Skip to Content

Odoo AI in Practice: A Sober Look at What It Can Do Today

November 15, 2026 by
Odoo AI in Practice: A Sober Look at What It Can Do Today
Vitalii Lypovetskyi



About this analysis. All tests were run on 14–15 November 2026 on a demo database (Odoo 20, testai2.odoo.com) with fictional customers, products and users. No real company data was used. The results describe the current release only. Odoo's AI is moving fast, and some of the limits described here may disappear with the next update.

Why I finally sat down to test it

Odoo's built-in AI can now create contacts, quotations and opportunities from a plain-language request, and it does this well. But it still cannot finish a business process on its own, and it sometimes reports success where nothing happened.


For months I kept hearing two opposite claims: "AI will replace the ERP consultant" and "AI in ERP is just a toy". I finally made the time to check for myself.


This is not a product comparison. I did not set out to prove one tool better than another. I wanted to answer three practical questions:


  1. What can Odoo's AI really do today, in daily sales and stock work?
  2. Where is the boundary, and where is it likely to move next?
  3. What does this mean for a business, and is there still work for the analyst who sits between the system and the client?

What I focused on

The scenarios were not invented for this test. They come from my experience with client requests: the everyday routine of people who have just started working in Odoo. Entering contacts and products, preparing quotations, receiving goods, checking the numbers. Everything a new user normally does by hand, I tried to repeat through the AI.


I ran 12 scenarios that mirror an ordinary working week: a new contact from a business card, a product, a quotation in euros, a stock receipt, a new salesperson and a new deal, then reports, corrections and diagnostics. I wrote every request in deliberately broken English, the way a busy user types. Each scenario was checked against six questions.


Block

The question I asked

Understanding

Does it understand a short, careless request with spelling mistakes?

Creating and editing data

Can it create and fill records correctly from one request?

Finishing the process

Can it confirm an order, issue an invoice, send an email, apply a stock count?

Truthfulness

When it says "done", is it really done? Does it invent results or make silent mistakes?

Reports and filters

Can it answer a management question with correct numbers?

Diagnostics

Can it see errors, logs and settings, and check its own work?

What I found, block by block

1. Understanding: strong

Every request was full of mistakes: "creat", "headfones", "comercial ofer", "aprove". The AI understood all of them correctly, fixed product names ("Wireles Blutooth Headfones" became "Wireless Bluetooth Headphones Pro") and wrote clean, sales-ready English descriptions.


Example request: "Pls creat new product in odoo. Name is "Wireles Blutooth Headfones Pro", price 89 euro, cost 45 euro…"



Why this block came first: English is not my native language. I first wrote requests in my own language, but the built-in translation did not always carry the meaning correctly. Sometimes it hit, sometimes it missed, and checking every translation would have distorted the test. So I switched to English, and the results became stable. Even then, my wording and pronunciation are far from a native speaker's, and not every request was understood exactly as I meant it. It reminded me of the early days of Siri, when you spoke English to your phone and it heard something else. I expect this to improve quickly, as the underlying language models are already capable. For this test, I kept to English and checked how well the AI copes with imperfect, non-native requests.


I also tested recognition from an image. I took a photo of a business card and asked the AI to create a contact and a company from it. Everything was recognised correctly and filled in. This part clearly works: there is no longer any need for a separate text-recognition service, it is all built in. In a future analysis I plan to test harder documents: requests for quotation, supplier price lists, and incoming supplier orders or invoices that have to become sales orders and invoices in your own system. For now, on simple examples, I like the result a lot.

2. Creating and editing data: strong

Getting to a correct quotation took a few steps. First the products were created, in the base currency. When I asked for a quotation in euros, the system itself pointed out that the euro was not set up, and helped me activate it and add a euro pricelist. The quotation was first in złoty, then recalculated in euros. Later, converting euros back into złoty also required adding an exchange rate.


This took some time, but the AI prompted me at each step. I know the system and know what to ask; a new client usually does not, and does not expect that settings, currencies or reference lists need to be filled in first. Here the AI nudges you in the right direction, and that is a big plus.


Later, in one request it created a new salesperson, a client company with a contact person inside it, a quotation with two products in euros (the demo's base currency was PLN) with an export tax, and an opportunity at the right pipeline stage. The quotation total (3,270 €) and margin (1,650 €, 50%) were correct. Every change came with a confirmation card first.


Why euros: I deliberately worked in a currency other than the base one. I did not configure anything in advance: no extra currencies, no exchange rates. I simply registered a demo and started working, the way a new client does, to see how the system reacts when a foreign currency appears and no rates are set.



One weak spot is about thinking in processes. When the AI created the opportunity, it did not create the quotation inside it: the two existed separately, and the opportunity showed no quotations at all. In real sales work a quotation usually belongs to a lead or an opportunity. The AI does not yet see this as one sales flow: it does exactly what you ask and does not suggest the next logical step. It also did not copy the contact's email into the opportunity. Both were fixed only after a separate request. This is my observation rather than a bug.


Summary: the AI did everything I asked correctly, and even helped me along the way. For today, that is a high score from me. With clients I usually explain these features in much more detail, but what Odoo did for me, as a potential client, is impressive. It deserves attention and is definitely worth using.

3. Finishing the process: not yet possible

Most processes in Odoo run through a chain of linked documents: quotation, sales order, delivery, invoice, manufacturing order eth... Each of them has to be confirmed, created and sent, that is, processed in the system, so that it is recorded correctly and the business logic behind it starts. This is everyday routine, and I tried to repeat it through the AI.


