PLATFORM
How our agents work.
The platform runs teams of AI agents across a whole company: support, sales, marketing, finance, HR, procurement, engineering, product, design, content and management. Every agent follows the same steps, and each step is written down so a reviewer can follow it.
7 industries · 11 departments · 163 roles in every industry
HOW EVERY AGENT WORKS
Seven steps, the same for every agent.
- Step 1Looks it up
It searches its department’s knowledge base: policies, playbooks, price lists.
- Step 2Reads the request
It turns the message into structured fields, such as the guest, the invoice or the requisition, and leaves blank anything it was not told.
- Step 3Checks the facts
It reads the records it needs from the company’s own system, such as the property management system, the core banking system or the ERP.
- Step 4Forecasts, when it decides
A decision-maker predicts what its proposal would do, such as cash after a payment run or occupancy after a rate change, before it proposes it.
- Step 5Works it through
It applies the rules it keeps, with the limits set in its settings.
- Step 6Asks a person
When the rules say so, it stops and shows the approver a plain summary. The approver can approve it, edit it, or send it back.
- Step 7Acts
Only then does it write to the system it acts in.
Three ways an agent thinks
Assistant · 84 per industry
Looks things up, works the request through and drafts the answer or the change.
Decision-maker · 51 per industry
Forecasts the effect of what it proposes before it recommends it.
Investigator · 28 per industry
Weighs competing explanations against the evidence before it concludes.
Built to be trusted
It never invents facts
When a record is missing, the agent says so and asks for it. It does not make up a number to fill the gap.
People stay in charge of money and commitments
Five agents always need their owner’s sign-off, in every industry: Payroll, Accounts Payable, Refunds and Returns, Contract Preparation and CV Screening.
Limits are settings, not code
Every agent has settings, such as a refund limit, a price tolerance or a goodwill cap. A manager changes them without a developer.
Every rule is written down
Each agent lists the rules it keeps, in plain language, on its own page.
Your systems, your choice
Agents name no vendor. They connect to whichever property management system, ERP, CRM or core banking system the company already runs.
Each company is separate
Every company runs as its own organisation, with its own people, roles, knowledge and model credentials.
ONE WORKFORCE, ELEVEN DEPARTMENTS
The same 163 roles in 22 teams, in every industry.
A company can start with one department and grow into the rest.
FROM A REAL RUN · HOTEL
We ran one agent in every department of every industry: 77 runs. Every one reached its approval step or finished.
This is an illustrative night audit for Hotel dated 2025-03-14, with all figures labelled as sample data. Room revenue (18,450.00), F&B (6,195.75), spa (960.00) and card/cash payments tie to the audit totals, but taxes are 429.35 short, so nothing posts. Nine lines, including the 1,500 deposit and 432.05 service charge, are parked in 9999 Suspense because no standing mapping exists. Once approved, the coded lines would be sent as suspense or draft entries only, with no revenue or tax journals posted until the questions are resolved.
Details are made up for the test, and the agents say so.
FROM A REAL RUN · BANKING
Banking: Customer Support Agent
A frustrated, unverified retail customer reports a card payment declined abroad and asks for a written explanation and a refund of a foreign transaction fee of about 4.50 GBP. This is logged as a complaint, and a possible fraud block is referred to the fraud team. No records were available from the core banking system and the knowledge base returned nothing relevant, so the cause of the decline and the fee tariff are unconfirmed. If approved, the complaint is logged and acknowledged within 1 business day, the customer is asked to verify without sharing card secrets, and any refund above 0 USD is decided by you after the fee is confirmed against the verified account.
Details are made up for the test, and the agents say so.
FROM A REAL RUN · HOTEL
Hotel: Guest Support Agent
No guest case exists, so nothing is being sent or offered. The request asks for an illustrative example, and the intake found no hotel, guest, reservation, dates, policy content or PMS records. Invented details were declined because they would be presented as facts. Once the missing details are supplied, a live case can be opened and the reservation, folio and policies checked before any reply.
Details are made up for the test, and the agents say so.
FROM A REAL RUN · IT COMPANY
IT Company: Front-End Agent
This is a planned TOTP two-factor enrolment flow for the Northwind Customer Portal settings page, estimated at 16 hours on story NW-4821 with design reference DS-ACCT-2. No code, bundle figures, visual diffs or hours have been measured because the repository is not connected and the knowledge base returned nothing relevant. Missing items are the SOW browser matrix, the contract WCAG level, the design link, estimate approval against the SOW change budget, repository access and base branch, the status of the verification and recovery-code APIs, and the settings-route bundle budget. Once these are supplied, the work proceeds and any bundle, visual-diff, WCAG or 25 percent overrun trigger will be raised for your approval before a pull request is opened.
Details are made up for the test, and the agents say so.
INSIDE THE PLATFORM
Every agent, with its state, runs, success rate, cost and outcomes.
Each industry’s Customer department, as its agent cards in the platform, grouped by team.






