Odoo AI: the complete guide for SMEs (Ask AI, AI agents, AI fields)

Since Odoo 19, artificial intelligence is no longer a gadget hidden in a corner of the text editor. It has become a full-fledged app, with a conversational assistant available everywhere, configurable agents that work on your real data, and fields that fill themselves in. On paper, it is appealing. In the day-to-day life of an SME, the real question lies elsewhere: what actually works, what is just a sales demo, and where do you start without putting your data at risk?

As an Odoo integrator, we have been deploying these building blocks for our clients since they were released. This guide sorts through them, feature by feature, with the points to watch that the vendor will never highlight in its presentations.

What exactly does “Odoo AI” cover?

When people talk about Odoo AI, they often lump very different things together. Here is the actual map of the building blocks available in version 19:

An important point for your IT department: Odoo does not run its own large language model. When you configure an agent, you choose the underlying model (ChatGPT or Gemini, in several versions). Your prompts and the data sent therefore pass through these providers, something to factor into your thinking on confidentiality from the start.

Ask AI: the assistant your teams will actually use

This is the simplest entry point. Press Ctrl+K anywhere in the database and type your question in plain language: “show me unpaid invoices older than 60 days”, “summarize this thread”, “translate this message into English”, “suggest a polite follow-up for this quote”. The assistant understands the context of the screen you are on, and the AI button at the top right even offers preconfigured prompts tailored to the open app.

Once the answer is generated, you get three options that are useful every day: send it as an email, save it as an internal note in the chatter, or simply copy it. For a sales team that spends its day in the CRM, it is a real time saver on repetitive writing tasks.

The limit to know about: in its basic version, Ask AI does not modify your database. It will not create a lead and will not write to your records. And when a request goes beyond what it can do, it does not return an explicit error message, it politely declines. Your users need to know this so they do not jump to the conclusion that “the AI doesn't work”.

You in 2024, still writing your 30 customer follow-ups by hand instead of asking Ask AI.

AI agents: where it gets serious

The Odoo 19 AI app lets you create custom AI agents. An agent is an assistant you give four things: a role (the system prompt), topics (the instructions that define what it is allowed to do), sources (the documents it can draw from), and tools (the concrete actions it can perform, such as creating a lead or opening a view).

In practice, here is what this lets an SME build:

• A customer support agent connected to the website live chat, fed with your product pages, your FAQ and your Knowledge articles, able to answer recurring questions and create a ticket when it doesn't know.
• A pre-sales agent that qualifies website visitors and creates leads directly in the CRM, with the right information in the right fields.
• An internal documentation agent restricted to your quality procedures, which answers your teams' questions without them having to dig through the shared drive.

Two settings make all the difference in how reliable the result is. First, the response style: analytical (precise and consistent), balanced, or creative (more human but more variable). For anything related to support and data, stick with analytical. Second, the “Restrict to Sources” option: it prevents the agent from answering outside the documents you have provided. It is your best protection against made-up answers in front of a customer.

Deployment always follows the same logic: you add the sources (PDFs, web links, documents, Knowledge articles), wait for their status to change from “Processing” to “Indexed”, then use the Test button to challenge the agent before exposing it to anyone. An agent with no assigned topic can only provide information; this is intentional, and it is a good starting point: start read-only, then add the write tools.

“The underlying trend is massive: yet according to Gartner, over 40% of agentic AI projects will be canceled by the end of 2027, due to uncontrolled costs and unclear business value. In other words, the technology is ready, but poorly scoped projects are not.”

AI fields and OCR: AI that works while you sleep

a) AI fields

This is the most discreet building block and probably the most useful for business management. Via Studio (or as a property field), you create a field whose value is generated by AI: text, HTML, number, date, checkbox, and even links to other records or tags. The prompt can reference the other fields on the record with the /field command. A real example: a “sales pitch” field on the customer record, generated from the company name, its industry and the qualification notes.

The fields are filled in two ways: manually, by clicking the field's AI icon, or automatically through a scheduled action that runs once a day and completes any AI fields that are still empty. Enabled by default, this action is worth your administrator knowing about: it consumes AI calls on everything that is empty, including your old records.

b) Accounting OCR

Older than the generative AI wave, vendor bill digitization remains one of the fastest returns on investment in Odoo. You forward the PDF bill by email or drag and drop it, the engine extracts the vendor, amounts, taxes and lines, and pre-fills the accounting entry. Billing works with IAP credits (in-app purchases), so it is a pay-per-use cost, marginal compared with the data entry time saved.

Your accountant on Monday morning, when the weekend's 47 vendor bills have pre-filled themselves.

Limits and governance: what the vendor won't tell you

Let's be direct, because as an integrator that is our job. Odoo AI is well designed, but it does not escape the usual rules:

• The quality of the answers depends on the quality of your data. An agent connected to a database riddled with duplicates, empty product records and outdated documents will produce wrong answers with great confidence. Data cleanup remains the prerequisite, as on any ERP project.
• Your data passes through third-party models. ChatGPT or Gemini, depending on your configuration. For most SMEs this is acceptable, but it must be a conscious, documented decision, consistent with your confidentiality commitments to your customers.
• Generative AI is not deterministic. The same question can produce two different answers. So you never trust it with a tax calculation, a price or a contractual commitment without human validation.
• The cost is pay-per-use. IAP credits for OCR, usage tied to model calls for the rest. Nothing ruinous, but a cron job that fills AI fields on 50,000 records deserves a cost estimate before you enable it.

“Adoption, for its part, is already here: according to McKinsey's State of AI 2025 survey, 88% of organizations now use AI in at least one business function. The question is no longer whether your competitors are getting on board, but whether they are doing it properly.”

How do you deploy AI in your Odoo without getting burned?

Our method with clients comes down to four steps, and it deliberately looks like a classic Odoo implementation, because that is what it is:

• 1. Choose ONE measurable use case. Not “put AI everywhere”, but “cut invoice data entry time in half” or “answer the 20 most frequent live chat questions”. One goal, one metric.
• 2. Prepare the sources. An up-to-date FAQ, complete product records, clean procedures. That is 70% of the work, and it is the part people always skip.
• 3. Test in a test environment. Never in production. You build the agent, challenge it with the real trick questions from your business, and adjust the topics and sources.
• 4. Deploy read-only, then expand. An agent that informs first, and acts later. Plus a monthly review of conversations to correct course.

Should you wait? No. The Odoo 19 AI building blocks are mature enough for everyday use in an SME, provided you treat them as a process project and not as a box to tick. Those who structure their data and their first agents now will have a head start that is hard to catch up with.

Article written by Armand Meot

Odoo integrator at Hors du Commun (Lyon). I help industrial and service SMEs overhaul their information systems, secure their data flows and succeed with their ERP rollouts.