Who can use what

This page is an overview for administrators: which group, capability and rule each AI feature depends on, and where to change it. The principle is always the same: AI works with your rights and can only be narrowed, never widened.

The three layers

  1. Groups. AI: User lets a person use the assistant and the one-click features. AI: Administrator lets a person configure providers, policies, agents and sources. Both are set on the user form under the AI category.

  2. Capabilities. Master switches under AI ‣ Configuration ‣ Behavior ‣ Capabilities decide which classes of tools exist at all: Ask (QA), Read data, Navigate UI, Write data, Delete data, Web access, MCP tools, Business actions, Customize UI & schema and Guided system setup. Write, delete, web, MCP, actions and customization are off after installation. A capability can be limited to groups.

  3. Access rules. Under AI ‣ Configuration ‣ Security ‣ Access Rules, per-model and per-field allow and deny rules sit on top of Odoo’s own rights. By default reading is allowed and creating, writing and deleting are denied. See Access policy.

What each feature needs

  • AI panel, writing assistant, AI search, command palette, dictation: AI: User, a chat model, the Ask (QA) capability. The assistant sees record data only if Read data and the access rules allow reading that model for you.

  • Changes made by the assistant in chat: Write data or Delete data, plus an allow rule for the model and fields, plus the write mode. Keep confirmation on.

  • One-click buttons in apps (assign a lead, describe a product, draft timesheets, file a document): the click is your decision, so they do not need the write capability, but they read data through the same access rules. If the AI is not allowed to read the model, you get a clear message instead of a result.

  • AI fields and AI server actions: the user chosen in Acts as (or the agent task’s agent user). See AI fields, server actions and automations.

  • Knowledge sources: sources are managed by AI administrators; each person only gets excerpts from sources and records they may read themselves (Knowledge sources).

  • MCP server: internal users in AI: User; only an allow-list of tools; changes always wait for your confirmation (MCP server).

  • Live chat agent: a dedicated service user that can only ask, read the data sources you list and hand over; it gets no write rights beyond creating a lead, if you turn that on (Live chat AI agent).

  • Website builder tools: Customize UI & schema and the website designer group; every change is a confirmation card and can be undone (Website AI). Importing a website also needs the Settings administrator right (Import a website).

  • AI triage of project tasks: reviewer in AI: User, the Write data capability and an allow rule for writing the task fields it changes (AI for projects with SLA).

Limits and traces

  • Rate and cost limits per person and per day are under AI ‣ Configuration ‣ Settings (see Configuration). Repeated policy violations earn strikes and a temporary ban.

  • Every model call is written to the AI log and every applied change to the change log, both under AI ‣ Monitoring ‣ Logs and AI ‣ Monitoring ‣ Mutation Logs (see Monitoring).

Tip

Roll out in steps: start with the assistant and the writing features for a pilot group, add knowledge sources, then enable one write path at a time and read the logs for a couple of weeks before widening it.