Every business platform now talks about "AI agents", and it is hard for decision makers to separate what actually works in the system from marketing. This article sticks to what the Odoo 20 release notes and the official agents documentation say: what an agent can do, how usage is charged, and what to check on data and access rights.
What is an AI agent in Odoo?
Agents appeared in Odoo 19 as assistants you can chat with, which learn from your documents and can take actions; Odoo 20 extends them. According to the documentation, each agent has:
- Instructions (system prompt): the agent's role, responsibilities and constraints.
- Knowledge (sources): uploaded files, web links, Documents files and Knowledge articles, indexed for retrieval. "Restrict answers to Sources" limits the agent to them.
- Skills and tools: skills define tasks, tools perform actions. The documentation says an agent with no skills can only provide information.
- Triggers: what invokes the agent, such as a user conversation or an automated or scheduled action.
What an agent can do in Odoo 20
- Create records, including from an uploaded instruction file such as a PDF.
- Update records on request.
- Agentic automation: agents can be summoned by automated and scheduled actions, with a chat showing how each run was handled.
- Work with files: upload files and link documents in chat, and ask about a file while previewing it.
- Generate images for websites and mailings.
- Voice: dictate requests directly to an agent.
- Self-update: agents can propose or make changes to their own instructions and skills.
- Live progress, clickable answers to agent questions and permission requests, and confirmation when the maximum tool-call limit is reached.
- Subagents: an agent can delegate to other agents.
- Accounting: ask the assistant for insights based solely on the accounting data in your database.
"Topics" were renamed "Skills", and agent conversations are now stored for up to 30 days after the last message.
Connecting through MCP
Odoo 20 lets external agents connect to your database through the Model Context Protocol. According to the MCP server documentation:
- Odoo acts as an MCP server; an external agent reads or modifies data through "tools", which are server actions.
- Authentication uses OAuth (recommended, for internal users) or an API key.
- The connection runs with the user's identity and permissions in the database.
- Five read tools are available by default; other tools stay hidden until an administrator exposes them.
- Each connection appears as an API key in the user's preferences; deleting it revokes access.
- The documentation warns that enabling Dynamic Client Registration (DCR) lets any application register as a client.
Cost: IAP credits
The Odoo 20 release notes state that all AI features now need credits, bought via IAP (in-app purchase), and the agents documentation confirms AI requests consume them. The Odoo pricing page also lists "Agentic AI" in the Custom plan. So:
- Start with one specific use case and monitor credit consumption during a pilot before scaling.
- Scheduled automations can consume credits without human involvement; scope their frequency carefully.
Data, privacy and access: what to review
We found no detailed statement on data processing in the release notes or the agents documentation. The Odoo 19 documentation says Odoo supports several versions of ChatGPT and Gemini models, so requests are processed by external model providers. Before enabling agents, review:
- Odoo's terms and privacy policy, and your own data obligations and applicable regulations, with your legal adviser.
- Sources: do not upload sensitive documents (payroll, contracts, personal data) to an agent that does not need them.
- Skills and tools: give each agent the minimum tools; an agent that can update records needs considered access rights and confirmations.
- MCP: expose write tools only when needed, avoid DCR without reason, and review access keys regularly.
- Human review: financial decisions, journal entries and contracts stay with a responsible person; the agent proposes and speeds up.
How to measure a first pilot
| Measure | Example |
|---|---|
| Time | Average task time before and after the agent |
| Quality | Share of outputs needing human correction |
| Cost | IAP credits consumed versus work completed |
| Risk | Actions rejected or requiring confirmation |
Suitable first use cases: drafting records from incoming files, answering staff questions from company policies uploaded as sources, or summarising overdue customers for review. For more on evaluating AI projects, read AI solutions for business.
FAQ
Are AI agents included at no cost in Odoo 20?
No. According to the release notes, all AI features need IAP credits.
Can an agent change data?
Yes, it can create and update records if given the necessary skills and tools, which is why this must be configured carefully.
Do I need Odoo 20 for agents?
Agents exist since Odoo 19, but capabilities described here, such as updating records and MCP, come from the Odoo 20 release notes. See our Odoo 20 guide.
Next step
If you are considering Odoo agents, Maham Pro, an official Odoo partner at Ready Partner level, can help choose a first use case, set access rights and sources, and define pilot metrics as part of our Odoo service and AI solutions. Start with a free 30-minute introductory session.
Sources: Odoo 20 Release Notes, AI agents (20.0), AI MCP server, AI agents (19.0), Odoo Pricing. Reviewed November 2026.