AI
Odoo and AI
What artificial intelligence already means for your ERP today, and where the marketing stops and practice begins.
The question is no longer whether AI gets a place in your business processes, but where it pays for itself. And you don't have to wait for a new Odoo version for that.
AI has gone from experiment to tool in a short time. In almost every process where text, documents, or patterns in data play a role, there's now an application that takes work off your hands. Invoices that read themselves in, deviations that stand out on their own, a question you ask in plain language instead of in a filter screen.
At the same time, there's a lot of noise. Every vendor now puts "AI" on its product sheet, and not every application pays for itself. The distinction isn't in the technology, but in the question of which concrete problem you solve with it, and whether you can measure the result.
Below is what's currently happening around Odoo and AI, what you can already use today, and where it goes wrong in practice.
What Odoo itself does: version 19
Odoo includes AI by default from version 19 of Odoo Enterprise. That makes artificial intelligence part of the product itself instead of something you bolt on separately, with the benefits that brings: one vendor, one license, support that comes along with the rest of the package.
What that means
Functionality that's built into the product, you don't have to build or maintain yourself. The improvement comes along with every upgrade. For standard applications, that's almost always the wisest route.
What it doesn't solve
It applies to Enterprise, not Community. It applies from version 19, while many organizations run on 16 or 17. And a version upgrade is a project in itself: testing, revisiting customizations, bringing users along. If waiting for version 19 means you do nothing for a year and a half, waiting is the most expensive option.
You don't have to wait: AI from Odoo Community 16
This is the point that usually gets buried in the discussion. AI in Odoo isn't a feature you switch on or off. It's a connection between your data and a language model. And that connection can simply be made from Odoo Community 16 onward. On older versions AI can also be connected, but the business case then quickly becomes weaker: you'd be investing in customization on a foundation that needs replacing anyway.
Odoo has a full-fledged API, an open data model, and an extensible architecture. Everything needed to feed documents, orders, relations, and procedures to a model is present. Whether that model comes from OpenAI, a European provider, or runs on your own infrastructure is a choice you make yourself. Not something the Odoo version decides.
That choice can also be steered through five levers: the AI model used, the cost per call, the region where processing takes place, whether or not your data is used to further train the model, and the speed the application needs. For invoice recognition you mainly weigh speed and price per document; for anything with personal or business-sensitive data you mainly weigh region and excluding training. We set those levers and document why, so it doesn't stay a loose technology choice but becomes a deliberate decision.
Why that's attractive
You start with the problem instead of the release schedule. You choose which model fits your requirements around privacy and cost. And you build experience with a well-defined application, so that when you later move to version 19 you know what you want, instead of discovering what the package happens to offer.
What to watch out for
What you build yourself, you maintain yourself. A connection needs attention at every Odoo upgrade and every change on the model's side. Keep the connection as thin as possible, therefore, and put the logic in Odoo, not in the integration.
Applications that already work today
No future visions, but things we encounter in practice that can be calculated.
Invoice recognition
Purchase invoices that read themselves in: supplier, amount, VAT, invoice number, and lines, linked to the right purchase order. This is the classic first step, because the volume is predictable and the time saved per invoice is directly measurable.
Fraud and anomaly detection
A model that knows your normal transaction pattern sees what deviates from it: a supplier with a changed account number, an amount that doesn't fit the history, an approval outside office hours. Not as a block, but as a signal for someone to review.
Searching in plain language
"Which retail customers ordered less this quarter than last year?", typed as a question instead of assembled from filters and groupings. This lowers the barrier to your own data enormously, especially for colleagues who aren't ERP specialists.
Agents on your own documentation and procedures
An assistant fed with your work instructions, quality manual, and internal agreements, that answers from them with source references. Especially valuable when onboarding new employees and for processes that occur rarely but must be exactly right.
Organizing and presenting data
Turning raw output into something usable: categorizing, summarizing, grouping, and writing a readable explanation. From an export file to a management summary without anyone losing an afternoon.
Pre-sorting email and tickets
Automatically sorting incoming messages by subject, urgency, and owner, including a draft reply for the common cases. For organizations with a service desk, this is usually the fastest-paying-off application.
We also see demand and inventory forecasting and cleaning up master data, duplicate contacts, inconsistent product data, as applications that pay for themselves remarkably fast, precisely because they tackle a problem that's been tolerated for years.
General helpdesk or specialist
There are roughly two ways to position AI in your organization, and the choice determines how much value you get from it.
As a general helpdesk
One assistant covering the whole system: answering questions, looking up data, explaining how something works. The barrier is low and everyone can use it. The downside is that such an assistant knows a bit about everything and isn't really good at anything, and that users drop off after two disappointing answers.
As a specialist in one knowledge domain
An assistant tailored to one domain, marketing for example: fed with your campaign history, target audiences, and tone of voice, able to assemble segments from your CRM, draft copy, and analyze what did and didn't work before. Narrower in scope, but with a noticeably better result.
Our advice: start specific
A narrow assistant that excels convinces faster than a broad one that's mediocre, and it's also measurable: you know exactly which process you wanted to improve and whether that succeeded. The support that creates is what makes a broader rollout possible later.
Where it goes wrong in practice
The projects that stall rarely stall on the technology.
Messy data
AI on unreliable data produces nonsense presented convincingly, and that's more dangerous than an error message, because nobody notices. Duplicate contacts and half-filled fields aren't a detail you fix later, but a prerequisite up front.
Where does your data go
As soon as you send company data to an external model, that data leaves your environment. That calls for a data processing agreement, clarity on whether your input is used for training, and a deliberate choice about hosting within the EU. For sensitive data, a model on your own infrastructure is a real option.
No human left in the loop
Anything with financial or legal consequences needs a human check. A model that prepares invoices is a gain; a model that posts invoices unseen is a liability question. Establish in advance who approves what.
AI as a replacement for process design
The most persistent misconception. A bad process you automate is a bad process that runs faster. AI is at its best when it sits on top of a process that already works. Not when it has to hide that the process doesn't.
Start small, and measure it
Choose one process where the pain is demonstrable and the volume predictable, put a well-defined application on it, and calculate what it delivers. That works better than a broad rollout, and it gives you the insight you need to decide what the next step should be.
We're happy to think along about which process lends itself to this within your Odoo environment, whether you run on Community 16 or are considering the move to 19.
Schedule a strategy session →