOM for one desk: the managing editor, the graphics lead, the social editor and a production teammate (an agent definition with the newsroom's skills and style guide); the agent reads the day's rundown and the archive the desk already has and GENERATES the shift's tasks as plan steps with owners and due dates (illustration brief for the fuel-price segment by 14:00, three headline alternatives for the editor, the social cut of the evening bulletin); the editor's decision lands in the Inbox as a proposal with a deadline (publish the AI-assisted piece or hold); the humans keep the byline, the fact check and the sign-off; the receipt records who approved what, which draft was machine-made and the spend against the desk's budget, which degrades to suggestions only before it blocks
The engagement on top
a custom engagement: the newsroom's own CMS, rundown system and asset library wired to the room through connectors loaded on this pull; a media lawyer or standards editor joins as the paid professional on the match for the AI-use policy; the media pack carries task templates (headline set, social cut, correction note); editors reach the room on WhatsApp
A business newspaper built a subscriber chatbot on its own article retrieval model, plus an in-house tool that drafts short digests that editors review
Where
JP, East Asia
Segment
enterprise
Problem
External models do not handle Japanese well; readers cannot form questions
Technology
Ask! NIKKEI (own retrieval model plus a switchable external LLM), Nikkei Tailor digest tool
Who uses it
Nikkei, Japan
Outcome
none stated
Effort
high
Risks
Rights exclusions; model drift when switching suppliers; no published accuracy measure.
How to adopt
Build or choose a retrieval step for your language, limit sources by age and rights, tell the model to use only retrieved articles, refuse when none are found.
How a room runs it
the ROOM for a subscriber assistant: the editor-in-chief, the legal counsel and a reader-answers teammate; the teammate GENERATES the content-rights plan as steps with owners and due dates (which articles may be used, which are excluded) and files each model change as a decision in the engineering lead's Inbox with a deadline; the humans keep the editorial line and the rights call; the receipt records the articles each answer used
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the article database and the app), and an applied-AI engineer joins as the paid professional on the match
Generative AI used for final on-screen visual effects in an original series, with the in-house studio
Where
US;AR, North America
Segment
enterprise
Problem
a building-collapse shot beyond the show's budget
Technology
generative AI VFX tools (Eyeline Studios, Netflix)
Who uses it
Netflix (USA), series El Eternauta (Argentina)
Outcome
VFX sequence completed 10 times faster than with traditional tools (co-CEO Ted Sarandos, Q2 2025 earnings call), and not feasible at the show's budget otherwise
Effort
medium
Risks
Creative-labour concerns; rights and likeness; visual artefacts.
Next: co-CEO Greg Peters said Netflix may use generative AI for spoken-request content search and to help brands create advertising source
How to adopt
Apply to one budget-limited shot, keep artists in the loop, disclose use, then widen.
How a room runs it
the production room: the producer, the VFX supervisor, the director and a pipeline teammate; the agent drafts shot options and GENERATES the review tasks per artist with due dates, and each shot sign-off becomes a proposal in the producer's Inbox with a deadline; the humans keep the creative approval; the receipt records the shot, the tool used and the approver
The engagement on top
connectors to the asset manager and review tool, a VFX supervisor on the match, a shot-review task template set
AI-generated three-bullet summaries on news stories for terminal users, reviewed by subject-matter experts
Where
US, North America
Segment
enterprise
Problem
professionals cannot read every story and must pick what to read in full
Technology
large language model summarisation inside the Bloomberg Terminal
Who uses it
Bloomberg (USA)
Outcome
Three-bullet AI summaries at the top of Bloomberg News stories on the terminal from 15 Jan 2025, following AI earnings-call summaries; subject-matter experts evaluate output (Bloomberg via Talking Biz News)
Effort
medium
Risks
Summaries that misstate a story; market-moving errors; source attribution.
How to adopt
Start with one content type, keep expert review of output, expand sources after quality checks.
How a room runs it
the newsdesk room: the editor, the reporter, the data editor and a summary teammate; the agent drafts the summary and GENERATES the check tasks per desk with due dates, and each correction or sensitive story becomes a proposal in the editor's Inbox with a deadline; the humans keep the editing; the receipt records the source, the summary and the checker
The engagement on top
connectors to the content system and the style guide, an editor on the match, a summary-check task template set
A tabloid publisher turned an internal hackathon into a fact-checked election Q&A chatbot for readers
Where
SE, Europe
Segment
mid
Problem
readers had questions on the EU elections that editors could not answer one by one at volume
Technology
LLM chatbot grounded in the newsroom's fact-checked material
Who uses it
Aftonbladet (Schibsted), Sweden
Outcome
none stated
Effort
medium
Risks
Wrong or misleading answers on a sensitive topic; reputational risk; no results published.
How to adopt
Needs a newsroom with a trusted archive and fact-checkers. Start with a bounded Q&A bot on one topic (an election) with answers grounded in vetted material. Running it needs editorial review and monitoring.
How a room runs it
the ROOM for a public service desk: the deputy editor, the fact-checker, a reporter and a reader-service teammate that answers only from the desk's checked material; the agent GENERATES the tasks the desk needs (add the new fact-check to the answer base by 12:00, review the ten questions it could not answer) and puts the unanswerable questions in the editor's Inbox as proposals (write a piece or decline) with an edition deadline; the humans keep what is published and any correction; the receipt records every answer's source and every question handed to a person
The engagement on top
a custom engagement: the publisher's own CMS and fact-check archive wired to the room through connectors loaded on this pull; a fact-checker joins as the paid professional on the match; reader questions can arrive over WhatsApp, a built channel
A national broadcaster automated video tagging, transcription and translation of its flagship news programme with AI, so the digital team publishes to YouTube with captions, chapters and descriptions the next day
Where
PH, Southeast Asia
Segment
mid and enterprise
Problem
manual tagging took 14 days per batch; slow transcription and translation; monthly-only reporting
Technology
AI transcription, translation and tag generation, YouTube auto-captioning/chaptering, Looker Studio reporting
Who uses it
ABS-CBN (TV Patrol), Philippines
Outcome
14 days of manual video tagging cut to under 20 minutes; transcription and translation from 2 hours to 15-30 minutes (stated as -87.5 %); reporting from monthly to next-day (ABS-CBN project, per Google News Initiative; measurer not named) the company’s own figure
Effort
low
Risks
Translation and transcript errors; vendor dependence. Source is the supplier's story, tools not named.
How to adopt
Needs an archive of programme video and a digital publishing channel. Start with automated transcripts, captions and chapters on one programme. Running it needs an editor to check output.
How a room runs it
the ROOM for the digital desk: the digital editor, the archive producer and a publishing teammate; each evening the agent reads the day's programme files and GENERATES the next morning's tasks as plan steps (tag review for segment 3, translated caption check, YouTube chapter proof) with owners and due times; a decision lands in the editor's Inbox when a tag or translation is low-confidence (accept or correct) with the publishing slot as deadline; the humans keep the editorial judgment on sensitive segments; the receipt records what was published, who reviewed the tags and the time saved against the previous manual cycle
The engagement on top
a custom engagement: the broadcaster's own media asset system and YouTube channel wired to the room through connectors loaded on this pull; a language editor or archivist joins as the paid professional on the match; the media pack carries the tag and caption review templates
Source: vendor · 2025 · vendor only; to corroborate
A newspaper launched an experimental generative-AI tool (Ask The Post AI, Nov 2024) that answers readers from articles the paper has published since 2016 and gives no answer when relevance is too low
Where
US, North America
Segment
enterprise
Problem
Readers cannot find answers in a large archive and generic chatbots hallucinate
Technology
Generative AI search over the newsroom archive with a relevance threshold
Index the archive, answer only from retrieved articles, refuse below a relevance threshold, label the tool experimental and ask readers to check sources.
How a room runs it
the ROOM for an archive assistant: the product editor, the standards editor and a reader-answers teammate; the teammate GENERATES the test plan as steps with owners and due dates (question sets, failure review, source checks) and files each widening of scope as a decision in the standards editor's Inbox with a deadline; the humans keep the editorial standard and the refusal rule; the receipt records every answer's source articles
The engagement on top
a custom engagement: the client's own systems stay theirs — the agent is wired to them through connectors loaded on this pull (the archive and the subscription system), and a newsroom product lead joins as the paid professional on the match
Subscriber chatbot answering from two decades of the publisher's archive with source footnotes, in beta for a professional tier
Where
GB, Europe
Segment
enterprise
Problem
readers want answers, not search results
Technology
Ask FT on Anthropic's Claude 3 with the FT's internal search
Who uses it
Financial Times (UK)
Outcome
Beta for a select number of FT Professional subscribers, answers drawn from two decades of FT content (FT, Mar 2024); no usage figures published
Effort
medium
Risks
Outdated or incorrect answers; archive rights; trust in the brand.
Next: the tool sits in FT Professional, which serves nearly 8,000 businesses, government bodies and educational institutions worldwide, the audience for a wider rollout source
How to adopt
Release to a small paying tier, show sources on every answer, audit answers, widen after quality checks.
How a room runs it
the reader-service room: the product editor, the archive editor, the subscriber-support lead and a chatbot teammate; the agent answers from the archive with footnotes and GENERATES the correction and gap tasks per editor with due dates, and each wrong or sensitive answer becomes a proposal in the editor's Inbox with a deadline; the humans keep the editorial standard; the receipt records the question, the sources and the answer
The engagement on top
connectors to the archive and the subscriber system, an editor on the match, an answer-audit task template set
Your own team builds it. Across the 9 adoptions here, 2 were low effort, 5 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs a listing platform with agents and listing history. Start with an AI description generator offered to agents, then a price assistant. Running it needs model upkeep and agent training.
A property portal gives its agents generative-AI listing descriptions, an ML listing-price assistant and a lead-quality tool; in Singapore a third of Q4 2023 listings used the description generator and over 90 % of active agents had tried lead management
Where
SG;MY;VN, Southeast Asia
Segment
mid (portal and agents)
Problem
agents spend time writing listings, pricing them and judging which leads are worth a call
Technology
generative AI listing text, ML price suggestion, lead-quality insight
Who uses it
PropertyGuru Group (Singapore, listed), its agents in Singapore
Outcome
about a third of Singapore listings used the description generator in Q4 2023; 17 % of listings used the price assistant; over 90 % of active agents had tried lead management (management statements on an earnings call)
Effort
high
Risks
Figures are from a machine transcript; inaccurate descriptions or prices; agent trust.
How to adopt
Needs a listing platform with agents and listing history. Start with an AI description generator offered to agents, then a price assistant. Running it needs model upkeep and agent training.
How a room runs it
the ROOM for an agency or a developer's sales desk: the principal, the sales agents and a listing teammate (an agent definition with the property skills); the teammate reads the listings, enquiries and viewings the agency already keeps and GENERATES the week's tasks for each agent as plan steps with owners and due dates (draft the description for unit 12B, call the three warm enquiries, book Saturday's viewings), and puts the price-change proposal in the principal's Inbox as a decision with a deadline; the humans keep the client conversation, the sign-off on price and the commission; the receipt records each listing published and each viewing held
The engagement on top
a custom engagement: the agency's own listing system or CRM wired to the room through connectors loaded on this pull, the licensed agent or marketing professional on the match, and the real-estate pack's task templates (listing, viewing, offer); enquiries arrive over WhatsApp, which is built
Source: analyst · 2024 · vendor only; to corroborate
Self-service AI platform inside the firewall with public tools blocked, used for data queries, document answers and translation
Where
US, global
Segment
enterprise
Problem
staff need answers from large troves of property data without leaking it
Technology
CBRE internal AI platform (later Ellis AI), GitHub Copilot
Who uses it
CBRE (USA)
Outcome
About 3,000 unique monthly users and about 150 new users a week by Nov 2023; AI-enabled facilities management at 20,000+ sites covering one billion square feet (CBRE via CIO Dive)
Effort
high
Risks
Data leakage if public tools are used; uneven quality; cost without measured return.
Next: CBRE says it is rolling out self-service capabilities broadly and tracks usage, cost and efficiency gains to guide more investment source
How to adopt
Block public tools, give a sandboxed internal platform, let teams self-serve, track usage and returns.
How a room runs it
the property-services room: the account director, the facilities manager, the client's occupier and a data teammate; the agent answers from building documents and GENERATES the work-order and review tasks with due dates, and each cost or vendor change becomes a proposal in the director's Inbox with a deadline; the humans keep the client decisions; the receipt records the question, the document source and the approver
The engagement on top
connectors to the building-document store and the work-order system, a facilities manager on the match, a work-order review task template set
A property portal launched a native search inside ChatGPT where users describe a home, refine and compare listings by conversation
Where
IN, South Asia
Segment
enterprise (portal)
Problem
discovery through filters is slow for first-time buyers
Technology
ChatGPT app integration over the portal's listing database
Who uses it
Housing.com (REA India), India
Outcome
none stated
Effort
low
Risks
Traffic leaks to the assistant's owner; no results published; stale listings shown in chat.
