From sign-up to your team answering in its inbox, in an afternoon.

Two ways to run it

The playbook is the same plain files either way. What differs is how much runs for you today.

1Sign up with GitHub

Each PlatAtlas organization is one GitHub account: your organization's, or your own if you start alone. Install the PlatAtlas GitHub App on it, sign in, and set it up; after that, its members sign in with GitHub. GitHub says who you are, and it is where your playbooks are kept: each one, with every email you forward to it, is a private repository in that account. PlatAtlas reads and writes them there and keeps no copy.

Each person who answers needs a GitHub account. An invitation link binds the GitHub account that opens it.

2Point at the answers your team already gave

Past requests with the column that says what was decided: paid or held, approved or sent back, whatever your people wrote. Fifty is enough to start; more is better.

On platatlas.com today, that is a spreadsheet export (CSV or Excel) you open at /demo/start/, or your email. Starting with your email? It needs no history. On your organization's page, Set up my inbox makes the playbook and gives you an address to forward to; your first answers become the history.

On your own server today

Your steward connects them, read-only, from wherever they live:

It only ever reads. Passwords stay on your side, never in PlatAtlas's files (a forwarded mailbox needs none); web pages are read politely, honouring each site's robots.txt.

Check your export now

Or go further: see the rules your team already follows and download your playbook, from the same file, in this browser.

Drop a CSV, Excel or JSON export here to see what PlatAtlas would make of it. It is read in this page and never uploaded.

3It becomes your team's playbook

1. It says what it found: which column holds the decision, and every column that looks like personal data (emails, phone numbers, addresses, account numbers). Those are left out unless you ask to keep them.
2. Your past answers become the test. An AI model is graded on how many of your own past answers it would have got wrong, and what it costs. On your own server that happens before you switch to it, and you keep the cheapest one that clears your limit; on platatlas.com the agent's check runs when someone who manages the playbook asks.
3. Nothing is decided alone until your steward allows it. Until then every request reaches a person. After that, by default 1 in 10 of what the AI may finish still does, and under a grant never fewer than 1 in 10.
4. It's yours from the first minute: a folder of plain files, kept in your organization's own repository.

Today you can do this in your browser at /demo/start/, where Create makes the playbook's private repository in your GitHub organization, or with one command on your own computer.

4Your steward's hour

Your steward is whoever knows how these requests should be handled. On one page, they write what each answer means in your team's words, next to the past requests the AI got wrong (this is where accuracy comes from); who may give which answers; and what, if anything, may finish without a person. That last one comes with evidence from your own history, and nothing finishes alone until the steward presses Allow.

On our made-up demo company (60 past invoices, Claude Sonnet 5), the evidence said: “at 0.7 or above, 27 of 29 past paid cases would have closed with no person, and 0 of those calls were wrong.” Bank-detail changes and possible duplicates stayed with a person: only 5 past cases of each, too few to lean on. Measured 2026-09-23Fictional company

One run of the sample inbox's check; a different run can land a case or two either way. The run behind it.

On your own server, the steward can also write a grant: one agent may finish one kind of answer. A grant needs at least 5 past cases of that kind with none wrong, lasts at most 180 days, and still sends at least 1 in 10 of the agent's answers to a person. Under a grant the agent signs each answer as itself, never as a person.

5Your team answers where it already works

New requests arrive on their own. On platatlas.com, that is the email you forward. On your own server, it is also anything that can send a web request (Power Automate, Zapier, a script), or PlatAtlas re-reading the same database, folder or mailbox for anything new.

6Every week, the check

The AI is graded again on your team's answers, including every one given since. If it misses more than your limit, the check fails and says which answers it gets wrong. If a later check finds a wrong answer in a situation an agent was granted, that grant is suspended. This runs today on your own server; on platatlas.com it comes next. There, an AI agent on a playbook is checked against your team's past answers when someone who manages the playbook asks.

Where your data goes

Said plainly. On platatlas.com, starting from an export, forwarded email and the browser inbox run today. An AI agent that guesses at each forwarded email runs there too, once someone who manages the playbook adds one: its guesses stay hidden and are only compared with your team's answers until that person chooses to show them as suggestions. The weekly check comes next there, and runs today on your own server. Files, SQLite, web pages and APIs are read in tests on real data; the other databases and mailboxes connect through their standard drivers but have not yet been pointed at a customer's system. Outlook, Teams, Slack and the Power Automate connection are built and tested against Microsoft's and Slack's rules; none has run at a customer yet. An answer records who made it: a person by name, or an agent as itself. It is not yet signed with a key that proves it.

Rather talk first? Write to us and tell us the request.