Custom AI solutions: which one fits your question?
The question is rarely "could we do something with AI". The question is which form fits, what it costs, and when it is not worth it. This page answers that in two questions.
What does an AI solution on Odoo cost?
A focused AI solution on top of Odoo falls into the same cost bands as any other custom work: a single automation usually stays under EUR 4,000, a focused solution with a few screens and a clean integration costs EUR 4,000 to EUR 15,000 and is live in one to three weeks, and a build across departments or systems starts at EUR 15,000 and runs in phases. What drives the price is rarely the AI part but where your data sits: everything inside Odoo is considerably cheaper than data that has to come from three systems and a folder of documents.
What happens when someone says: we want to do something with AI
Almost every conversation starts the same way. Someone saw a demo, or the board asked "what are we actually doing with AI", and a question lands on the table that sounds like a brief but is not one. We tend to ask an awkward question back: what work would you like to see disappear tomorrow? That usually produces a silence, and then the real answer.
More often than not the problem turns out to be smaller and more concrete than the word AI suggests. Someone exports three reports to Excel every Monday to pull out a single number. The inside sales team answers the same five questions about lead times, every day. Two hundred invoices a month arrive and get retyped by hand. Those are not AI problems, they are workflows with a hole in them. That the hole happens to be fillable with a language model is an implementation detail.
Sometimes the answer is bigger than expected instead. Then it turns out people are not waiting on a report but on a decision, and the report was only the excuse. In that case an assistant that fetches numbers faster does not help, and you have to look a level up. That is also a good outcome for a first conversation, even though it earns us no project.
What we are really doing in that first session is reducing the question to something you can build and measure. Not "AI for the sales department", but "the quote request that currently gets retyped four times goes through correctly in one pass". Small enough to build in weeks, large enough to notice.
Why the AI part is rarely the expensive bit
The expectation is nearly always that the model makes it expensive. In practice it is the cheapest component. Calling a language model costs cents per question, and the logic around it is faster to write than ever with the current generation of development tools. Where the money goes is the layer underneath: the data.
If everything sits in Odoo, it is simple. We read from the same model your sales, stock and accounting already live in, with the permissions you already configured. No synchronisation, no intermediate store, no second version of the truth. That is exactly why an ERP that is open enough saves so much, and why this is not possible with many other packages, or only at considerable cost.
If the data sits outside it, the work changes character. First there has to be a reliable way in: an integration that does not fall over, field names that stay consistent, an agreement about what happens when two systems claim different things. That is classic integration work, and it is the reason the same functionality is live in two weeks one time and becomes a phased project of months the next.
That is why the decision tree above does not ask about your budget but about where your data lives. That single answer predicts the price better than any feature wish. And it is also the most honest way of saying: if your data is scattered and messy, do not buy AI. Clean up first. It is cheaper, and you benefit from it either way.
Which AI solution fits your question?
Two questions. The first decides the form, the second mostly decides the price. No form, no email: you see the outcome immediately.
What do you want to solve?
The form follows from the problem, not from the technology. Pick whatever comes closest.
Where does the data for that live?
This is the biggest cost driver. The AI part is rarely the expensive bit; getting data out of separate systems is.
The five forms, and when each fits
Almost every AI question that reaches us falls into one of these five. Each one also says when it is the wrong choice.
-
AI assistant on your own data
An assistant that answers questions in plain language by reading your own Odoo data, with the source attached so you can check the answer.
+ When people wait on a report, or when the same question is worked out by hand every month.
- Not when what you actually lack is a good dashboard. A fixed report is cheaper and more reliable than an assistant reinventing it every time.
-
Answer engine with sources
A search and answer layer that draws on your own content and documents, can compare things, and always shows where the answer came from.
+ For a knowledge base, product documentation or support that keeps getting the same fifty questions.
- Not when your documentation itself is outdated or contradictory. Then you build a machine that gives the wrong answer with authority.
-
AI automation inside the process
A step in your workflow that reads documents, extracts data, or classifies and forwards something, without anyone retyping it.
+ For incoming invoices, applications, forms or emails that are moved across by hand today.
- Not at low volumes. If it happens ten times a month, doing it ten times is cheaper.
-
Custom app as a shell around Odoo
Your own application for the process that does not fit the standard, running on the same Odoo data rather than beside it.
+ For planning, field work, grant management or an industry-specific process that off-the-shelf software does not match.
- Not when the standard almost does it. Then adapting your process is nearly always cheaper than maintaining an app.
