How to Set Up Your First AI Agent in Odoo 20

The safest way to start with Odoo 20's AI agents is also the most effective one: build a single agent, for a single, narrow job, and learn how the pieces fit together before you build a second. Odoo's own documentation gives the same advice — before creating an agent, have a clear use case in mind, because it drives everything else: the system prompt, the skills, and the sources. This guide walks through creating your first agent end to end, using only what Odoo documents for the AI app, and finishes with how to test, run, and supervise it.

Diagram showing the setup flow for a first Odoo 20 AI agent: system prompt, assigned skills, sources, testing, and automation

Before you start

Three prerequisites, all easy to miss:

The AI app must be installed. Many AI features in Odoo work without it, but creating and customizing agents requires the AI application.
You need IAP credits. In Odoo 20, all AI features require credits bought via in-app purchase, and no subscription plan includes them. Odoo does not publish a simple per-credit price for budgeting, so plan to measure your first agent's real consumption instead of guessing in advance.
Pick the use case first. “Answer questions about our support Knowledge articles” or “create leads from incoming briefs” are good first agents: bounded, easy to verify, and low-risk if the output is imperfect. Look at Odoo's preconfigured agents as worked examples of how a finished agent is put together.

Step 1: Create the agent

Navigate to the AI app and click New. Enter an Agent Name and, if you like, a short description of what the agent is for. In Odoo 20 the agent also gets an avatar based on its assigned role, which helps users tell agents apart in chat and in automation reviews.

Resist the urge to make this first agent general-purpose. A name like “Support Knowledge Assistant” or “Lead Intake Agent” forces the right thinking for the next steps: every later field on this form is about this job, not about AI in general.

Step 2: Choose a response style

Next, select a Response Style. This manages the overall tone the agent uses when responding. The documented options are:

Standard — a balanced, professional tone with clear, direct answers at a moderate level of detail. The sensible default for a first agent.
Nuanced — more sophisticated and contextualized responses that explore multiple perspectives; useful where the subject matter has real subtleties.
Snappy & Creative — punchy and engaging, prioritizing brevity and wit; better for outward-facing content than for operational work.
Slow & Rigorous — precision above all: the agent breaks down complex steps, cites its reasoning, and verifies claims; a good fit when correctness matters more than speed.

Step 3: Write the system prompt

The System Prompt is the agent's overall role, purpose, and behaviour — its core mission statement. It is applied continuously, and it influences how the agent interprets messages, selects tools, uses sources, and decides whether more action is needed before it responds.

For a first agent, a good system prompt covers four things in plain language:

Who the agent is — “You are the support knowledge assistant for our support team.”
What it is responsible for — the specific tasks from your use case, stated as outcomes.
The boundaries — what it must not do, what it should escalate to a human, and any rules it must always follow.
How it should answer — format, level of detail, and what to do when it does not know (say so, rather than improvising).

The documentation notes that a system prompt's specificity and length vary with the use case, and suggests the preconfigured agents as a guide. Write yours, then expect to revise it after testing — prompt refinement is normal, not a sign you did it wrong.

Step 4: Assign skills — this is where the agent gets its powers

Select one or more Skills for the agent. A skill is a collection of instructions and tools — the actual functions the agent can perform in Odoo, such as creating a lead or opening a view. The skills preconfigured in the AI app include View Builder, Information Retrieval, Create Records, Update Records, Generate Images, Web Search, and Website Builder; Create Leads is also available if the CRM app is installed. The tools available in your database vary with the apps you have installed.

One warning from the documentation deserves to be quoted in every setup guide: if an agent is not assigned any skills, it can only provide information — it cannot complete tasks or make changes to the database. That cuts both ways:

If your new agent politely answers questions but never does anything, missing skills are the first thing to check.
If your first agent only needs to answer questions from sources, you can deliberately assign no action skills and keep it read-only while you learn.

Start narrow. Each skill you add widens what the agent can touch, so give a first agent the smallest set that covers its one job. You can add skills later; un-scoping an agent that has already been acting broadly is harder.

Step 5: Add sources, and decide on Restrict to Sources

If the agent needs knowledge to work from, scroll to the Sources tab and click Add a source. Sources can be:

Uploaded files, such as PDFs;
Web links — paste the URL; multiple links can be added at once, one per line;
Documents from the Documents application;
Knowledge articles from the Knowledge app.

Sources are indexed so the agent can retrieve relevant parts when it works. After adding one, watch its Status field: a source is ready when the status moves from Processing to Indexed, and the Active toggle must be on for the source to count. (If a source fails to upload, the documentation notes it may be a conflict with the selected provider.)

Then make the governance decision: tick Restrict to Sources to limit the agent to only respond based on the resources you provided. For a support or policy agent, this is usually the right call — answers stay grounded in your documents. For a general internal helper, you might leave it off. Either way, choose deliberately.

Step 6: Test the agent before anyone else uses it

Click the Test button to open a conversation window and try the agent. Test like a sceptic:

Ask the questions a real user will ask, including the badly phrased ones.
If the agent has action skills, ask it to perform its task on a harmless example and check the record it created or updated.
Ask something outside its scope and confirm it declines or escalates the way your system prompt says it should.
Attach a file — click the + icon in the message window, then Attach Files — and ask about its contents. Files uploaded in a chat act as temporary sources for that conversation only.
If it also works from previewed documents, open a document in preview, click the AI icon at the top of the page, and try the default prompts or a custom question.

Expect to loop back to the system prompt and skills after the first round. That iteration is the real configuration work; the form itself takes minutes.

Step 7: Put it to work — and supervise it

Once the agent behaves, it can be used in chat — including by voice, since requests can be dictated — and, in Odoo 20, it can be summoned by automated and scheduled actions so it can work without anyone prompting it. That is where supervision matters most:

Review the runs. Check the automation chats to see how the agent handled automated tasks, and ask for changes based on any errors it faced.
Expect permission prompts. If an agent reaches its maximum tool call limit, it prompts the user for confirmation before proceeding; agents also surface questions and permission requests that you can answer by clicking buttons.
Keep sources current. If a source's content changes after indexing, reprocess it: from the agent's kanban card, open the menu, click Configuration, and on the Sources tab click Reprocess next to the changed source.
Remember the audit window. Conversations with agents are stored for up to 30 days after the last message — long enough to review a new agent's early work, not a permanent archive.
Watch the credits. Every run consumes IAP credits, and scheduled agents consume them without anyone watching a chat window. Check consumption like any metered supplier.

Common first-agent mistakes

No use case. A vague agent produces vague behaviour; the fix is upstream, in the job definition, not in more settings.
Skills too broad, too early. Adding Create Records and Update Records “just in case” gives a new agent write access it has not earned. Scope first, expand after testing.
Sources added but not Indexed or not Active. An agent cannot use a source that is still processing or toggled off — check the status before blaming the prompt.
Stale answers. If the underlying document changed, reprocess the source; agents answer from the indexed version, not from the file you edited an hour ago.
Skipping the test window. The Test button exists precisely so the first surprising behaviour happens in front of you, not in front of your customers.

Done this way, a first agent is a small, supervised experiment: one job, the minimum skills, grounded sources, and a human reading its early work. That is the foundation you want before you let Odoo 20's agents run on a schedule.