How to adopt
Publish listing search to a chat assistant as a second door, send users to the portal for contact and detail, and track which chat searches become enquiries.
How a room runs it
the ROOM for an agency or a portal's sales desk: the principal, the sales agents and a listing teammate (an agent definition with the property skills); the teammate reads the listings, enquiries and viewings the agency already keeps and GENERATES the week's tasks for each agent as plan steps with owners and due dates (answer the enquiries that matched the new search wording, refresh the descriptions that read thin, book Saturday's viewings), and puts the pricing or advertising-spend proposal in the principal's Inbox as a decision with a deadline; the humans keep the client conversation, the sign-off on price and the commission; the receipt records each listing refreshed and each viewing held
The engagement on top
a custom engagement: the agency's own listing system or CRM wired to the room through connectors loaded on this pull, the licensed agent or marketing professional on the match, and the real-estate pack's task templates (listing, viewing, offer)
Source: press · 2026 · vendor only; to corroborate
ChatGPT Enterprise for about 2,000 staff, 150 AI promotion leaders in 85 departments and about 500 custom GPTs in three months
Where
JP, East Asia
Segment
enterprise
Problem
labour shortage and slow internal information work in a large developer
Technology
ChatGPT Enterprise, custom GPTs
Who uses it
Mitsui Fudosan (Japan)
Outcome
About 2,000 employees on ChatGPT Enterprise from 1 Oct 2025; about 500 custom GPTs created within three months; estimated working-time cut of over 10% (company estimate)
Effort
medium
Risks
Confidential property data in prompts; unmanaged custom-assistant sprawl; unverified time-savings estimate.
Next: Mitsui Fudosan plans to expand AI use to address labour shortages, use internal data better and improve management decisions source
How to adopt
Name AI promotion leaders per department, run in-person training, let teams build custom assistants, track usage.
How a room runs it
the development room: the project lead, the leasing manager, the accounting owner and a data teammate; the agent drafts property summaries and press text and GENERATES the review tasks per department with due dates, and each external statement or figure becomes a proposal in the lead's Inbox with a deadline; the humans keep the decisions; the receipt records the draft, the data used and the approver
The engagement on top
connectors to the property database and the document store, a department lead on the match, a custom-assistant task template set
A national property portal lets shoppers search listings by describing what they want in plain language, including commute time, affordability, schools and nearby places, and offers saved searches with alerts
Where
US, North America
Segment
enterprise (national portal)
Problem
shoppers must translate a wish into many filters
Technology
natural-language search over listings, saved-search alerts
Who uses it
Zillow Group, United States
Outcome
none stated
Effort
high
Risks
Mis-read queries return wrong homes; fair-housing exposure when search touches schools and neighbourhoods; no accuracy data published.
How to adopt
Needs structured listing data (price, commute, school, location). Start with natural-language search in one app, keep the filter search beside it, and review the queries it misreads.
How a room runs it
the ROOM for an agency or a portal's sales desk: the principal, the sales agents and a listing teammate (an agent definition with the property skills); the teammate reads the listings, enquiries and viewings the agency already keeps and GENERATES the week's tasks for each agent as plan steps with owners and due dates (answer the enquiries that matched the new search wording, refresh the descriptions that read thin, book Saturday's viewings), and puts the pricing or advertising-spend proposal in the principal's Inbox as a decision with a deadline; the humans keep the client conversation, the sign-off on price and the commission; the receipt records each listing refreshed and each viewing held
The engagement on top
a custom engagement: the agency's own listing system or CRM wired to the room through connectors loaded on this pull, the licensed agent or marketing professional on the match, and the real-estate pack's task templates (listing, viewing, offer)
A national property portal opened a beta of conversational AI search on its homepage: the shopper describes what they want and gets a shortlist from live listings, beside the normal search
Where
GB, Europe
Segment
enterprise (national portal)
Problem
shoppers cannot express a vague brief in filters
Technology
Google Gemini models, conversational search over listings
Who uses it
Rightmove, United Kingdom (with Google Cloud)
Outcome
none stated
Effort
medium
Risks
Beta only, no outcome data; wrong shortlists erode trust in the portal; dependence on one model supplier.
How to adopt
Start with a limited beta beside the existing search, measure which briefs convert to enquiries, and widen only after the misses are reviewed.
How a room runs it
the ROOM for an agency or a portal's sales desk: the principal, the sales agents and a listing teammate (an agent definition with the property skills); the teammate reads the listings, enquiries and viewings the agency already keeps and GENERATES the week's tasks for each agent as plan steps with owners and due dates (answer the enquiries that matched the new search wording, refresh the descriptions that read thin, book Saturday's viewings), and puts the pricing or advertising-spend proposal in the principal's Inbox as a decision with a deadline; the humans keep the client conversation, the sign-off on price and the commission; the receipt records each listing refreshed and each viewing held
The engagement on top
a custom engagement: the agency's own listing system or CRM wired to the room through connectors loaded on this pull, the licensed agent or marketing professional on the match, and the real-estate pack's task templates (listing, viewing, offer)
A large developer applies OCR integrated with ChatGPT to turn complex images and documents into text, uses AI to sort incoming customer conversations by priority, and plans a generative-AI chatbot on its property e-catalogue
Where
ID, Southeast Asia
Segment
enterprise
Problem
slow document processing, slow answers to buyers, buyers finding the right product
Technology
OCR plus an LLM, AI conversation triage, planned generative chatbot, Microsoft cloud
Who uses it
Sinar Mas Land (Indonesia): Chief Transformation and Data Officer Mulyawan Gani, CIO Stefanus Mulianto
Outcome
none stated
Effort
medium
Risks
Data sent to a third-party model; extraction errors; chatbot is only planned; no results published.
How to adopt
Needs scanned documents and customer chat logs. Start with OCR plus a language model to extract text, and AI sorting of customer conversations by priority. A catalogue chatbot is planned, not live. Running it needs integration and human review.
How a room runs it
the ROOM for a developer's sales and customer desk: the head of sales, the sales agents, the legal/admin clerk and a buyer-care teammate; the teammate reads the documents (booking forms, ID scans, contracts) and the buyer chats the developer already holds and GENERATES tasks as plan steps (verify this booking form by Friday, call the buyer who asked about handover, send the instalment schedule), putting any discount or exception in the sales head's Inbox as a decision with a deadline; the humans keep the contract signature, the discount above the envelope and the handover; the receipt records every document checked and every buyer answered
The engagement on top
the developer's own ERP, e-catalogue and CRM wired to the room through connectors loaded on this pull; the notary or property lawyer as the paid professional on the match; the pack's templates for booking, instalment and handover; buyers reach the room over WhatsApp
Your own team builds it. Across the 7 adoptions here, 1 were low effort, 3 medium and 3 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs a drone operator and survey software. Fly surveys of a site to produce maps and 3D models for planning and progress. Running it needs trained pilots and permits.
Video analytics on site cameras flags missing protective gear and workers under suspended crane loads in real time to the safety team; rolled out across at least 122 sites in six markets including Indonesia
Where
multi, Southeast Asia
Segment
SMB and mid (contractors) and public (HDB)
Problem
manual inspection and watching camera feeds is prone to fatigue and error
Technology
computer vision on CCTV, real-time alerts
Who uses it
Ailytics (Singapore) with contractors in Singapore, Hong Kong, Malaysia, Indonesia, Abu Dhabi, Israel; HDB supplied training data
Outcome
up to 50 % fewer manual inspections; 32,000 hazards in 2023 (CEO's claims)
Effort
medium
Risks
Worker surveillance and privacy; false alerts; figures are the CEO's own.
How to adopt
Needs site cameras and a video analytics service plus a safety team that responds to alerts. Start with protective gear detection on one site. Running it needs alert triage and worker notice.
How a room runs it
the ROOM for a site's safety week: the site manager, the safety officer, the foremen and a safety teammate; the teammate reads the alerts and incident log the contractor's camera system already produces and GENERATES tasks as plan steps (toolbox talk on crane zones Monday, re-brief gang 3, close alert 114 with a photo), and the decision to stop work in a zone lands in the site manager's Inbox with a deadline; the humans keep the stop-work call, the discipline and the regulator report; the receipt records each alert acted on and each briefing held
The engagement on top
the contractor's camera-analytics feed and incident register wired to the room through connectors loaded on this pull; a safety professional (WSH officer) as the paid professional on the match; the pack's toolbox-talk and incident templates; foremen work over WhatsApp, which is built
AI engineers on live jobsites attending project meetings, reviewing drawings and payment applications, inside a Jobsite of the Future programme
Where
US, North America
Segment
enterprise
Problem
document-heavy project management and drawing conflicts found late
Technology
AI design review, voice scheduling, computer vision, document AI (Trunk Tools)
Who uses it
Suffolk Construction (USA, 3,500 employees)
Outcome
A Midwest pilot on a multi-billion-dollar project reportedly saves 40+ hours a month; voice scheduling cuts multi-day processes to hours; data lake of about 293 TB (Suffolk via Construction Digital)
Effort
high
Risks
Wrong drawing or code reading; subcontractor data ownership; safety decisions left to AI.
Next: Suffolk plans to expand the model across higher education, healthcare, aviation, gaming, mission-critical and mixed-use projects and standardise technologies that pass sign-off source
How to adopt
Pilot on one large project, embed AI engineers in meetings, standardise what passes sign-off, then expand by sector.
How a room runs it
the project room: the project executive, the superintendent, the subcontractors and a coordination teammate; the agent reads drawings and requisitions and GENERATES the RFI and submittal tasks per trade with due dates, and each change order or payment application becomes a proposal in the executive's Inbox with a deadline; the humans keep the safety and cost decisions; the receipt records the document, the finding and the approver
The engagement on top
connectors to the document platform and scheduling tool, a project engineer on the match, an RFI and submittal task template set
AI-native roadmap with three AI agents for tender review, contract risk and site-data insight, rolling to all new projects from 2026
Where
KR, East Asia
Segment
enterprise
Problem
heavy tender and contract paperwork and scattered site data
Technology
AI-ITB Reviewer, AI-Contract Manager, AI-Project Expert (AIPEX) built with AWS
Who uses it
Samsung C&T Engineering & Construction (South Korea), with AWS
Outcome
Three AI agents co-developed with AWS announced at the 2025 AI Day; roll-out to all new construction projects from 2026 and a 2028 target for AI-driven workflows (company statements, Nov 2025)
Effort
high
Risks
Missed contract risks; over-trust in tender summaries; proprietary data exposure.
Next: from 2026 AI-ITB Reviewer and the other agents roll out to all new construction projects, with AI-centred workflows across the business by 2028 source
How to adopt
Start with the tender-review agent on proprietary data, roll to all new projects, add contract and site-data agents.
How a room runs it
the bid-and-contract room: the bid manager, the contracts lawyer, the project director and a tender teammate; the agent reads the tender documents and GENERATES the clause-review tasks per specialist with due dates, and each risk clause becomes a proposal in the bid manager's Inbox with a deadline; the humans keep the bid decision; the receipt records the clause, the finding and the approver
The engagement on top
connectors to the tender repository and the contract system, a contracts lawyer on the match, a tender-review task template set
A general contractor ran drone flights for progress monitoring on a single aerial-intelligence platform nationwide, replacing a mix of apps
Where
US, North America
Segment
enterprise (general contractor)
Problem
several apps made flight operations confusing for new pilots
Technology
drones, aerial-intelligence software
Who uses it
Skanska USA, United States (software from Measure)
Outcome
none stated
Effort
medium
Risks
Supplier lock-in; airspace and pilot licensing; no quantified benefit published.
How to adopt
Pick one flight-and-processing platform, train a pilot per region and compare monthly captures against the schedule.
How a room runs it
the ROOM for a project: the project manager, the site supervisors and a site teammate (an agent definition with the construction skills); the teammate reads the survey, camera or inspection output the contractor already collects and GENERATES the week's tasks for each supervisor as plan steps with owners and due dates (re-fly zone B, close out the three flagged hazards, walk the north bay), and puts the schedule or safety call in the project manager's Inbox as a decision with a deadline; the humans keep the site work, the safety sign-off and the spend; the receipt records every inspection closed and every hazard cleared
The engagement on top
a custom engagement: the contractor's own drone, camera and project-management systems wired to the room through connectors loaded on this pull, the safety or surveying professional on the match, and the construction pack's task templates (inspection, hazard, progress)
Source: press · 2020 · vendor only; to corroborate
A general contractor tests camera-carrying humanoid robots that patrol a tower site on set routes, with analysis of the video by multimodal language models planned
Where
JP, East Asia
Segment
enterprise (general contractor)
Problem
inspection rounds take supervisors' time in a labour-short trade
Technology
humanoid robots, cameras, multimodal LLM video analysis
Who uses it
Shimizu Corporation, Japan (with partners including Sony Group)
Outcome
none stated
Effort
high
Risks
Early robots on a live site; no results yet; the analysis step is only planned.
