Odoo 20 is the release where AI stopped being a button in the editor. In Enterprise it reads your records, fills fields, runs inside automation rules, answers customers in live chat and files your vendor bills. Release notes say all of that. What they do not say is how often it is right, where your data goes when it works, and what each answer costs.
So we ran it. Two Odoo Online databases on 20.0 Enterprise, between 22 and 29 October 2026, every answer checked against the database it came from. This is the inventory: what ships, what we measured, and where it needs a human to check the result.
This address used to hold our Odoo 19 Community teardown, “five AI features, four of them phone home”. Odoo 20 put its new AI (agents, AI fields, MCP) in Enterprise only, and this publication now covers Enterprise. The Community findings, re-run on 20.0, are in the Odoo AI Feature Map.
Nearly every AI feature in Odoo 20 Enterprise goes through one service,
Odoo AI, billed in In-App Purchase credits. On both of our databases the model
behind it was Google's gemini-3.8-flash, chosen by Odoo, not by
us. The AI respects each user's access rights. It was right on most of what we
checked, and wrong in ways that look exactly like being right. One thing still
runs entirely inside your database: predictive lead scoring.
What ships, and what we measured
Ask AI: questions about your own data
A button in the top bar opens a chat window. Ask a question in plain English and the agent searches your database: on our runs, between ten and thirteen tool calls per question (load a skill, read field definitions, search, group). Asked how many contacts the database held, it named the three that exist. Asked for revenue by salesperson, it got nine totals out of ten exactly right. The tenth was off by $1,000, presented with the same confidence as the others.
Asked for a pie chart of sales by customer country, it opened the right report and added a filter nobody asked for: order date in the last 365 days. On our data the result did not change. On yours it might.
AI fields
Any field added with Studio can be ticked Autofill and given a prompt; Odoo fills it from the other fields of the record. We created a “main need” selection on opportunities and ran it on eight of them. Six matched what a person would have picked. A field created through code, without the AI widget, shows no button at all. A prompt that does not insert the record’s fields returns “unknown”, because the model only sees what the prompt hands it.
AI inside automation rules
Automation rules gain an action of type AI: the rule hands the record to the model, which decides whether to call a tool. One detail cost us an afternoon: the tool is only ever called if its Use in AI box is ticked. Without it, every run logs “No tool was executed” and no error. With it, two urgent leads out of six were raised to top priority, and every ambiguous one was left alone. That is the behaviour you want, and the one a test with only clear cases would never show.
Agents with their own sources
You can build an agent, give it instructions and attach sources: a file, a link, a Knowledge article. We gave one a five-question FAQ about deliveries and returns and asked it about a warranty the FAQ never mentions. With short instructions it invented one, two years, in a fluent sentence. With Limit answers to sources switched on, it said the sources did not cover it. The switch is off by default.
Helpdesk and live chat
Helpdesk teams get an AI option that finds similar past tickets, off until
someone enables it on each team. Similarity runs on embeddings
(gemini-embedding-2) computed by a daily scheduled job, so a team
switched on this morning finds nothing until that job has run. A live chat
channel can hand visitors to an agent instead of a scripted bot; asked a
question its sources did not answer, ours said so and offered a callback.
Documents and vendor bills
We dropped a fictitious vendor bill into the Documents inbox. In under eight seconds it was renamed, moved to the purchases folder and turned into a draft bill. Supplier, reference, date and totals were exact. The three invoice lines became one “Expenses” line, and the due date followed the supplier’s default terms rather than the 30 days printed on the PDF. Those fields are filled by Odoo’s document digitization service, billed separately, not by the agent that sorted the file.
The MCP server
Odoo 20 exposes /mcp, so an outside assistant (Claude, ChatGPT,
a coding agent) can query the database with an API key scoped to MCP. By
default the tools are read-only, and the client sees exactly what its user is
allowed to see: our test user, restricted to its own documents, saw 13 records
out of 62. In that mode Odoo’s own AI is not involved: no AI session is
recorded on the Odoo side, and the reasoning happens in the outside model.
It respects access rights, literally
We asked the same question from two accounts: how many opportunities does the CRM hold, across all teams? The administrator got 17, worth $320,200, which is exact. A salesperson restricted to their own documents got zero, and was told that none had been converted yet. Zero was correct for what that user can see; the sentence around it was not. The AI does not leak data across access rights. It also does not tell you when the answer is limited by them.
Where your data goes
Odoo’s IAP privacy page states that “the AI provider is selected by Odoo automatically for Odoo v20+” and lists OpenAI and Google (Gemini). It says what travels: your prompt, the relevant conversation, the action you ask for and context from the record or screen you are working on. It also says that “information submitted through these providers is not stored and not used to train AI models”, that Odoo’s IAP service may store the conversation temporarily, and that inference residency is global.
What we measured is narrower and more useful: on both databases, on every
session we inspected, the provider recorded by Odoo was Google, model
gemini-3.8-flash. We found no setting in 20.0 to pick another one.
If your compliance review needs a named processor, that is the name today, and
Odoo can change it.
What it costs
Odoo AI draws on In-App Purchase credits, attached to one database. The service page announces 20 free credits. After 56 AI sessions of testing, our balance read 17.0066. Measured per action:
- an AI field filled on one record: 0.0021 credit;
- a question answered from an agent’s sources: 0.0054;
- an agent triggered by an automation rule: 0.0075;
- an analytical Ask AI question with a dozen tool calls: 0.041.
Two traps. The balance shown through the API is not stored and goes stale; only Settings › View my services refreshes it. And the balance moved again after our tests had stopped, so measure after it settles. On pricing, Odoo lists agentic AI under its Custom plan: check which plan you are on before you enable the AI app.
The one that stays
Predictive lead scoring is still in crm, and it still makes no
network call: a frequency table updated by every won or lost opportunity. Odoo
20 now shows an AI icon next to the probability field, which makes it easy to
mistake for Odoo AI. It is not. It is local, free, and switched on the day CRM
is installed, with a start date set eight days before installation and seven
fields chosen for you, country and language among them. Worth a decision rather
than a default.
How to check this on your own database
Every Ask AI conversation is stored as a session, and each model response records its provider. A read-only query over XML-RPC is enough:
events = models.execute_kw(DB, uid, KEY, 'ai.session.event', 'search_read',
[[]], {'fields': ['metadata'], 'order': 'id desc', 'limit': 50})
{(e['metadata'] or {}).get('provider_metadata', {}).get('model')
for e in events} - {None}
# on our databases, October 2026: {'gemini-3.8-flash'}
Run it after asking one question in the top bar. If the set changes, the processor in your privacy notice changes with it.
What this changes in practice
Three things, none of them “do not use it”.
Check the answers that matter. The AI was right far more often than wrong, and its mistakes were not flagged: a total off by $1,000, an extra date filter, an invented warranty. Before a number from Ask AI reaches a customer or a board, check it against the report it came from.
Switch on the guard rails yourself. Limit answers to sources, Use in AI on each tool, per-team activation in Helpdesk: the safe settings exist, and several are off by default.
Name the processor. Your staff’s questions and the records they are looking at go to a model chosen by Odoo. Say so in your privacy notice, with the provider you measured.
To reproduce any of this, an Odoo Enterprise trial includes the AI app. The method behind each result, step by step, is what our AI Playbooks for Odoo 20 are for.