How it works, in four pictures.

Every example here is Larkspur Garden Co., a made-up company whose accounts payable team uses PlatAtlas. Fictional The suggestions come from runs of AI models on its invoices; the failing check is a stand-in with no AI model.

1

One request, start to finish.

The AI suggests. Your team's rules decide what happens next, and anything they don't clearly allow waits for a person.

A REQUEST ARRIVES Invoice email Corvel Fleet Parts$768.66 PLATATLAS SUGGESTS Hold: bank details changed98% sure THIS RECORD'S RULES A bank change goesto a person every time, however sure it is. FINISHES ON ITS OWN Only answers your teamallowed, e.g. "Pay it"when 97%+ sure. WAITS FOR A PERSON Dana, AP lead,decides. It's written down,with Dana's name.
The AI was 98% sure, and the invoice still went to a person, because a change of bank details is the classic way payments get stolen and this record's rules say so. What may finish on its own is written by your team, not decided by the AI.

Source: the made-up company's rules (thresholds.yaml in the example record) and one AI suggestion from a run on its invoices.InferredFictional

2

Who may decide what, wherever they work.

Outlook, Teams, Slack or a web browser only show the request and pass on who clicked. PlatAtlas checks that person against your team's list before anything is written, so the rule is the same everywhere.

WHERE THEY CLICK Outlook Teams Slack Web browser YOUR TEAM'S LIST Who clicked?May they decide this one?Refused: nothing is written. Sam, AP clerk Pay it · hold for receipt · price question. Not bank changes. Dana, AP lead · Priya, Controller Anything, including bank changes. The assistant (Ask Atlas, or any AI agent) Reads, explains, drafts an email for you to send. Has no button to press. An agent signs only what your steward grants, as itself. The browser works today. Outlook, Teams and Slack are built and tested, and not yet running at a customer.
If Sam clicks "Hold: bank details changed", PlatAtlas refuses before anything is written, in every tool. Because the list lives with your team's rules and not inside Teams or Slack, switching tools doesn't change who may decide. Each person who answers in the browser signs in with a GitHub account.

Source: the made-up company's list of people (people.yaml in the example record).DeclaredFictional

3

Your team's know-how is yours. The AI is rented.

Only the bottom layer is your organization's asset. Everything above it can be swapped without touching it.

WHERE PEOPLE WORK · RENTED Outlook · Teams · Slack · a web browser · any AI assistant that can read it THE AI MODEL · RENTED Claude, a small open model, or next year's. Switch with one setting, then grade it with the same check. YOUR TEAM'S PLAYBOOK · OWNED A folder of plain files your organization keeps. What each answer means · who may decide what · what goes to a person · every past decision, and why On your own server, or on platatlas.com with the files in your own GitHub repository. Inside your Microsoft 365: designed, not built.
When someone retires, their successor starts from the playbook. When a vendor changes, you remove it and the playbook is untouched. If your office runs on Microsoft 365, keep it; the playbook renders into it. Anyone can open the files with a text editor: see the example record.

4

A check that can fail.

The AI is graded against 60 past invoices whose answers are already written down. The team's limit is 5% wrong. Above it, the check fails and says so.

A stand-in, no AI model(Deterministic stand-in) 34 misses · 56.7% · fails Claude Sonnet 5, the sampleinbox's past invoices 2 misses · 3.3% · passes the team's limit (steward's bar: 5%)
The stand-in answers "pay" for every invoice, so it misses every one whose answer was something else. It failed, and said so. That is the point: a check that can't fail proves nothing. Switching to a different model is a one-line change, and it is graded by the same check. Every run behind every number on this site, with its date, is on The runs.

Source: Measured 2026-09-22 · 60 past invoices · a deterministic stand-in, not an AI model · made-up company. Measured 2026-09-23 · 60 past invoices · Claude Sonnet 5 · the sample inbox · made-up company.MeasuredFictional

See it for yourself.