How to adopt
Start with a fixed patrol route in a safe area, compare robot footage against a supervisor's round, and keep a person on every judgement.
How a room runs it
the ROOM for a project: the project manager, the site supervisors and a site teammate (an agent definition with the construction skills); the teammate reads the survey, camera or inspection output the contractor already collects and GENERATES the week's tasks for each supervisor as plan steps with owners and due dates (re-fly zone B, close out the three flagged hazards, walk the north bay), and puts the schedule or safety call in the project manager's Inbox as a decision with a deadline; the humans keep the site work, the safety sign-off and the spend; the receipt records every inspection closed and every hazard cleared
The engagement on top
a custom engagement: the contractor's own drone, camera and project-management systems wired to the room through connectors loaded on this pull, the safety or surveying professional on the match, and the construction pack's task templates (inspection, hazard, progress)
A specialist foundations contractor trialled an AI video system on two major projects that flags missing PPE, entries into restricted zones and plant-people conflicts to managers on a dashboard
Where
GB, Europe
Segment
enterprise (specialist contractor)
Problem
manual observation misses breaches
Technology
AI video analysis, tracking and camera footage
Who uses it
Bachy Soletanche, United Kingdom (with AITIS and Iknaia)
Outcome
none stated
Effort
high
Risks
Worker privacy and union concerns; false alerts; no published detection rates.
How to adopt
Trial on one or two sites with cameras already fitted, agree what counts as a breach, and check alerts against supervisors' own walks before extending.
How a room runs it
the ROOM for a project: the project manager, the site supervisors and a site teammate (an agent definition with the construction skills); the teammate reads the survey, camera or inspection output the contractor already collects and GENERATES the week's tasks for each supervisor as plan steps with owners and due dates (re-fly zone B, close out the three flagged hazards, walk the north bay), and puts the schedule or safety call in the project manager's Inbox as a decision with a deadline; the humans keep the site work, the safety sign-off and the spend; the receipt records every inspection closed and every hazard cleared
The engagement on top
a custom engagement: the contractor's own drone, camera and project-management systems wired to the room through connectors loaded on this pull, the safety or surveying professional on the match, and the construction pack's task templates (inspection, hazard, progress)
Source: vendor · 2023 · vendor only; to corroborate
Drone aerial surveys and LiDAR mapping of project sites produce topographic maps and 3D models used to plan and measure work progress
Where
ID, Southeast Asia
Segment
enterprise (state-owned contractor)
Problem
manual surveying is slow and imprecise on large or vegetated sites
Technology
survey drones, aerial photo mapping, LiDAR, 3D site models
Who uses it
Wijaya Karya (WIKA), Indonesian state contractor, with Terra Drone Indonesia
Outcome
one survey covered about 347 hectares and produced maps with accuracy under 10 cm (WIKA survey at Sungai Duri, as reported by Kontan from the drone supplier's account)
Effort
medium
Risks
Aviation permits; data processing cost; dated (2019); no AI.
How to adopt
Needs a drone operator and survey software. Fly surveys of a site to produce maps and 3D models for planning and progress. Running it needs trained pilots and permits.
How a room runs it
the ROOM for a project: the project manager, the site engineer, the QS and a site-progress teammate; the teammate reads the survey outputs and the schedule the contractor already has and GENERATES the fortnight's tasks as plan steps (re-survey block C by the 20th, reconcile quantities with the QS, raise the variance note), and the claim-or-variation decision lands in the project manager's Inbox with a deadline; the humans keep the survey flight, the measurement sign-off and the client certificate; the receipt records each survey received and each variance resolved
The engagement on top
the contractor's schedule and cost system and the drone vendor's map folder wired to the room through connectors loaded on this pull; a licensed surveyor or quantity surveyor as the paid professional on the match; the construction pack's progress and variation templates
Your own team builds it. Across the 7 adoptions here, 0 were low effort, 3 medium and 4 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs sensors, CCTV and traffic data on a common platform. Build a command centre dashboard with flood and landslide alerts. Running it needs operators and sensor maintenance.
In-house CRE language model released to the whole workforce, with 11,000 users in the first 48 hours
Where
US, global
Segment
enterprise
Problem
slow lease review and space analysis across 80+ countries
Technology
JLL GPT (in-house model fine-tuned on CRE data)
Who uses it
JLL (USA), JLL Technologies
Outcome
Released to 103,000+ employees; more than 11,000 used it in the first 48 hours; a lease review for a pet clause went from hours or days to seconds (JLL figures via CIO Dive, Aug 2023)
Effort
high
Risks
Wrong lease readings; client data confidentiality; build cost.
Next: JLL plans data mining for automated facility management, price modelling for investors and matchmaking for leasing transactions source
How to adopt
Build or fine-tune on internal CRE data, release widely after a successful test, collect feedback, add use cases.
How a room runs it
the portfolio room: the asset manager, the lease analyst, the client's facilities owner and an analysis teammate; the agent reads the lease and drafts the summary and GENERATES the review tasks per analyst with due dates, and each lease exception becomes a proposal in the manager's Inbox with a deadline; the humans keep the investment and leasing decisions; the receipt records the clause, the source and the reviewer
The engagement on top
connectors to the lease repository and the space-utilisation data, an asset manager on the match, a lease-review task template set
An industrial-estate operator opened a Smart City Command Center that visualises CCTV, traffic-management and environmental-sensor data, with landslide and flood alerts, for estate management and its tenants
Where
ID, Southeast Asia
Segment
enterprise (estate manager)
Problem
estate staff react late to floods, landslides, traffic and security events across a large estate
PT Samakta Mitra (Sinar Mas Land IT subsidiary) with NEC Indonesia, Deltamas / GIIC estate, 205+ tenant companies
Outcome
none stated
Effort
high
Risks
Supplier conflicts of interest; surveillance and privacy; no results published.
How to adopt
Needs sensors, CCTV and traffic data on a common platform. Build a command centre dashboard with flood and landslide alerts. Running it needs operators and sensor maintenance.
How a room runs it
the ROOM for an estate's operations: the estate manager, the security and facilities supervisors, a tenant-relations officer and an estate teammate; the teammate reads the alerts the command centre already raises and GENERATES the tasks as plan steps with owners and due dates (inspect drain 4 before the rain, notify tenants in zone B, close the gate log), and a decision (close the road, call the contractor, spend above the envelope) lands in the estate manager's Inbox with a deadline; the humans keep the field response and the spend above the envelope; the receipt records each alert, who acted and when, which tenants were told
The engagement on top
the estate's command-centre and CCTV feeds, work-order system and tenant list wired to the room through connectors loaded on this pull; a facilities or safety professional as the paid professional on the match; the facility pack's incident and tenant-notice templates; supervisors and tenants are reached on WhatsApp, which is built
Source: vendor · 2026 · vendor only; to corroborate
An AI-native rental-operations platform runs specialised agents for portfolio data, rent collection and lease renewals (leases, insurance, affordability checks, payments), with property managers keeping approvals
Where
AE, Middle East
Segment
SMB and mid (property managers)
Problem
property managers cannot grow admin teams as fast as portfolios; renewals and collections are repetitive
Technology
AI agents, open-finance affordability checks, pay-by-bank orchestration
Who uses it
Rentify (UAE) and its property-manager customers; portfolios over AED 22 billion
Outcome
the platform supports portfolios of more than AED 22 billion (about USD 6 billion) in real estate assets (the company's own scale figure, not a measured change; not attributed on the page)
Needs property, lease and payment data in one place. Start with rent collection and renewal reminders with manager approval. Running it needs integration to payment systems.
How a room runs it
the ROOM for a rental portfolio: the property manager, the landlord, the leasing and collections staff and a renewals teammate; the teammate reads the leases, rent ledger and arrears the manager already keeps and GENERATES the tasks as plan steps (send renewal offers for the 14 leases ending in 60 days, chase the three overdue units, book the inspection), and a rent change or a legal notice lands in the landlord's or manager's Inbox as a decision with a deadline; the humans keep the approvals, the tenant conversation and the notice; the receipt records each renewal signed and each payment received
The engagement on top
the manager's rental ledger, lease store and bank feed wired to the room through connectors loaded on this pull; a property lawyer or licensed manager as the paid professional on the match; the rental pack's renewal and arrears templates; tenants are reached on WhatsApp
Source: press · 2026 · vendor only; to corroborate
Brokerage AI assistant for 83,000 agents that drafts outreach, lists prospects and bundles about 100 tools behind one prompt
Where
US, North America
Segment
enterprise
Problem
agent admin work and lost sales opportunities
Technology
Compass AI Assistant
Who uses it
Compass (USA)
Outcome
Launched to 83,000 agents in July 2026; within weeks about 15,000 agents had 97,000+ conversations (Compass figures via Fortune)
Effort
medium
Risks
Low adoption by agents; fair-housing and disclosure errors; AI-altered listing images.
Next: the president of growth expects usage to rise as agents adopt the platform, with executive visits and training driving adoption source
How to adopt
Launch with a live demo, train agents, track conversations per agent, expand tool connections.
How a room runs it
the listing room: the agent-of-record, the buyer or seller, the transaction coordinator and a marketing teammate; the assistant drafts outreach and property lists and GENERATES the follow-up tasks per party with due dates, and each price or disclosure change becomes a proposal in the agent's Inbox with a deadline; the humans keep the client relationship; the receipt records the draft, the data used and the approver
The engagement on top
connectors to the CRM, the listing service and e-signature, a transaction coordinator on the match, a listing-to-close task template set
A kos platform sells owners a smart-key that locks a room by phone and can lock it when payment is overdue past a tolerance, plus a Smart-Kost system for room management, contracts, resident security and caretaker task monitoring
Where
ID, Southeast Asia
Segment
SMB (kos owners)
Problem
owners lose keys, chase late rent and cannot see what caretakers do
Technology
digital door locks linked to bookings and billing, property-management software
Who uses it
Mamikos (Indonesia) and its kos-owner partners; SinggahSini works with 1,000+ owners managing 14,000 rooms
Outcome
none stated
Effort
medium
Risks
Locking out residents over payment disputes; tenancy law; no AI claimed.
How to adopt
Needs smart locks, resident contracts and billing data. Fit locks in a few rooms and tie the lock to payment status with a tolerance period. Running it needs maintenance and resident support.
How a room runs it
the ROOM for a kos owner: the owner, the caretaker, a bookkeeper and a rental teammate; the teammate reads the bookings, rent ledger and lock log the platform already holds and GENERATES the caretaker's tasks as plan steps (clean room 7 before the new tenant Friday, collect the key, repair the tap), the rent reminders to tenants over WhatsApp, and a decision in the owner's Inbox (extend a late tenant, apply the fine) with a deadline; the humans keep the key hand-over, the tenant relationship and the decision to evict; the receipt records each reminder, each task done and each payment received
The engagement on top
the platform's booking and billing data and the owner's bank statement wired to the room through connectors loaded on this pull; a property manager or notary as the paid professional on the match; the rental pack's check-in, check-out and reminder templates
A property-management software supplier added an embedded generative-AI assistant, a message inbox that sorts resident messages and suggests replies, and a workflow engine for repeatable processes
Where
US, North America
Segment
SMB and mid (property managers)
Problem
manual reports, replies and recurring workflows
Technology
generative AI assistant, message triage, workflow automation inside the management system
Who uses it
AppFolio, United States
Outcome
users save over 10 hours a week on their task lists, on average; each Realm-X message took 26 seconds less than one written without it (AppFolio's own figures, measurement method not stated) the company’s own figure
Effort
medium
Risks
Self-reported savings; wrong replies to residents; plan limits on workflow features.
How to adopt
Choose a management system that embeds the assistant, start with message replies and reports, and have staff approve every outgoing reply at first.