-
Configurator or calculator
A tool that guides the visitor through a choice or calculation and ends with something you can use: a configuration, an estimate, a request.
+ For a composed product, a price driven by many variables, or a question that otherwise ends in a quoting round.
- Not when your offer fits in one sentence. Then a tool is a barrier rather than a help.
What a project actually looks like
We never start by building. The first step is a half-day session in which we draw the process as it really runs today, including the detours nobody writes down. Something always surfaces that nobody had mentioned: an intermediate step in Excel, an email that counts as approval, an exception for two customers. Those are precisely the things that decide whether a solution survives contact with reality.
After that we build the smallest piece that is already worth something. Not a prototype you can only look at, but something running on your own data that your colleagues can use. Usually that stands within a week or two. That is deliberate: your judgement about whether this works is far better once you have used it than once you have seen it.
Then comes the phase that rarely appears in a quote but makes the difference: adjusting on the basis of real use. The first version gives answers that are correct but not in your language. Or it is too cautious, or too confident. That shaping often takes as long as the building, and anyone who does not plan for it ends up with a solution that is technically finished and humanly not trusted.
Finally we agree who owns it. An AI solution nobody reviews slowly turns into something people work around. One person who spot-checks now and then and is allowed to say "this is wrong" is enough, but without that person it is only a matter of time.
How ElizaKnows came about
ElizaKnows did not start as a product but as an irritation. At one client the same question came round every Monday morning: how are we doing. The answer was in Odoo, but it took someone an hour to get it out, and by the time it circulated it was already a day old. There were dashboards, but nobody looked at them, because the question was slightly different every week.
We did not build another dashboard. We built a layer that takes the question in plain language, works out for itself which data belongs to it, and returns the answer with the underlying records attached, so you can check where it came from. That last part turned out to matter more than we expected: without those source records nobody trusted the number, and with them the discussion suddenly became about the business instead of about the report.
What we learned along the way is that the difficulty is not in answering but in refusing. A model that always says something is dangerous. It has to be able to say that the question cannot be answered with the data available, or that the answer becomes unreliable because an entry is missing somewhere. We built that behaviour in explicitly, and it is the reason we dare to put it somewhere decisions are made.
It now runs at several clients and we use the same underlying idea for other questions. But the starting point has not changed: not AI because it is possible, but because an hour disappeared every Monday.
When AI is not the answer
We would rather say this up front. In these cases we advise against it, even when we could build it:
- The process is not defined yet. AI on an unclear process only makes the confusion faster.
- The data is wrong. An assistant on messy data gives the wrong answer with great confidence.
- It happens too rarely. Below a few times a week, automating rarely pays back.
- There is no owner. Without someone judging whether the output is right, nobody will trust it.
- The standard almost does it. Odoo now has AI features of its own; checking what you already have is free.
None of these is final. They are simply cheaper to solve before you build than after.
What we already built
Updoo is our collection of shells around Odoo. Not demo environments but things running at clients. They are here to show that we build this ourselves.
Frequently asked questions about AI on Odoo
What does it cost to have an AI agent built?
A focused AI solution on top of Odoo usually costs EUR 4,000 to EUR 15,000 and is live in one to three weeks. A single automation stays below that; a build across several departments or systems starts at EUR 15,000 and runs in phases. The largest cost is rarely the AI part, but unlocking data that sits outside Odoo.
What exactly is an AI agent?
An AI agent is software that understands a question in plain language, works out for itself which data or actions it needs, and carries those steps out. The difference with a chatbot is that an agent does not only answer but can also look something up, prepare it, or push it through in your system.
Can AI reach my own Odoo data without that data leaking?
Yes. The data stays in your own Odoo environment and only what is needed to answer the question goes to the language model. Per solution we record which fields may be reached and who may see which answers, so the authorisations from Odoo stay intact.
Does Odoo already have AI features of its own?
Yes, and that is the first thing we check. Odoo now includes features for invoice recognition and in-app assistance among others. Custom work is only worthwhile once the standard does not cover your question, for instance because it concerns your own process or data outside Odoo.
How long before something like this runs?
A single automation is often live within days. A focused solution with its own screens and an integration takes one to three weeks. A build across several systems runs in phases of several weeks. In every case we deliver a small working part first, so you can judge whether it returns what you hoped.
Not sure which form fits?
Put your question to someone who has built this before. We will say honestly if the standard already does it, or if it will not return enough.
Odoo Gold Partner · Amsterdam · we build this ourselves, no reselling