The result: this is the biggest limit today. The AI could not confirm an order, create an invoice, send a quotation by email, cancel a document or apply a stock count. When asked directly, it explained that it can only read and edit fields, and has no tool to press Odoo's business buttons. It then gave step-by-step instructions to do it manually.



In my view, this is a current limitation on Odoo's side, and it will most likely be solved in future releases and updates. For now, though, it has to be taken into account: we are not fully autonomous yet. The AI prepares everything, but a person still has to go into the system and press Confirm, Create Invoice and Send.

4. Truthfulness

Anyone who has worked with AI agents knows they do not always answer correctly. They can hallucinate or simply make mistakes, and trusting them fully can become a problem. I wanted to check the same thing in Odoo.


The problem is there, and it is nothing new. When I asked for something, the AI reported that everything was fine; when I opened the record, the change had not happened. The AI never invented data, but four times it reported success that did not happen:


  • It said the stock was "added to the main warehouse". In fact it only recorded a count; the real stock stayed at 0.
  • It offered to cancel a quotation, and after I chose that option it said it could not.
  • It said it had cleared personal filters in the pipeline. The filter stayed, and the board was empty.
  • It said an amount was "updated to 12,000 PLN". The amount had not changed.



None of these were lies in the human sense. The AI described what it intended, not what the system did. But the effect for a user is the same: "done" does not always mean done.


A second, quieter example is a silent mistake. It is not a serious problem as such: the system makes such mistakes from time to time, I can see them when I check, but it does not flag them itself. I asked for an opportunity of "12,000 euro". It was saved as 12,000 złoty, the company's base currency, roughly four times less, with no warning. When I asked directly, the AI admitted the mistake and explained the cause correctly. It also discovered that no euro exchange rate existed in the system (1 EUR = 1 PLN). It did not suggest a fix until I gave a precise instruction ("set 1 EUR = 4.25 PLN"), which it then carried out perfectly.



Lesson: on a direct, precise instruction it is accurate. It does not yet think ahead about consequences.


I am more than sure this will be improved. I have met the same issue in AI integration projects, and the fix lies in the agent's logic on the server. But today it exists, and it has to be taken into account. For anything important, open the record and check it yourself straight away. Full trust is not possible yet.


My conclusion: this will not disappear until there are clear rules for how the AI must act in specific situations. A problem that shows up for one client in case A may not appear for another client in a slightly different case B, and that does not mean it is solved. It is solved only when we tell the AI: in this situation, act exactly like this. Set up rules for your own business, and trust will grow.

5. Reports and filters: a real strength

"Show me all quotations created this week with amount more than 100 euro" opened the actual Odoo list with the right filters applied. A report by salesperson (count, total, margin, margin %) was correct to the cent. For users who never learned to build filters and groupings, this alone saves real time.


One caution: the amount filter compares numbers, not currencies, so 100 złoty and 100 euro are treated as the same.


My conclusion: building filters and showing information on screen is a strong tool, and it really helps clients. Many people who meet the Odoo interface for the first time struggle with how to filter, group or build the right custom filter. Now the AI can do it for them. And when something is too complex for a filter, it shows the result as a table right inside the chat. That is already a good result, and this tool is worth working with.

6. Diagnostics: honest, but limited

Here my goal was to check whether I am still needed as an analyst when this tool is available. I gave it the kind of requests I usually start with when I connect to a client's system: find where an error happened and why, check the settings, compare, test and verify.


Asked to look for errors, it found two bounced emails (sent automatically to a fictional address) and linked each to the right document. It said clearly that it cannot see scheduled actions or server logs. It did not invent anything.


A side finding: the AI's own actions trigger real system emails. Assigning a salesperson automatically notified her. On a live database, colleagues and customers will receive messages started by AI.


The same check confirmed that the quotation I had "sent" earlier was never actually sent.


I consider this a good result. First, you can focus on the problems it found and work from there, the way any specialist does. Second, if you have no time or skills to dig in, you already know what to hand over to a specialist.


Where the boundary is, and where it is moving

Today the line is clear: the AI works with data (reading, creating, editing, filtering, calculating) but does not yet run processes (confirming, invoicing, sending). The pace is fast: version 19 brought the first AI features, version 20 already creates and edits data, and the remaining limits will most likely be solved in coming releases. Quality, speed and trust will grow, and built-in hints will help users understand the system as they work.

I expect Odoo's AI to stay focused on the user: everyday work, faster and with fewer clicks. Replacing the analyst is already technically possible, but even then the same issues remain: trust, truthfulness, rules. They are solved case by case. In our projects, for example, data passed from one system to another is checked at every stage; if something is wrong, agents fix it or call a person, and the data does not move on. That is why companies will build their own AI servers with their own rules, across several systems, not only Odoo. I have already built such solutions for clients, and I think this is the future.

Sooner or later we will rely less on specialists. For now, I am glad that is not yet the case.


Summary

For business owners

Odoo's AI is worth using today. It removes routine data entry, builds reports and filters on request, and guides new users through settings they did not know existed. Let it prepare the work, then check and confirm it yourself. Over time it will also make you more knowledgeable about your own system, the same way a business consultant does when working directly with a client.

For analysts

The routine part of the job is moving to AI. The value stays in what AI does not do reliably: process design, configuration and context (one missing exchange rate distorted a whole pipeline), integration with other systems, and responsibility for the result. AI will also make clients more knowledgeable, the same way a project does: they come back with better requests. The analyst is not being replaced; the work is moving one level up.

Odoo AI in Practice: A Sober Look at What It Can Do Today
Vitalii Lypovetskyi November 15, 2026