How a room runs it
the ROOM for a portfolio: the property-management principal, the leasing and maintenance staff and a portfolio teammate (an agent definition with the property-management skills); the teammate reads the leases, messages and work orders the firm already keeps and GENERATES the week's tasks for each staff member as plan steps with owners and due dates (reply to the overdue-rent tenants, schedule the plumber, send the renewal offers), and puts the rent-change or vendor-spend proposal in the principal's Inbox as a decision with a deadline; the humans keep the tenant relationship, the legal notice and the approval; the receipt records each message sent and each work order closed
The engagement on top
a custom engagement: the firm's own property-management system wired to the room through connectors loaded on this pull, the licensed property manager on the match, and the property pack's task templates (rent, renewal, repair)
Source: vendor · 2024 · vendor only; to corroborate
A housing group backs a start-up building an AI platform that automates landlords' administrative processes and organises their documents
Where
DE, Europe
Segment
SMB (landlords and small managers)
Problem
manual routine work and many tools and channels
Technology
AI property-management platform, document handling
New company with no published results; tenant data protection under GDPR; automated notices need human sign-off.
How to adopt
Pilot with a small landlord portfolio, load its leases and documents, and compare admin hours before and after.
How a room runs it
the ROOM for a portfolio: the property-management principal, the leasing and maintenance staff and a portfolio teammate (an agent definition with the property-management skills); the teammate reads the leases, messages and work orders the firm already keeps and GENERATES the week's tasks for each staff member as plan steps with owners and due dates (reply to the overdue-rent tenants, schedule the plumber, send the renewal offers), and puts the rent-change or vendor-spend proposal in the principal's Inbox as a decision with a deadline; the humans keep the tenant relationship, the legal notice and the approval; the receipt records each message sent and each work order closed
The engagement on top
a custom engagement: the firm's own property-management system wired to the room through connectors loaded on this pull, the licensed property manager on the match, and the property pack's task templates (rent, renewal, repair)
Source: vendor · 2026 · vendor only; to corroborate
A township developer states it is optimising AI for traffic management, smart security and smart-building systems in BSD City, with IoT, a digital twin, drainage monitoring and command control for traffic and security
Where
ID, Southeast Asia
Segment
enterprise
Problem
running a township: traffic, security, drainage and buildings handled by separate teams
Technology
IoT, digital twin, AI, data analytics, Microsoft cloud
Who uses it
Sinar Mas Land, BSD City (Indonesia): CEO Technology Advisory Irvan Yasni
Outcome
none stated
Effort
high
Risks
Statements of direction only; cost overrun; privacy of residents.
How to adopt
Needs IoT infrastructure, drainage and traffic data and a control centre. Start with one function such as traffic management. Running it needs operators and data integration.
How a room runs it
the ROOM for a building or township management office: the property manager, the building engineer, the security chief, the tenant-service staff and a building teammate; the teammate reads the building systems' alerts and the tenant requests already logged and GENERATES the week's work orders as plan steps (service chiller 2 by Thursday, check reservoir drain, answer the three open tenant requests), and a repair above budget lands in the property manager's Inbox as a decision with a deadline; the humans keep the repair work, the vendor contract and the owner's approval; the receipt records each work order closed and each tenant answered
The engagement on top
the building management system, helpdesk and vendor contracts wired to the room through connectors loaded on this pull; a facility engineer as the paid professional on the match; the pack's work-order and tenant-notice templates; tenants reach the room over WhatsApp
Your own team builds it. Across the 8 adoptions here, 0 were low effort, 5 medium and 3 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs a bank partner for QR codes and a donation account per mosque. Start with QR codes on donation boxes in a few mosques. Running it needs treasurers trained and monthly reporting.
Saudi Arabia's Ministry of Hajj and Umrah runs the Nusuk app (100+ services in 20 languages) and says AI at the Grand Mosque has improved crowd management and emergency response; Nusuk issued over 13 million Rawdah permits in 2024, up from 4 million in 2022
Where
SA, Middle East
Segment
public (national ministry for pilgrimage)
Problem
millions of pilgrims a year to move safely through the holy sites
Technology
mobile app, permit platform, AI video monitoring, smart cards
Who uses it
Saudi Ministry of Hajj and Umrah
Outcome
Over 13 million Rawdah permits issued through Nusuk in 2024, up from 4 million in 2022; 18.5 million pilgrims and Umrah performers in 2024 (the minister's figures, as reported)
Effort
medium
Risks
Figures are the ministry's own; AI claims are not detailed; surveillance and privacy at the holy sites.
How to adopt
Needs a booking and permit process that can be done online and a multilingual help channel; first step is the permit and itinerary flow for one group; takes integration with the authority's platform.
How a room runs it
the ROOM for a pilgrimage operator's office: the group leader, the guides and a pilgrim-services teammate; the teammate GENERATES the groups' permit, booking and itinerary tasks for the guides as plan steps with owners and due dates; a decision block in the group leader's Inbox for each change of schedule, with a deadline; what the humans keep: the care of pilgrims, the on-site calls and the sign-off; the receipt records every finished task and every decision with who took it
The engagement on top
a custom engagement: the operator's own bookings and the ministry's permit platform stay theirs, wired to the room through connectors loaded on this pull, the pilgrimage-logistics adviser joins as the paid professional on the match
AI adviser copilot, Caddy, answering from approved sources with a supervisor check, built in-house and rolling out nationally
Where
GB, Europe
Segment
public (charity network)
Problem
advice queues and long calls
Technology
Caddy (Citizens Advice Manchester team)
Who uses it
Citizens Advice (Stockport, Oldham, Rochdale, Trafford), United Kingdom
Outcome
In a months-long trial across four local teams the average call length fell from 8 minutes to 4 (Newsio, Mar 2025)
Effort
medium
Risks
Wrong advice about benefits or debt; client confidentiality; reliance on one source library.
Next: Caddy is expected to roll out across Citizens Advice nationwide, with the UK technology secretary pointing to HMRC as another use source
How to adopt
Pilot in a few offices, restrict to approved sources, keep supervisor checks, publish the code.
How a room runs it
the advice room: the adviser, the supervisor and the client's case; the agent finds the answer in the approved sources and drafts the reply, and a supervisor approves each response as a proposal; the receipt records the question, the source and the checker; the human adviser keeps the advice
The engagement on top
connectors to the case system and the approved guidance library, a supervisor-review rule, an adviser training pack
AI that transcribes handwritten and typed historical record images into searchable text, so records are findable before volunteers index them
Where
US, North America
Segment
large (faith-based genealogy non-profit)
Problem
billions of record images that nobody can search
Technology
Full-Text Search AI (FamilySearch)
Who uses it
FamilySearch, United States (a non-profit run by The Church of Jesus Christ of Latter-day Saints)
Outcome
Over 600,000 historical genealogy books being digitised and many more collections added to AI full-text search in 2025 (FamilySearch, Jan 2025) the company’s own figure
Effort
medium
Risks
Transcription errors in names; privacy of living persons; volunteer trust.
Next: full-text search is being added to many more collections in 2025, in more handwriting languages, plus oral genealogies from 25 African countries source
How to adopt
Run the model on one collection, sample transcription accuracy, add volunteer correction.
How a room runs it
the research room: the volunteer indexer, the family researcher and an archive teammate; the agent transcribes and proposes name matches, and volunteers correct them as proposals in an Inbox queue; the receipt records the image, the reading and the corrector
The engagement on top
connectors to the image archive and the indexing app, a correction-queue template, a quality-sampling rule
Tirumala Tirupati Devasthanams runs an AI-based integrated command and control centre over more than 6,000 cameras (about 300 with face recognition) that turns video into heat maps, people counts and alerts for the darshan queues; free-darshan waits are reported to have fallen to 11 to 13 hours
Where
IN, South Asia
Segment
public (large temple trust)
Problem
crowds of 70,000 to 100,000 pilgrims a day and waits of up to 26 hours
Technology
video analytics, location-intelligence platform, GPUs, facial recognition
Who uses it
Tirumala Tirupati Devasthanams (TTD); Kloudspot; donors; Andhra Pradesh government
Outcome
Free-darshan waits down to 11 to 13 hours and about 9 hours on average during the Vaikunta Ekadasi festival (as reported by The Week, from TTD statements; attributed in part to cancelling some scheduled darshans)
Effort
high
Risks
Facial recognition and privacy; unverified figures; reliance on cameras in crowds; the cost of the hardware.
How to adopt
Needs cameras with network coverage, a floor plan of the queues and an operations room; first step is people counting on the busiest queue; takes hardware, staff protocols and donor or budget funding.
How a room runs it
the ROOM for a temple's operations team: the executive officer, the queue supervisors and a crowd-flow teammate; the teammate GENERATES the shift's tasks for the supervisors (open a compartment, slow an entry, send staff to a hall) as plan steps with owners and due times; a decision block in the officer's Inbox when a queue exceeds its target wait, with a deadline and the camera counts attached; what the humans keep: the pilgrim-facing calls, the safety decisions and the sign-off; the receipt records every finished task and every decision with who took it
The engagement on top
a custom engagement: the trust's own cameras, booking system and festival calendar stay theirs, wired to the room through connectors loaded on this pull, the crowd-safety adviser joins as the paid professional on the match
AI chatbot on the parish union website, answering in several languages from a staff-built knowledge base
Where
FI, Europe
Segment
public (church body)
Problem
members cannot find service times, events or help outside office hours
Technology
Leadoo chatbot with a large language model
Who uses it
Helsinki Parish Union (19 parishes, about 305,000 members), Finland
Outcome
Nearly 44,000 interactions and about 8,500 full conversations in the first year; monthly interactions now above 9,100 (WCC report of the union's keynote, May 2026)
Next: the team is raising answer quality and satisfaction ratings, with monthly interactions already above 9,100 source
How to adopt
Staff curate the knowledge base, add a person-handover for sensitive topics, track satisfaction.
How a room runs it
the parish room: the secretary, the pastor and the volunteer coordinator; the agent answers from the staff-built knowledge base and generates the follow-up tasks (a callback, a funeral enquiry) with dates, and any pastoral or sensitive request becomes a proposal for a pastor; the receipt records the question, the answer and the owner
The engagement on top
connectors to the website and the events calendar, a multilingual knowledge-base pack, a sensitive-topic handover rule
Retrieval-based Catholic answer engine limited to church documents, with pilots to digitise pontifical libraries
Where
US, North America
Segment
SMB (faith publisher and start-up)
Problem
finding authentic church teaching quickly
Technology
Magisterium AI (Longbeard)
Who uses it
Longbeard, United States and Rome; pilots with the Pontifical Oriental Institute and the Salesian Pontifical University
Outcome
Over 180,000 people in 165 countries and a corpus that grew from about 600 to more than 8,000 documents since the July 2023 launch (founder figures, Apr 2024)
Effort
low
Risks
Perceived spiritual authority; misquotation; user numbers are the company's own.
Next: Longbeard is working with two pontifical institutions to digitise library holdings for the engine source
How to adopt
Fix the corpus first, force citations, refuse anything outside scope, review with clergy.
How a room runs it
the study room: the priest, the catechist and the parish volunteer; the agent answers only from the approved corpus with citations, and pastoral advice stays out of scope; a doctrinal answer used in a homily or a class becomes a proposal the priest reviews; the receipt records the question and the cited text
The engagement on top
connectors to the parish document library, a citation-first answer template, a scope guard that refuses personal spiritual direction
AI agents for a church-management software company: event reminders, volunteer scheduling with load balancing, round-the-clock member FAQs, giving prompts; claimed after 60 days volunteer attendance 61 to 89 % and admin time 15 to 5 hours a week
Where
US, North America
Segment
SMB and mid (churches)
Problem
volunteer no-shows, manual admin, low digital giving
Technology
AI agents for scheduling, support and follow-up
Who uses it
an unnamed Dallas church-management software client of Auralis AI
Outcome
volunteer attendance 61 to 89 % after 60 days; admin time 15 to 5 hours a week (vendor's case study claim) the company’s own figure
Effort
medium
Risks
Vendor claims; member data privacy; pastoral relations replaced by automation.
How to adopt
Needs a church-management system with member and volunteer data. Start with event reminders and volunteer scheduling. Running it needs a staff owner.
How a room runs it
the ROOM for a congregation office: the pastor or priest, the admin and a volunteers teammate; the agent GENERATES the weekly roster as tasks for volunteers with due dates and the reminder messages for the admin to release, and files any change of giving campaign as a decision for the board with a deadline; the clergy keep pastoral care and every message in their name, the agent sends nothing pastoral; the receipt records who served and who approved
The engagement on top
connectors loaded on this pull to the congregation's member database and calendar; a church administrator as paid professional; the pack carries roster templates; channel: WhatsApp
Source: vendor · 2026 · vendor only; to corroborate
AI chatbot trained only on the organisation's official resources to help families and leaders and to find a local unit
Where
US, North America
Segment
large (national youth organisation)
Problem
parents and volunteers cannot find answers or a unit
Technology
Scoutly (Scouting America)
Who uses it
Scouting America, United States
Outcome
Launched on Scouting.org and BeAScout.org, available 24 hours; no usage figures published (Aug 2025) the company’s own figure
Effort
low
Risks
Child-safety and privacy rules; wrong policy answers; no published results.
