Most ERP “AI” features of the last few years were really just drafting assistants. They would summarize a thread, suggest a reply, or rephrase a paragraph, and then a human still had to do the actual work in the database. Odoo 20, unveiled on 24 September 2026 at Odoo Experience in Brussels, takes a different position. Its AI agents do not stop at suggestions: they create records, update records, work from files you hand them, and run inside automations when nobody is chatting with them at all. That shift — from assistant to agent — is the headline AI story of the release.
What an AI agent actually is
Odoo's documentation defines an AI agent as a smart virtual assistant that can understand natural language, perform tasks, and assist users by interacting with Odoo tools. Strip away the marketing and an agent is a configured object with four parts:
- A purpose and a system prompt. The system prompt is the agent's mission statement: its role, responsibilities, and behaviour. It is applied continuously and drives how the agent interprets requests, selects tools, and decides whether more action is needed before answering.
- Skills. A skill is a collection of instructions and tools that defines what the agent can do in a specific context. Tools are the actual functions the agent can perform in Odoo — creating a lead or opening a view, for example — and they are what turn conversation into real action.
- Sources. The information the agent can draw on: PDFs, web links, documents from the Documents app, and articles from the Knowledge app. Sources are indexed so the agent can retrieve relevant passages.
- An identity. In Odoo 20, agents get avatars based on their assigned role, which makes it easier to tell at a glance which agent you are talking to, or which one acted inside an automation.
One practical point: many AI features in Odoo work without the AI app, but creating and customizing agents requires the AI application to be installed.
What agents can do
The Odoo 20 release notes list concrete capabilities that, taken together, describe a worker rather than a chatbot:
- Create records. You can ask an agent to create anything for you — including by uploading a file of instructions, such as a PDF, and having the brief become records in the database.
- Update records. Agents can be asked to update existing records, not just create new ones.
- Answer questions about a file. While previewing a document, you can ask questions about it; the agent's responses are grounded in the file being previewed, combined with any sources assigned to it.
- Work with files in chat. Files can be uploaded and documents linked inside agent chats; uploaded files act as temporary sources for that conversation, so the agent can summarize or extract details from them.
- Generate images. Agents can generate images and enriched buttons for websites and mailings.
- Build views with time periods. When building views, agents can filter by time period — weeks, quarters, and so on — which makes requests like “show me last quarter's” workable in natural language.
- Take voice instructions. Requests can be dictated directly to an agent instead of typed.
Two supporting details make this usable day to day: agents show specific, live updates on what they are doing — analyzing the request, consulting sources, preparing an action — instead of a generic “AI is thinking” message; and when an agent mentions supported records in a live chat, the chat can show rich, clickable cards, so you can jump straight to what it created or changed.
Skills: the difference between answering and acting
The single most important sentence in Odoo's agent documentation is this one: if an agent is not assigned any skills, it is only able to provide information. It cannot complete tasks or make changes to the database. Autonomy in Odoo 20 is therefore not a default — it is something you hand to an agent one skill at a time.
The AI app ships with preconfigured skills, including:
- View Builder — interprets a natural-language query and opens the appropriate view, including time-based grouping such as by day, month, or year.
- Information Retrieval — tools to retrieve information about the models in your database.
- Create Records and Update Records — the tools that let an agent write to the database.
- Generate Images — loaded when the user asks to generate or enhance an image.
- Web Search — lets the agent explore the internet in real time.
- Website Builder — tools for editing Odoo website pages.
- Create Leads — automated lead creation, available only when the CRM app is installed.
The tools available in a database vary with the apps installed, and several tools can be assigned to a single skill. Each skill also carries its own instructions — covering its purpose, rules, and step-by-step workflows — so two agents with the same skill can still behave differently.
If the Restrict to Sources option is enabled on an agent, it can only use information from its uploaded, active sources. For a support agent answering from your Knowledge articles, that restriction is the difference between grounded answers and improvisation. And when a source changes, you can ask the agent to reprocess it so it uses the most up-to-date version.
Agents inside automation: where Odoo 20 really changes
Odoo 19 already had configurable agents you could chat with. Odoo 20's change is where agents are allowed to run. The release notes call it agentic automation:
- Agents can be summoned by automated and scheduled actions. An automation rule or a schedule can call an agent the way it would call any other step in a process — no human needs to open a chat.
- Runs are reviewable. You can check the automation chats to see how an agent handled a task, and ask for changes based on the errors it ran into.
- Agents can update themselves to best serve their use cases, according to the release notes — a good reason to review automation chats regularly.
- Agents can become model defaults. When an agent is asked to interact with a specific model, it suggests becoming the default agent of that model, so future work on that model routes to it.
This is the “software that does the work” part of the story: a chat assistant waits for a person, while an agent inside a scheduled action does not. The trade-off is that unattended work needs the same governance as any automated process — narrow skills, good sources, and a habit of reading the run logs.
The guardrails Odoo built in
Because agents write to the database, the release notes also describe the checks around them:
- Tool call limits with confirmation. If an agent reaches its maximum tool call limit, it prompts the user for confirmation before proceeding — long jobs stop and ask instead of continuing.
- Interactive responses. You can respond to an agent's questions, permission requests, and suggestions by clicking buttons, rather than typing replies — approvals stay quick and explicit.
- Stored conversations. Conversations with agents are stored for up to 30 days after the last message was sent, giving you a window to audit what was asked and what was done.
Underneath all of this, the normal rules of Odoo still apply: an agent works with the tools its skills provide. Scoping an agent well — the right skills, the right sources, restricted sources where accuracy matters — is the main safety decision, and it is made at configuration time, not at run time.
The meter is running: IAP credits
There is a cost dimension that earlier Odoo AI features did not force you to think about. The release notes are blunt: all AI features now need credits, which can be bought via IAP (in-app purchase). No Odoo plan includes IAP credits, so AI usage is a separate, usage-based cost on top of your subscription — and Odoo does not publish a simple per-credit price list to budget against, so treat agent work like any metered service: estimate the volume per agent and per workflow, then watch actual consumption once you start.
One related choice has been simplified: you select an AI provider, and Odoo automatically uses the best AI model for the task at hand. Default prompts can also carry context and dynamic values, keeping recurring agent tasks consistent without retyping instructions.
Agents in the wider Odoo 20 AI picture
Agents are the centre of Odoo 20's AI story, but not the whole of it. The same release adds an AI accounting assistant that answers questions and produces insights based solely on the accounting data in your database; an AI website assistant that creates pages, edits snippets, adds images, and builds forms and menus; and a native MCP (Model Context Protocol) connector, which lets an external AI system connect to your Odoo database and work within the user's existing access rights — in effect, bringing outside AI in rather than running Odoo's own.
The through-line is consistent: in Odoo 20, AI is expected to produce outcomes in the database, not just text about them. Whether you use it through chat, automations, or an external tool connected by MCP, the configuration questions are the same — what can it do, what can it see, and how will you review what it did. Get those three answers right, and an agent stops being a novelty and starts being staff.
Further reading
- Odoo 20 Release Notes — Odoo (see the AI section)
- AI agents — Odoo documentation
