---
title: How the advisor works
description: What the AI advisor can read, why it cites facts and never computes them, why every action is a proposal, and how budgets and permissions bound it.
order: 3
updated: 2026-09-16
keywords: AI, advisor, proposals, citations, read-only, budget, permissions
---

The advisor is a language model with a very short leash. Understanding the leash is the point.

## It reads the same figures you do

When you ask "why is food cost up?", the advisor does not do arithmetic. It calls read-only tools that return figures already computed by the product, the same ones the dashboard shows, scoped to your permissions. The answer streams with citations to those figures, and you can click each one. A number that cannot be bound to a figure is not shown; the advisor is built to say "I don't have that data, here is how to get it" instead.

## It never writes

The advisor has no tool that changes anything. When it suggests an action, the suggestion becomes a proposal card: the exact action, its details, a risk class and a before-and-after diff marked exact or approximate. A person approves it, edits it with a reason and a rollback note, or dismisses it. The approval runs the action under that person's name and permissions, through the same checks as if they had clicked the button themselves.

This is enforced in code, not in the prompt. Even a manipulated question cannot make the advisor post, pay or file. The worst outcome is a blocked or escalated answer.

## It respects who is asking

The advisor sees only what your login can see. A manager for one outlet cannot ask about another. Kitchen and counter screens do not show the advisor at all.

## It is budgeted

Each organisation has a usage budget with a reserved share for the morning briefing, so chat cannot starve it. When the budget is exhausted the advisor says it is unavailable rather than cutting an answer short or silently switching to a cheaper model.

## It reads invoices the same way

Invoice reading uses the same provider layer. The draft it produces is never accepted by itself; a person checks it, and only documents read at 95% confidence or higher are offered for bulk accept.

## What it is not

It is not autonomous, it does not run your restaurant, and it does not learn from your data to serve other customers. It is a faster door into the same truth, with a person on the other side of every action.

## Related

- [Ask the advisor and approve a proposal](/docs/how-to/ask-the-advisor)
- [How ChowChirp keeps numbers trustworthy](/docs/explanation/trustworthy-numbers)