Next: Scoutly is the primary source for the new Artificial Intelligence merit badge due in mid-September 2025 source
How to adopt
Limit the model to official content, add find-a-unit and registration help, review logs.
How a room runs it
the membership room: the parent, the unit leader and the council volunteer; the agent answers from official resources and generates the follow-up tasks (a registration step, a leader contact) with dates, and anything that touches a child's record becomes a proposal for the council; the receipt records the answer and the source
The engagement on top
connectors to the membership system and the official handbook library, a registration-help template, a youth-data privacy rule
The Religious Ministry integrates its zakat management information system with national and provincial zakat institutions (LAZ); 80 LAZ (45 national, 35 provincial) were briefed in February 2024
Where
ID, Southeast Asia
Segment
public and community (zakat bodies)
Problem
zakat institutions on separate systems; reporting and oversight
Technology
zakat management information system, integration with LAZ systems
Who uses it
Ministry of Religious Affairs; 80 LAZ
Outcome
none stated
Effort
high
Risks
Uneven capacity of bodies; data quality; no AI.
How to adopt
Needs a ministry system and agreed data format for zakat bodies. Integrate national then provincial bodies. Running it needs training and reporting rules.
How a room runs it
the ROOM for a zakat body: the director, the distribution manager and a zakat-desk teammate; the agent reads the collection and beneficiary lists the body keeps and GENERATES the monthly report to the ministry as a draft for the manager and the verification visits for field officers with due dates; each allocation of funds is a proposal in the sharia supervisor's and director's Inbox with a deadline and the human decision recorded; the receipt shows donor to beneficiary
The engagement on top
connectors loaded on this pull to the body's donor and beneficiary system and the ministry's reporting; a zakat auditor or sharia adviser as paid professional; the pack carries verification task templates; WhatsApp for donors
BSI and the Indonesian Mosque Council (DMI) pilot mosque digitalisation in 37 Aceh mosques: QRIS codes on donation boxes send infak and zakat straight to the mosque's account, a mosque finance and activity app, and crowdfunding for waqf projects; about 300,000 of nearly 800,000 mosques are registered in the Religious Ministry's SIMAS
QRIS, mobile banking, mosque management app, crowdfunding
Who uses it
Bank Syariah Indonesia, DMI, mosque committees; Ministry of Religious Affairs (SIMAS)
Outcome
DMI secretary-general says mosques using QRIS can raise income up to 10 times (an official's estimate, no data)
Effort
medium
Risks
Fraud with fake QR codes; digital divide; claim of 10x is unmeasured. No AI.
How to adopt
Needs a bank partner for QR codes and a donation account per mosque. Start with QR codes on donation boxes in a few mosques. Running it needs treasurers trained and monthly reporting.
How a room runs it
the ROOM for a mosque: the chair of the takmir, the treasurer, the imam and a mosque-admin teammate; the agent reads the donation records the mosque already has and GENERATES the monthly report for the congregation as a draft for the treasurer, the Friday roster tasks for volunteers with due dates and a proposal for the chair when a spend exceeds the committee's envelope; the committee keeps every fund decision and the imam all religious guidance, the agent gives none; the receipt is the open book the congregation can read
The engagement on top
connectors loaded on this pull to the mosque's bank statement export and SIMAS records; a community accountant as paid professional; the pack carries roster and report templates; channel: WhatsApp group of the congregation
BAZNAS launched Zaki, a chatbot on LINE that calculates a donor's zakat, explains programmes and answers questions with natural-language processing; built with Artdigi
Where
ID, Southeast Asia
Segment
public (zakat body)
Problem
donors unsure how much zakat to pay; low access to information
Technology
NLP chatbot on LINE
Who uses it
BAZNAS (national zakat agency), PT Artdigi
Outcome
none stated
Effort
medium
Risks
Advertorial source; unknown whether it still runs; wrong religious calculations.
How to adopt
Needs a messaging channel and a list of programme content and zakat rules. Build a chatbot that calculates zakat and answers questions. Running it needs content updates.
How a room runs it
the donor-facing door of a zakat room: the donor asks on WhatsApp, the teammate calculates from the donor's own figures and hands over to a human amil for anything about rulings; the amil keeps all religious and allocation decisions; the receipt shows what the donor was told and what was paid
The engagement on top
channel: WhatsApp (built); a sharia adviser as paid professional; the pack carries the zakat-calculation task template
Pilot AI virtual engagement officer that contacts donors by email and text, positioned as support for human relationships
Where
US, North America
Segment
public (diocese)
Problem
thin staff time for donor follow-up
Technology
Maria (Givzey)
Who uses it
Diocese of Bridgeport, United States
Outcome
Pilot launched 25 March 2026 with a small donor group; no results published yet the company’s own figure
Effort
low
Risks
Donor consent and privacy; tone with grieving or giving donors; unproven results.
Next: the diocese says it will explore other uses of AI to support personal connection source
How to adopt
Start with a small consenting donor group, approve every message, review replies weekly.
How a room runs it
the stewardship room: the development officer, the pastor and the donor's record; the agent drafts personal follow-ups and generates the call tasks with dates, and every outbound message to a donor becomes a proposal the officer approves until the rules are proven; the receipt records the message, the opt-in and the approver
The engagement on top
connectors to the donor CRM and the messaging channel, a consent and opt-out rule, a message-approval template
Your own team builds it. Across the 12 adoptions here, 4 were low effort, 6 medium and 2 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Needs stores with network connectivity. Roll out SD-WAN, then sensors for security, cooling and shelf stock. Vision AI for queues is named as a future step. Running it needs network operations.
Personal shopping assistant for customers plus Sage, an AI assistant for associates (schedules, time off, pay, policies), on Google Cloud
Where
US, North America
Segment
enterprise
Problem
shoppers planning meals and savings, and store staff juggling scheduling admin
Technology
Gemini Enterprise for Customer Experience (Google Cloud); Sage
Who uses it
Kroger, United States
Outcome
Sage grew from 30 use cases five months after the early-2024 start to more than 95 use cases, used by up to 150,000 associates (Grocery Dive, Jan 2026)
Effort
high
Risks
Wrong pay or policy answers; labour-relations; customer data.
Next: Kroger will use Customer Experience Agent Studio to analyse store calls and keeps improving Sage source
How to adopt
Start with associate-facing tasks; grow use cases in steps; add the shopper assistant after.
How a room runs it
the store room: the store leader, the associates, the pharmacist and a scheduling teammate; the agent drafts the shift plan and GENERATES the request tasks per associate with dates, and each schedule change or time-off approval becomes a proposal in the leader's Inbox; the receipt records the request, the rule applied and who approved
The engagement on top
connectors to the workforce-management and loyalty systems, a store-operations teammate, a policy knowledge set
Marketplace AI for sellers: a store chat assistant, smart listing generation, translation into 15 languages and AI-prompted buyer reviews across Indonesia and five other markets
Where
multi, Southeast Asia
Segment
SMB and mid (online sellers)
Problem
sellers spend time on chat, listings and reviews
Technology
generative AI chatbot (LISA), listing and translation generation
Who uses it
Lazada (Alibaba group), sellers in Indonesia, Malaysia, Philippines, Singapore, Thailand, Vietnam
Outcome
none stated (the source's 38 % and 40 % figures could not be re-read: page returned 403)
Needs a seller account on the marketplace. Switch on the chat assistant and listing generator. Running it needs review of generated content.
How a room runs it
the ROOM for an online shop: the owner, one packer and a seller teammate (an agent definition with the marketplace's vocabulary); the agent reads the orders and chats the shop already has and GENERATES the day's tasks (pack these 14 orders by 15:00, restock the top seller, answer the three unresolved buyer questions) as plan steps with owners and due times, and a DECISION for the owner (discount, reorder, pause an ad) with a deadline; the owner keeps the price, the supplier call and the spend above the envelope; the receipt records each order handled
The engagement on top
a custom engagement: the shop's marketplace store, stock sheet and WhatsApp Business wired to the room through connectors loaded on this pull, an e-commerce operator as the paid professional on the match, the retail pack's templates
Source: vendor · 2025 · vendor only; to corroborate
Staff assistant on a large language model rolled out to 125,000 employees in eight countries to ease routine work and give frontline teams back operational time
Where
FR, Europe
Segment
enterprise
Problem
routine written and information tasks take staff time
Technology
generative AI chatbot (Google Gemini)
Who uses it
Carrefour, France with rollout in eight countries (Spain, Poland and Romania named)
Outcome
rollout complete in eight countries to 125,000 employees; no quantified result (Carrefour's group CTO, via Retail Systems)
Effort
medium
Risks
Data leakage; unmeasured benefit; staff over-trust in outputs.
How to adopt
Needs a staff chat tool, usage rules and training. Start with one country and a small set of tasks, then widen. Running it needs a policy for what data staff may paste in.
How a room runs it
the ROOM for a store group: the store manager, the department leads and a store-operations teammate (an agent definition with the chain's procedures and vocabulary); the agent reads the tasks the team already tracks and GENERATES the leads' tasks (rewrite the promotion signs, answer the supplier query, prepare the weekly rota draft) as plan steps with owners and due times, and a DECISION for the store manager (staffing, markdown, supplier escalation) with a deadline, landing in the manager's Inbox as a proposal; the humans keep the customer-facing and staffing decisions; the receipt records each draft and who approved it
The engagement on top
a custom engagement: the group's intranet, rota and stock systems wired to the room through connectors loaded on this pull, a retail-operations professional on the match, the retail pack's templates
Chatbot answering common customer queries while call-centre staff are retrained as remote interior design advisers, selling design by phone and video
Where
NL, Europe
Segment
enterprise
Problem
routine queries crowding out higher-value advice
Technology
Billie (IKEA's AI assistant)
Who uses it
Ingka Group (largest IKEA franchisee), Netherlands
Outcome
Billie handled 47 percent of customer queries since its 2021 launch; about 8,500 call-centre workers trained as interior design advisers; remote design sales 1.3 billion EUR in 2022, 3.3 percent of revenue (Ingka via Reuters, Jun 2023)
Effort
medium
Risks
Wrong answers on delivery or returns; workforce anxiety; figures are 2023.
Next: Ingka aims to raise remote design sales to 10 percent of revenue by 2028 source
How to adopt
Automate the routine queries first; retrain staff for advice; grow the paid advice channel.
How a room runs it
the design room: the adviser, the customer, the store team and a planning teammate; the agent answers the routine queries and drafts the room plan and GENERATES the follow-up tasks per customer with dates, and each quote becomes a proposal in the adviser's Inbox; the receipt records the plan, the quote and who approved
The engagement on top
connectors to the order and catalogue systems, a design-adviser teammate, a retraining track for service staff
In-store and in-app shopping assistant with store-specific knowledge (aisle maps, local stock, climate-appropriate advice), taking text, voice-to-text, images and several languages
Where
US, North America
Segment
enterprise
Problem
home-improvement shoppers unsure which product fits their project and where it sits in a large store
Technology
Magic Apron, The Home Depot's AI-powered shopping assistant, in Store Mode of the mobile app and via in-store QR codes
Who uses it
The Home Depot, United States
Outcome
No outcome measured is published. The release states Magic Apron is live in all 2,000+ U.S. stores and fields millions of questions per month; pilot customers reported finding products faster and more purchase confidence (The Home Depot, 27 Aug 2026) the company’s own figure
Effort
high
Risks
Wrong product-fit or stock answers; customer data; reliance on store data being current.
Next: the release names no further Magic Apron step beyond the store-localized launch source
How to adopt
Start with online and in-app product Q&A; add store-level stock and maps; extend to in-store access by QR code.
How a room runs it
the store room: the shopper, the store associate and a product-knowledge teammate; the agent answers from the store's own catalogue and stock and, when it cannot, files an assist request to an associate instead of guessing; the receipt records the question, the sources used and the store it was answered for
The engagement on top
connectors to the product catalogue and store inventory, a store-knowledge set per location, an associate hand-off teammate
Store-network modernisation: SD-WAN for all stores from one dashboard; IoT for security, cooling-unit temperature and shelf-stock sensors; Vision AI for customer behaviour and queues named as next
Where
ID, Southeast Asia
Segment
enterprise
Problem
unstable store connectivity; cold-chain and stock losses; high operating cost
Technology
SD-WAN, IoT connectivity (Indosat Business), IoT sensors, Vision AI (planned)
Who uses it
Alfamart (thousands of stores), Indonesia
Outcome
none stated
Effort
high
Risks
Cost across many stores; vision AI not yet deployed; no results published.
