Odoo 20 AI Agents: What They Do and What They Cost

AI in ERP has mostly meant chat boxes that summarise things. Odoo 20 goes further. Its agents can take actions: create records, update them, send files and run as part of automations. That's useful. It also deserves more thought than turning on a chat feature.

What Odoo 20 agents can do

From the release notes, agents can:

• Create and update records, for example drafting a record from an uploaded PDF
• Send and upload files
• Generate images for websites and mailings
• Run from automated and scheduled actions, so an agent can be one step in a workflow
• Ask questions about a file while you preview it
• Show live progress, accept voice dictation, and ask for permission through buttons

There's also an AI provider selector with automatic model selection, default prompts with dynamic values, and model-bound agents tied to a specific record type. Agent "topics" have been renamed skills.

Stored agent conversations are kept for up to 30 days after the last message.

MCP: letting outside AI tools in

Odoo 20 adds an MCP connection to the database. MCP (Model Context Protocol) is the standard many AI assistants use to talk to external tools. In practice, it means an assistant outside Odoo can look up and work with Odoo records, within the user's access rights.

That's powerful for power users and a new thing for IT to govern. Decide who's allowed to connect, with which user, and log it.

The cost: IAP credits

This is the part people miss. Every AI feature in Odoo 20 requires IAP credits, bought on top of your subscription. Usage-based pricing is fair, but it's unpredictable until you've measured it.

Before rolling out:

1. Turn on AI for a small group first
2. Watch credit consumption for a few weeks
3. Set a budget and an owner for it
4. Expand to more users based on real numbers

Where we'd use agents

Good early candidates are repetitive, low-risk and easy to check:

• Drafting records from documents. Vendor bills, purchase orders or leads from PDFs and emails, with a human confirming before posting.
• Summarising and routing. Reading incoming requests and suggesting the right team or category.
• Content first drafts. Product descriptions, website copy, mailing text.
• Internal Q&A. "What's the status of order X?" from people who don't live in Odoo.

Where we'd be careful

• Anything that posts to accounting without a human approval step
• Stock adjustments and inventory valuation
• Customer-facing messages sent automatically without review
• Agents running on scheduled actions across large record sets, where one bad prompt affects thousands of records

A good rule: let agents prepare, let people approve. Odoo's permission prompts help, but design the workflow so the approval step is real, not a reflex click.

Governance checklist

• Named owner for AI usage and credit budget
• List of agents, what each can do, and which models they touch
• Agents tested with a restricted user, not an administrator
• Approval steps before financial or stock-affecting actions
• MCP connections limited to approved users and tools
• Review of agent output for the first few weeks

Our view

Odoo 20's agents are a real step forward, and for a lot of admin work they'll save time. Treat them like a new team member: give them clear jobs, limited permissions and someone checking their work at first.

For broader context, see AI in ERP and our AI ERP automation service. For everything else in the release, start with Odoo 20: what's new.