Assistant: it reads your data, proposes, you approve.
In the office
Assistant: questions on your data, proposed actions, approval
Knowledge base: procedures, owner, review date
Mailbox: invoices extracted and matched, anomalies, audit log
On the tablet
No AI during service: the tablet stays the tablet
The night's anomalies are read in the office
Orchestrator and agents
One entry point, twelve specialised agents: sales, reservations, guests, purchasing, invoices, mail, stock, costs, HACCP, staff, knowledge. Each only has the tools of its domain.
Autonomy policies
Every proposed action waits for your approval by default. Sales, payments, closes and fiscal data are out of the agents' reach, by construction.
Mailbox and invoices
The restaurant's mailbox (IMAP) is read by the mail agent: supplier invoice extracted, matched to the purchase order, prices compared to the price list.
Knowledge, anomalies, log
Procedures with an owner and a review date. Every night, deterministic checks open anomalies (sales outside a Z, stock, forbidden payments, failed jobs). The audit log says who did what, by scope.
Inside the app
app.brigades.io/mail
Mail: the supplier invoice is extracted and matched to the order.
app.brigades.io/knowledge
Knowledge: procedures, owner, review date.
app.brigades.io/anomalies
Anomalies: the night's checks, open or already resolved.
app.brigades.io/audit
Log: who did what, by scope.
What changes
A question about yesterday's service gets a figure, not a feeling.
The invoice arrives matched to the order; you only confirm.
What drifts at night is visible in the morning, with its cause.
A model of your choice, with your key
Anthropic, OpenAI, Mistral or Google: the provider and the key are the restaurant's, with a capped monthly budget and alerts at 80 % and 100 %. The data sent to the model is that of the question asked.