How to adopt
Needs stores with network connectivity. Roll out SD-WAN, then sensors for security, cooling and shelf stock. Vision AI for queues is named as a future step. Running it needs network operations.
How a room runs it
the ROOM for a store cluster: the area manager, the store heads and a retail-operations teammate (an agent definition with the replenishment and cold-chain vocabulary); the agent reads the alerts the company's sensors and store systems already raise and GENERATES the store heads' tasks (check freezer 3 in store 0412 before 15:00, restock shelf B) as plan steps with owners and due times, and a DECISION for the area manager when a unit keeps failing (call the technician, move stock) with a deadline; the humans keep the floor work and the spend above the envelope; the receipt records each alert, its fix and who closed it
The engagement on top
a custom engagement: the chain's POS, stock system and sensor feeds wired to the room through connectors loaded on this pull, a retail-operations professional on the match, the retail pack's templates; store staff over WhatsApp
Generative AI shopping assistant in the retailer's app: summarises reviews, answers product questions, recommends and compares items and helps plan purchases; reordering and service booking announced
Where
US, North America
Segment
enterprise
Problem
shoppers struggle to compare products and read many reviews
Technology
retail-specific large language models, agents for comparison and personalisation
Who uses it
Walmart, United States
Outcome
none stated (the assistant has no usage figure; Walmart's own survey: 27 % of respondents prefer AI recommendations)
Effort
medium
Risks
Wrong or invented product claims; low shopper trust in agents for a whole trip; vendor-reported.
How to adopt
Needs a product catalogue with reviews and a channel (app or site) to host the assistant. Start with review summaries and product questions for one category, with staff checking answers. Running it needs review of wrong answers.
How a room runs it
the ROOM for a shop: the owner, one assistant and a shopper-help teammate (an agent definition with the catalogue's vocabulary); the agent reads the product questions and reviews the shop already receives and GENERATES the day's tasks (answer these five open questions, fix the listing that draws returns, reorder the top seller) as plan steps with owners and due times, and a DECISION for the owner (discount, reorder, drop a line) with a deadline, landing in the owner's Inbox as a proposal; the owner keeps the price, the supplier call and any spend above the envelope; the receipt records each answer given and who approved it
The engagement on top
a custom engagement: the shop's store platform, catalogue and customer chat wired to the room through connectors loaded on this pull, an e-commerce operator as the paid professional on the match, the retail pack's templates
AI chat commerce on WhatsApp: automated order handling in the messaging app, multi-language and dialect replies, cross-sell analytics
Where
ID, Southeast Asia
Segment
SMB
Problem
small sellers cannot staff chat all day; orders are processed by hand
Technology
AI chat-commerce platform on WhatsApp
Who uses it
Wappin (vendor) for UMKM online sellers, Indonesia
Outcome
none stated
Effort
low
Risks
Wrong orders; customer data on a third-party platform; vendor interview only.
How to adopt
Needs a WhatsApp business account and a product catalogue. Start with automated order replies and handoff to staff. Running it needs conversation review.
How a room runs it
the ROOM for a small seller: the owner, one or two admins and a shop teammate; the agent reads the chats and the order sheet the shop already has and GENERATES the admin's tasks (confirm payment for these five orders, pack by 14:00) as plan steps with owners and due times, and a DECISION for the owner when a buyer asks for a discount or a bulk price, with a deadline; the owner keeps the price and the money; the receipt records each order from chat to dispatch
The engagement on top
a custom engagement: the shop's WhatsApp Business number and order sheet wired in through connectors loaded on this pull, an e-commerce operator on the match, the retail pack's templates; WhatsApp is the door
Source: press · 2024 · vendor only; to corroborate
Warehouse management system linking brand principals, logistics partners and fulfilment for sellers and the partner warung network
Where
ID, Southeast Asia
Segment
SMB and mid (warung network)
Problem
slow, fragmented distribution to small shops
Technology
warehouse management system, fulfilment centres
Who uses it
Kioson (KIOS) and its GudangPintar.id arm, Indonesia (80 warehouses, mainly Central Java, in July 2021)
Outcome
80 warehouses, mainly Central Java, and about 175 GudangPintar.id users (90 % online sellers, 10 % FMCG brands) in July 2021 (Kontan's reporting; the 1,000-warehouse target is a plan)
Effort
high
Risks
Dated (2021); no AI; warung counts not shown.
How to adopt
Needs partner warung and logistics integration. Link principals, logistics and fulfilment in one warehouse system. Running it needs partner onboarding.
How a room runs it
the ROOM for a warung supplier: the distributor owner, the route sales lead, the warung owners over WhatsApp and a supply teammate; the agent reads the order and stock sheets the distributor already has and GENERATES the sales lead's tasks (visit these 12 warungs, restock the fast movers) as plan steps with owners and due dates, and a DECISION for the owner on credit terms for a warung, with a deadline; the humans keep the credit call and the visit; the receipt records each order and delivery
The engagement on top
a custom engagement: the distributor's order system and stock sheet wired in through connectors loaded on this pull, a retail distribution professional on the match, the retail pack's templates
Seller Assistant on the marketplace: onboarding help, listing-quality recommendations, short videos made from one product image and replies to buyer inquiries
Where
multi, Latin America
Segment
SMB and mid (marketplace sellers)
Problem
sellers spend time on listings, media and buyer questions
Technology
generative AI assistant (Seller Assistant)
Who uses it
MercadoLibre, Latin America (countries not stated in the release)
Outcome
daily active users of the Seller Assistant grew more than 40 % month on month in March 2026 (company's own figure, earnings release of 2026-05-07) the company’s own figure
Effort
low
Risks
Platform lock-in; growth rate has no absolute base; vendor-reported results.
How to adopt
Needs a seller account on the marketplace. Switch on the assistant for listing checks and buyer replies. Running it needs review of generated answers.
How a room runs it
the ROOM for an online shop: the owner, one packer and a seller teammate (an agent definition with the marketplace's vocabulary); the agent reads the orders and buyer questions the shop already has and GENERATES the day's tasks (pack these 14 orders by 15:00, answer the three open buyer questions, fix the two weak listings) as plan steps with owners and due times, and a DECISION for the owner (price, restock, pause an ad) with a deadline, landing in the owner's Inbox as a proposal; the owner keeps the price, the supplier call and spend above the envelope; the receipt records each order handled
The engagement on top
a custom engagement: the shop's marketplace store, stock sheet and WhatsApp Business wired to the room through connectors loaded on this pull, an e-commerce operator as the paid professional on the match, the retail pack's templates
Demand prediction for a noodle-stall chain: polynomial regression on sales history of six menus as a stock and sales planning aid (R2 0.91 for the best menu, only fair for drinks)
R2 0.91 for the best menu, only fair for drinks (study)
Effort
low
Risks
Small dataset; demand shifts; method not a deployment.
How to adopt
Needs a few months of per-item sales history. Fit a simple regression per menu and compare with the owner's own stock plan. Running it needs monthly refresh.
How a room runs it
the ROOM for a food stall or small chain: the owner, the cook-in-charge and a stock teammate; the agent reads the sales history the owner already keeps and GENERATES the buying list for the cook-in-charge before each market day as plan steps with due times, and a DECISION for the owner on a large purchase above the envelope, with a deadline; the agent states the basis and its uncertainty, never a promise; the humans keep the buying and the cooking; the receipt records each list and what was wasted or ran out
The engagement on top
a custom engagement: the stall's cashier or notebook sales data wired in through connectors loaded on this pull, the F&B pack's templates; the owner over WhatsApp
Virtual home-improvement advisor for customers, built with OpenAI on the retailer's own expert advice, giving project guidance and product recommendations
Where
US, North America
Segment
enterprise
Problem
DIY and project shoppers needing how-to advice and the right materials without waiting for a store associate
Technology
Mylow, built with OpenAI; desktop and mobile web
Who uses it
Lowe's, United States
Outcome
No outcome measured is published. The release states Mylow launched on desktop and mobile web, open to MyLowe's Rewards members or via lowes.com/Mylow (Lowe's, 5 Mar 2025) the company’s own figure
Effort
high
Risks
Wrong how-to or safety advice; product recommendations that steer purchases; customer data.
Next: Lowe's said Mylow would get voice capabilities and arrive in the Lowe's app later in 2025 (stated 5 Mar 2025) source
How to adopt
Start with a web advisor over your own expert content; add budget and location filters; add voice and the app after.
How a room runs it
the project room: the shopper, the Lowe's-style expert associate and a product teammate; the agent answers how-to questions from the retailer's expert content and proposes a materials list, and anything it cannot answer becomes an assist request to an associate; the receipt records the question, the sources and the products suggested
The engagement on top
connectors to the catalogue and delivery options, an expert-advice knowledge set, a budget and zip-code filter on recommendations
Your own team builds it. Across the 12 adoptions here, 4 were low effort, 3 medium and 5 high; start with the first step above and one owner.
Hire an outside team
A consultancy or forward-deployed engineers from a vendor build it inside your company. Ask for the data they need, who approves each step, and what you keep when they leave. Who builds for each field →
Work with us
A consultation first; then we switch on, connect or change how the work runs, with our engineers in your team for the larger engagements. Every step goes through your approvals and leaves a record. How an engagement runs →
Your own model for this field
Keep your model private. Our engineers adapt an open model to your field's data with you and host it with us; your data and the weights stay yours. Offering it to others in your field is your choice, free for now, open to invited partners first. Your own model →
Companies building AI for one field, from the established leaders to this year's seed rounds. Funding rounds as reported, each with its source. Hiring one of them is the second route in each field above.
Churches use its app and AI tools to turn sermons into clips and posts and to answer member questions.
Customers include Saddleback Church; Harvest Christian Fellowship source
Not done yet
Where AI could go next in each field: our view, not an adoption. No one we know of has done these in this way yet; each says why now, what it takes and what could go wrong.
Agriculture & plantations
A pest-intelligence pool between rival estates, shared under masking
Neighbouring estates that compete on everything else pool their pest and disease sightings through an agent that strips every record down to a grid cell and a date, so each estate's agronomist sees an outbreak front moving towards them before it arrives. Each estate decides what it shares and when; the agronomist still decides what to spray.
Why now
Drone and scout imagery already produce the sightings, but sharing has never happened because the raw data reveals yields and management failures to a competitor; masking that is enforced by policy and proven in a record removes the reason to refuse.
What it takes
Two or three estates willing to sign a sharing rule, their existing scouting or imagery feeds, an agronomist on each side and a masking rule each estate's board has approved.
What could go wrong
Masking that is too coarse makes the signal useless and masking that is too fine leaks block-level yields; one estate may free-ride on the others' scouts.
First step
Map what each estate's scouting data would reveal if shared, agree the coarsest grid that still predicts an outbreak front, and pilot one season on one pest with a short advisory engagement to write the sharing rule.
Harvest-crew pay settled daily by an agent, a receipt per worker
On estates that pay harvesters by the bunch or the kilo, an agent reconciles the weighbridge tickets, the harvest logs and the crew rosters each evening and proposes the day's pay run with a receipt each worker can see on their phone. The field supervisor approves only the lines the agent flags and the manager approves anything above the day's envelope.
Why now
Weighbridges, geotagged harvest logs and mobile money already exist on most estates, and deforestation due-diligence rules in export markets now ask growers to show how the people in the supply chain were treated, not only where the fruit came from.
What it takes
Clean weighbridge and roster data, a pay rule the crews or their union have agreed, a mobile-money rail and a supervisor willing to approve exceptions rather than totals.
What could go wrong
A bad weighbridge feed underpays a crew for a week before anyone notices, or supervisors treat the agent's proposal as the decision and stop checking the flagged lines.
First step
Take one division's last month of tickets, logs and pay sheets and reconcile them by hand in a short engagement; where the three disagree is where the agent and the worker's receipt earn their place.
Technology & professional services
Two companies' agents negotiate the routine terms of a contract
For NDAs, supply agreements and renewals, each company's agent negotiates within a playbook its counsel has approved, trading clauses until both land inside both playbooks, and hands counsel only the terms that fall outside. Counsel on each side decides the exceptions and signs; the record shows every offer and counter-offer both sides saw.
Why now
Contract playbooks are already written down in most legal departments, but no general counsel will let a model bind the company without an approval gate and a record that stands up in a dispute; a policy-driven gate with an append-only record is now ordinary engineering.
What it takes
Both parties' playbooks encoded as explicit rules, a shared record both counsel accept as evidence, and a clear line on which terms are never delegated.
What could go wrong
Two agents converge on terms that are inside both playbooks but commercially foolish, or one side quietly widens its playbook so that nothing ever reaches a person.
