AI team and permissions
An AI team is a set of agents staffed on one app. They read the app's data and inbox, and they propose actions in it. What they may do on their own is a decision you make per kind of action, with a money limit you set, and everything they do is recorded.
Staffing a team#
Every business system suggests a starter team for its trade: a receptionist, a follow-up agent for quotes and invoices, a bookings agent for a salon, a work-order agent for a garage. New apps are staffed automatically when they are created; the Team screen shows who is on the team and lets you add, remove or pause.
Agents are built on the same foundation as the Agent Builder: a persona, the app's data as tools, the Brain as memory. The difference is that a staffed agent can only act inside its app, through proposals.
What agents can do#
| Read | Act (through a proposal) |
|---|---|
| Open conversations and messages | Reply to a customer |
| Bookings, orders, jobs, invoices | Confirm, reschedule or cancel a booking |
| Customers and their history | Draft and send a quote or a reminder |
| The app's facts in the Brain | Mark an order ready, a car collected, a visit done |
| Live numbers: visits, sign-ups, sales | Record a fact in the Brain, or a standing rule |
The exact set depends on the business system. Agents never get the app's database credentials or secrets, and they never execute anything that is not on this list.
Permissions: ask, auto, off#
For each kind of action, choose from the app's Permissions section:
| Setting | Meaning |
|---|---|
| Ask (default) | The agent proposes; you approve or reject |
| Auto | The agent does it and tells you afterwards with a quieter push |
| Off | The agent must not propose this kind at all |
Rules can be narrowed: auto, but only below 50 euro, ask for this agent, auto for that one. These are the floors that always hold whatever you set:
- Anything high or critical risk is never automatic. Money kinds (refunds, cancellations with a charge, budget changes) are high risk by definition, and any amount above a threshold is too.
- Automatic actions are capped per day. Past the cap the agent asks again.
- Nothing is automatic while the app is in sample mode or while the team is paused.
The money mandate#
An action that moves money is automatic only if you set an explicit limit and the amount is below it. "Auto-refund up to 30 euro" is a mandate; "auto-refund" without a number is not accepted. The limit is yours to raise or lower at any time, and the weekly report shows how much went out automatically.
The trust ladder#
You do not have to configure any of this up front. After you have approved the same kind of action five times without editing it, the weekly report offers to switch that kind to automatic, with a limit derived from the amounts you approved. Accept in one tap, or ignore it. The offer repeats only if your behaviour keeps saying yes.
Audit trail#
The Audit Logs tab of the app's dashboard, and the app's History on the phone, record every action with:
- who proposed it (which agent) and who decided (you, a policy, or an automatic rule),
- the surface the decision was made on (web, phone, lock screen, assistant) and whether Face ID was used,
- the risk level, the amount if any, and a fingerprint of the exact parameters executed.
Denied proposals are recorded as well, with the rule that denied them. Export the log as a file from the dashboard.
Pausing the team#
Each app has a Pause AI team switch (phone: the app card's menu; web: the Team panel). While paused, agents stop proposing, loops stop advancing, and anything that was automatic waits for your approval instead. Nothing is deleted.
Lean by design#
Agents do their unattended work (sweeps, scheduled checks, briefings) on the most economical capable model and skip a cycle when nothing in the app has changed since the last one. Interactive work you trigger keeps the full-strength model. You will not see this; it is why a quiet week costs almost nothing.