First step
Choose one high-volume, low-stakes agreement type and have counsel write down the playbook as if a junior would follow it unsupervised; the gaps in that document are the engagement.
Pay-equity review by an agent that never sees a name
At each pay review an agent works on salaries with identities masked to role, grade, tenure and location, proposes adjustments that close unexplained gaps within the budget the board set, and writes a reasoned note per proposal. The remuneration committee decides, and the record shows what the agent saw and did not see.
Why now
Pay-transparency rules in several jurisdictions now require employers to explain gaps and give workers a right to information, and masking that is enforced and provable makes the review possible without exposing individual pay to a model or a wider circle.
What it takes
A payroll extract with a masking rule HR and legal have approved, a budget envelope from the board and a committee that will read the agent's reasons rather than its totals.
What could go wrong
The masked fields still allow re-identification in small teams, or the agent explains away a gap with a factor that is itself a proxy for the protected characteristic.
First step
Run a re-identification test on last year's pay data under the proposed mask before anything else; a short engagement that ends with a masking rule the works council accepts is the right first deliverable.
Other sectors
A courier's own agent files the appeal against a deactivation
When a platform suspends a courier or driver, an agent on the worker's side gathers their trip record, maps it to the platform's published rules, drafts the appeal and asks the worker whether to file; the platform's reviewer answers in the same room and the record shows what each side saw and when. The worker decides to file; the platform's person decides the outcome.
Why now
Platform-work and data-protection rules now give workers a right to human review of automated decisions and to the data behind them, which makes a two-sided record a requirement rather than a courtesy.
What it takes
Worker access to their own trip and rating data, the platform's rules in a form an agent can cite, a union or association that hosts the worker's side and a platform willing to answer in a shared record.
What could go wrong
Platforms treat the appeals as adversarial volume and route them to a slower queue, or the worker's agent over-argues weak cases and the platform's reviewers stop trusting any of them.
First step
A driver association can start by having one person map a month of real deactivation cases to the platform's rules; where the rules are missing or vague is the brief for the agent and for the conversation with the platform.
A household's standing budget that an agent spends across providers
A family sets a monthly envelope for cleaning, repairs, groceries and school runs; an agent books and pays across several providers inside it, re-books when a cleaner cancels, and brings anything above a line to whichever parent holds that budget. The family decides who holds which envelope and what the agent may never buy.
Why now
Household-service marketplaces, instant payments and agentic booking all exist, but no family hands an agent the card without a per-category budget, an approval line and a receipt they can check, and those are now cheap to build.
What it takes
Linked payment rails, two or three providers that accept agent bookings, a budget rule the family agrees and a receipt the family actually reads.
What could go wrong
The agent optimises for convenience and quietly drifts spend to the easiest provider, or a shared envelope becomes a new source of argument between the people in the household.
First step
A household-services company could pilot a budget-held concierge with ten families and one engagement to design the envelope, the approval lines and the receipt.
Health & care
A personal care budget run by an agent, with the family as approvers
For an older person receiving a personal budget or direct payment, an agent schedules agency and family carers, pays each shift within the care plan, flags missed visits and drafts the quarterly return for the funder. The person and their family approve any change to the plan, and the funder's officer reads the record instead of the paperwork.
Why now
Personal budgets are long established in several social-care systems but their administration falls on families, and the combination of mobile payments, carer check-in apps and a record a public funder will accept now makes delegation to an agent defensible.
What it takes
A care plan written as rules, carer check-in data, a payment rail the funder accepts, and a funder willing to audit a record rather than a shoebox of receipts.
What could go wrong
An agent that pays on check-in data pays for visits that did not happen, or the family becomes a rubber stamp and the person's own voice is lost.
First step
A home-care provider or a council could take twenty existing personal budgets and reconstruct their last quarter as a record; a short engagement that shows the funder what an auditable return looks like is the gateway.
Trial recruitment on masked records, consent before a name appears
A trial sponsor's agent searches a hospital's records with every identifier masked, proposes candidates to the treating clinician who decides who is asked, and only the patient's consent releases a name to the sponsor. The record proves no identity left the hospital before that consent.
Why now
Structured records, eligibility criteria written as code and provable masking make the search possible without the sponsor seeing the data, which is the exact objection ethics boards and privacy officers have raised against recruitment tools.
What it takes
An ethics approval for the masked search, eligibility criteria rewritten as rules, a clinician in the loop for every approach and a hospital that will host the agent inside its own walls.
What could go wrong
Masked records still identify patients with rare conditions, or clinicians approve lists without reading them and recruitment becomes a search engine with a signature.
First step
Start with one trial that is under-recruiting and ask the ethics board which masking it would accept; a consultation that turns the eligibility criteria into testable rules is the first piece of work.
Education & training
Prior learning mapped by an agent, credit awarded by a faculty board
A working adult's portfolio, employment records and certificates are mapped by an agent to the learning outcomes of a programme, with a proposed credit per module and the evidence beside it. The faculty board decides each award and the record travels with the learner to the next institution.
Why now
Micro-credential and skills frameworks now describe outcomes in a form an agent can match against, and the cost of assessing prior learning by hand is what keeps most institutions from offering it at all.
What it takes
Programme outcomes written explicitly, a learner's evidence in digital form, a board that will decide on an agent's dossier and a registrar willing to record credit with its evidence attached.
What could go wrong
The agent rewards learners who can describe their work in academic language rather than those who can do it, or boards approve wholesale and the credential loses its value.
First step
Pick one programme with a queue of mature applicants and run the mapping on ten real portfolios in a short engagement; compare the agent's dossier with the board's own judgement before anything is awarded.
A support-plan steward that sees the diagnosis so teachers need not
An agent holds a student's special-needs support plan, tells each teacher only the accommodation they must apply in their class with the diagnosis masked, notices when a plan was not applied and drafts the review for the parents and the coordinator. The coordinator and the parents decide every change.
Why now
Support plans are a legal duty in many systems and are routinely under-applied because they sit in a file nobody reopens; masking that is enforced lets the plan reach every classroom without spreading sensitive information.
What it takes
Plans in structured form, a timetable and assessment feed, a masking rule the parents have agreed and a coordinator who will act on the agent's flags.
What could go wrong
Teachers feel surveilled by a system that records what they did not do, or the plan hardens into a checklist and the child's changing needs are missed.
First step
A school could take its existing plans for one year group and ask which fields each teacher actually needs; a consultation that produces that masking rule and a review calendar is the start.
Public sector & NGOs
Flood mutual aid between districts, agents trading within authority
During a flood each district's agent posts what it can lend and what it needs — pumps, boats, shelter places, volunteers — and agrees swaps within a pre-approved authority; the incident commander approves anything beyond it. The record shows where every public asset went and when it came back.
Why now
Early warning, risk maps and incident lines are built; what is missing is the logistics between jurisdictions, which stalls on who is allowed to authorise what, and that is a policy and record problem rather than a sensing problem.
What it takes
A pre-agreed authority matrix between districts, asset registers in digital form, radios or phones that reach the agent and a commander who trusts the record enough to act on it.
What could go wrong
An agent moves assets to the district that asks loudest rather than the one in greatest need, or the authority matrix is set so tightly that everything escalates and the room becomes another queue.
First step
Two neighbouring districts could reconstruct last season's mutual aid from messages and logs in a short engagement; the gaps in authority and record are the design brief.
Bid evaluation by an agent that never learns who submitted the bid
In public procurement the evaluation agent scores technical bids with supplier identities, prices and letterheads masked, writes a reasoned score per criterion, and the panel decides with the agent's notes beside their own. The record proves what the agent could not see and what the panel changed.
Why now
Electronic tendering produces bids as documents, and evaluation is the stage where discretion and influence concentrate; provable masking and an append-only record now make blind scoring checkable by an auditor or a court.
What it takes
Tender criteria written as scorable rules, a masking step that strips identifying material from bid documents, a panel willing to have its own changes recorded and a procurement law that allows assisted evaluation.
What could go wrong
Bidders learn to signal identity through the masked text, or panels defer to the agent's score and the human decision becomes a formality.
First step
A procurement office could rescore three closed tenders blind and compare with the original scores; a consultation that reports where the two diverge is the honest starting point.
Mining, energy & utilities
A village micro-grid treasury trading power inside committee tariffs
In a village with rooftop solar, batteries and a grid connection, an agent buys and sells power between households and the grid within tariffs and daily limits the village committee approved, pays each household from the pooled account and brings any tariff change to a vote. The committee decides the rules and the record shows every trade.
Why now
Solar and battery costs keep falling, meters and inverters already report every interval and dynamic tariffs are spreading; the obstacle is that nobody has let a program move a village's money without approvals and a record the members can inspect.
What it takes
Smart meters on every connection, a pooled account with a payment rail, a committee that will set tariffs as rules and a regulator that tolerates peer trading.
What could go wrong
The agent's tariffs favour households with batteries and widen inequality inside the village, or a software fault moves money the committee never authorised.
First step
A cooperative or a developer could model one village's year of meter data against a draft tariff rule in a short engagement before any money moves.
Permit-to-work drafted by an agent, signed by the authorised person
At a mine or a plant, an agent assembles each permit to work — isolations confirmed, gas readings in range, competencies current, conflicting permits on the same area — and drafts it for the authorised person, who signs or refuses. No work order opens without the signed permit on the record.
Why now
Isolation systems, gas detectors and competency registers are already digital and permit errors remain a cause of fatal incidents; the gate is what has kept this manual, and a gate with a record is what makes it safe to assist.
What it takes
Live feeds from isolation and gas systems, a competency register, permit rules written as checks and an authorised person who signs the permit rather than the dashboard.
What could go wrong
A stale sensor feed produces a permit that looks complete, or signers trust the draft and stop walking the area.
First step
Take one area's last hundred permits and check which conditions could have been verified from existing data; a short engagement on that gap decides whether the agent helps or merely adds paper.
Hospitality, travel & events
A group trip's pooled budget, booked and split by an agent
Friends travelling together pool a budget; an agent books flights, rooms and dinners inside it, splits every charge on a receipt each member sees and asks the member who holds the envelope before any booking above the line. The group decides who holds which envelope; the agent never spends beyond it.
Why now
Instant payments, split-payment apps and agentic booking interfaces all exist separately; what no one has put together is a shared purse with approval lines that a group of adults will trust with their money.
What it takes
Payment rails that support a pooled account, booking channels that accept agent orders, and a group that agrees the envelope and the approver before the trip.
What could go wrong
The agent books the most convenient option and the group discovers the receipt afterwards, or one member's late approval holds up a fare that then disappears.
First step
A travel agency could offer a managed group purse for ten trips with one engagement to design the envelope, approval and receipt rules.
Front-desk service recovery from an envelope, not a manager's mood
When a guest's room is not ready or a booking was walked, an agent offers a remedy — upgrade, credit, transfer — inside a daily recovery envelope the general manager set, records the reason and the guest's reaction, and brings anything beyond the envelope to the duty manager. The duty manager decides exceptions; the owner sees what recovery cost and why.
Why now
Guest-messaging agents are common and property systems expose room inventory and rates, but letting an agent give money or rooms away has stalled on budget control and accountability, both of which an envelope and a record provide.
What it takes
Property-system access, a recovery policy written as amounts and conditions, a daily envelope from the owner and staff who still speak to the guest.
What could go wrong
Guests learn the envelope and complain to trigger it, or staff hide behind the agent's offer when a conversation was what the guest wanted.
First step
A hotel could price its last quarter of informal remedies and compare with what an envelope would have paid; a short engagement on that comparison tells the owner whether the rule is worth writing.
Finance, insurance & fintech
A cooperative's loan committee reads files prepared blind by an agent
In a savings cooperative or rural bank, an agent prepares each loan file with the member's name and connections masked — repayment history, cash flow, collateral, group standing — and a reasoned recommendation. The committee decides, and the record shows what each committee member saw and how they voted.
Why now
Cooperative records are moving onto shared systems, regulators are pressing small lenders toward scored credit, and the main governance failure in small lending is lending to friends, which masking that is enforced and recorded addresses directly.
What it takes
Member records in digital form, a masking rule the committee and the regulator accept, a credit rule the board has approved and a committee willing to decide without the name.
What could go wrong
In a small community the file identifies the member anyway, or the committee votes on the agent's recommendation and the agent becomes the lender in all but name.
First step
A cooperative could re-run last year's loan decisions blind and compare outcomes; a consultation that reports the difference and drafts the masking rule is the first piece of work.
An SME treasury agent that moves idle cash inside a signed mandate
A small company's agent watches its accounts, sweeps idle cash into deposits, pays approved suppliers on the best date and keeps a buffer, all inside a written mandate the owner signed with amounts and counterparties. Anything outside it becomes a request the owner approves on the phone, and every move is on the record the accountant reads.
Why now
Open-banking and payment interfaces let software move money and small firms lose yield and pay late because nobody has time; what has stopped delegation is the lack of a mandate enforced in code and a record the accountant and the bank accept.
What it takes
Bank interfaces with payment initiation, a mandate written as rules, an accountant in the room and a bank willing to treat the mandate as the authority.
What could go wrong
A compromised account becomes a compromised agent with payment rights, or the mandate is set wide to avoid interruptions and the owner stops reading the record.
First step
Have the accountant write the mandate the owner would actually sign, then test it against the last six months of transactions in a short engagement before any payment right is granted.
Sport, fitness & personal
A youth athlete's data trust, with scouts seeing masked summaries
A young athlete's training, medical and performance data sits in a record the family controls; clubs' and scouts' agents see masked summaries by age band and position, and an approach, a trial or a transfer offer reaches the family only through consent they can withdraw. The family decides what is shown and to whom.
Why now
Wearables and academy systems collect more on minors than ever, child-data rules are tightening, and the sports industry has no accepted model for who owns a young player's data; masking plus consent on a record is that model.
What it takes
Academies willing to feed the record, a masking rule approved by the federation and the parents, and scouts who accept that a summary is the first contact.
What could go wrong
Families with agents game the summaries, or federations build the trust and then quietly read the raw data.
First step
A federation or an academy could pilot the record for one age group with a consultation to write the masking and consent rules the parents sign.
A power-of-attorney steward for the years when admin becomes too much
For a person losing capacity, or simply overwhelmed, an agent pays bills, renews insurance, books appointments and answers routine letters inside a mandate the person and their attorney set. The attorney approves anything outside it, and the person and, when needed, a bank or a court can read the record.
Why now
Lasting powers of attorney are common, their administration is heavy and prone to abuse, and the pieces — payment interfaces, document reading, a mandate enforced in code with an append-only record — exist together for the first time.
What it takes
A legally recognised mandate, access to the person's accounts and correspondence, an attorney who approves rather than delegates, and a bank that accepts the record as authority.
What could go wrong
The attorney uses the agent to spend outside the mandate and the record is ignored, or the person is sidelined from decisions they could still make.
First step
An elder-law practice or a care provider could pilot the steward for a handful of clients, starting with a consultation on which decisions the person keeps and what the attorney may never delegate.
Logistics, transport & trade
Demurrage and detention settled by three agents against one record
The shipper's, the carrier's and the terminal's agents reconcile container timestamps, free-time rules and exceptions against one shared record and propose the demurrage settlement; each side's person approves or disputes a line. The record is the evidence when they do not agree.
Why now
Electronic bills of lading, port community systems and container tracking produce the timestamps, and demurrage disputes are a chronic drain settled by email; a shared record with approvals turns a fight into a reconciliation.
What it takes
Timestamps from carrier and terminal systems, the tariff and free-time rules in machine-readable form, and three parties willing to accept one record as the reference.
What could go wrong
Carriers decline to feed a record that reduces their revenue, or the agents settle quickly on terms a person would have challenged.
First step
A forwarder or a shipper could reconstruct last year's demurrage invoices against the timestamps they already hold; a short engagement that quantifies the disputes is the business case.
Customs declarations drafted by an agent, signed by a licensed broker
An agent reads invoices, packing lists and certificates, classifies the goods, assembles the declaration and masks commercial values from the forwarder and the trucker who only need the clearance status. The licensed broker signs each declaration and remains liable, and the record shows every field's source.
Why now
Single-window systems accept electronic declarations and classification models are competent, but the broker's licence and liability mean no one can let a program file; drafting with a signing gate and a record keeps the licence where the law puts it.
What it takes
Document feeds from the shipper, a classification rule the broker trusts, masking for the parties who should not see values and a broker who signs line by line at first.
What could go wrong
A confident misclassification is signed by a tired broker, or customs treats agent-assembled declarations as higher risk and inspects more.
First step
A brokerage could run the agent beside its clerks on one lane for a month without filing; an engagement that measures where the two differ sets the signing rule.
Manufacturing & industry
Engineering change control drafted by an agent, approved by quality
When a drawing or a supplier part changes, an agent drafts the change order — affected assemblies, open orders, work instructions, supplier notices, test requirements — and routes it. The quality manager and the customer's engineer approve, and the record is the trail the certification auditor reads.
Why now
Product-data and ERP systems hold the links an agent needs, and quality standards for regulated supply chains already demand exactly this trail, which is why changes queue for weeks behind engineers who write it by hand.
What it takes
A product-data system with bills of material, change rules written as checks, a quality manager who approves the agent's draft and a customer willing to receive notices from a record.
What could go wrong
An agent misses a downstream assembly and a changed part ships unapproved, or engineers approve drafts in bulk to clear the queue.
First step
Take the last fifty change orders and check how many links the agent could have found from existing systems; a short engagement on that audit shows where the value and the risk sit.
A machine with its own maintenance budget, spent within an envelope
Each critical asset has a maintenance envelope for the year; its agent orders spares, books technicians and schedules downtime inside it from the condition data and the production plan, and brings anything above the envelope or outside the plan to the plant manager. The manager decides exceptions; finance reads the receipt per machine.
Why now
Predictive maintenance already says what will fail; what has not happened is letting the prediction spend money, because no plant lets a model buy parts without a budget, an approval and a record.
What it takes
Condition data per asset, a spares catalogue and supplier interfaces, a budget per asset that finance has set and a manager who approves exceptions rather than every order.
What could go wrong
The agent spends the envelope on false alarms, or a cheap part is ordered while the real failure mode is missed.
First step
Choose five assets with the worst unplanned downtime and let the agent propose, not order, for a quarter; an engagement that compares its proposals with what was actually done settles the envelope.
Media, creative & marketing agencies
Rights clearance for every asset in the cut, bought within the budget
Before a film, campaign or episode is delivered, an agent checks every asset in the cut — music, footage, faces, fonts, generated elements — against the licences the production holds, proposes the purchases needed within the clearance budget and attaches each asset's provenance to the deliverable. The producer approves the purchases and the lawyer decides the hard cases.
Why now
Content-credential standards and licence marketplaces make rights machine-readable, generated assets have made provenance a contractual question, and clearance remains a manual spreadsheet that fails at the worst moment.
What it takes
A licence register in digital form, edit-system exports listing the assets, a clearance budget from the producer and a lawyer who reviews the flagged cases.
What could go wrong
The agent clears an asset on a licence that does not cover the territory or the term, or buys rights nobody needed to meet the deadline.
First step
Audit one finished production's assets against its licences in a short engagement; the misses in that audit are the case for the agent.
A verification room where the source stays masked from the editor
A newsroom's agent helps a reporter verify leaked documents — metadata, cross-references, consistency with public records — while the source's identity stays masked from everyone including the editor until the editor decides to publish. The record proves who saw what, which protects the source and the paper.
Why now
Document-analysis models are competent and legal pressure on newsrooms to reveal sources is growing; masking enforced in software with a provable record gives a reporter a way to work with colleagues and tools without widening the circle.
What it takes
A newsroom security policy that defines the mask, a self-hosted deployment, a reporter who still meets the source and an editor who accepts deciding without the name.
What could go wrong
The documents themselves identify the source regardless of masking, or the record becomes a discoverable map of the newsroom's sources.
First step
A newsroom could run a tabletop with its lawyers on one past story to decide what the record must and must never contain; that consultation precedes any software.
Construction & property
Subcontractor pay released down the chain when the client pays
When the client's payment reaches the main contractor, an agent releases each subcontractor's certified amount within retention rules the same day, holds only the lines in dispute and shows every subcontractor its own line on the receipt. The project manager approves deviations and the client can see that its money reached the trades.
Why now
Progress is measured by drones and diaries, payment applications are digital and late payment down the chain is a chronic cause of insolvency; what has been missing is a rule that pays automatically with a record and a gate for exceptions.
What it takes
Certified payment applications in digital form, retention and set-off rules written as checks, a payment rail and a main contractor willing to let its payment behaviour be visible.
What could go wrong
The main contractor sets the dispute flag liberally and nothing is released, or a rule pays a subcontractor whose work has since failed inspection.
First step
A client or a main contractor could trace one project's payments from client to trades over the last year in a short engagement; the delays found are the design brief.
An apartment building's committee with an agent that runs the fund
In a strata or owners' association, an agent collects service charges, runs small maintenance tenders, pays contractors inside the committee's budget and prepares the reserve-fund review; owners vote on anything above the budget and read every payment on the record. The committee decides the budget and the exceptions.
Why now
Building-management software and payment rails exist, but owners' committees are volunteers who neither trust a manager's spreadsheet nor have time for the admin; an agent inside a budget with a ledger every owner reads is a new middle.
What it takes
A register of owners and units, a budget approved at a general meeting, a payment account and a few contractors who accept agent-run tenders.
What could go wrong
The committee delegates everything and stops meeting, or a contractor learns the budget rule and prices to it.
First step
A managing agent or a committee could rebuild last year's accounts on a record with per-owner receipts in a short engagement; the owners' reaction tells whether the fund can be delegated.
Community & civic
Zakat and relief distribution decided on blind files an agent prepared
A mosque, church or relief committee receives applications for help; an agent prepares each case with the applicant's identity masked — need, household, prior support, verification — and a recommendation against the committee's rules. The committee decides, the applicant's dignity is kept and donors see aggregate receipts of where their money went.
Why now
Collection has gone digital while distribution is still decided in rooms where everyone knows everyone; masking that is enforced and a record donors can read answer the two complaints most often made about charitable distribution.
What it takes
Application forms in digital form, a distribution rule the committee has written down, a masking rule the religious authority accepts and a committee that decides without the name.
What could go wrong
In a small congregation the case identifies the person anyway, or the rules exclude the people who most need discretion.
First step
A committee could re-decide last year's cases blind and compare; a consultation that writes the distribution rule and the mask is the first piece of work.
Participatory budgeting where the agent costs every proposal
Residents propose projects for a town's participatory budget; an agent costs each one against the municipality's rates, checks it against zoning and procurement rules, flags what is not allowed and why, and publishes the list. Residents vote, the council's officer approves the final allocation and the record follows each project to completion.
Why now
Participatory budgets are widespread but stall at the costing stage, where officials cannot assess hundreds of ideas; an agent that costs under published rules with a record of its reasoning makes the stage cheap and checkable.
What it takes
Municipal unit rates and rules in a form an agent can cite, an officer who reviews the costings, and a council willing to publish the record including what was rejected and why.
What could go wrong
The agent under-costs popular projects and the council takes the blame, or residents learn to shape proposals to the costing rule rather than to need.
First step
A municipality could cost last cycle's proposals with the agent and compare with its officers' figures in a short engagement before opening the next cycle.
Retail, F&B & e-commerce
Small sellers' agents pool orders to reach a supplier's price break
Several independent shops or marketplace sellers let their agents pool orders for the same goods until a supplier's price break is reached; each seller approves its own share and pays only that, the supplier sees one order and the record shows who paid what. Each seller decides the ceiling and whether to join.
Why now
Supplier catalogues and payment splitting exist, and small sellers lose margin to distributors because they cannot coordinate; shared money with per-member approvals and a receipt is the governance that makes a buying group possible without a cooperative's overhead.
What it takes
Supplier catalogues with price breaks, a pooled payment arrangement, sellers who accept the agent's timing and a rule for what happens when one drops out.
What could go wrong
One seller's non-payment strands the group's order, or the pooled volume attracts suppliers who later refuse to sell to members singly.
First step
A distributor or a seller association could run a buying-group pilot on ten products with an engagement to design the pooling rule, the approvals and the dropout rule.
Allergen change control from the supplier's substitution to the menu
When a supplier substitutes an ingredient, an agent reads the new specification, updates the allergen matrix for every dish that uses it, drafts the label and menu changes and tells the kitchen. The head chef approves the change, and the record shows, for any customer who asks, what was in the dish on the day.
Why now
Allergen-labelling law has tightened in several countries after fatal incidents, supplier specifications are digital, and the failure is almost always a substitution nobody traced to a dish.
What it takes
Recipes and supplier specifications in structured form, a change rule the chef accepts, and front-of-house staff who read the record rather than guess.
What could go wrong
A supplier's specification is wrong and the record faithfully carries the error, or the chef approves in bulk during service.
First step
A restaurant group could trace one month of substitutions through its recipes by hand in a short engagement; the untraced ones are the argument for the agent.
Talk to us.
Your company, its size and the first job you would hand to an assistant. We answer by email, nothing else.
