Choosing an AI firm in Belgium.
What an engagement costs, what an audit must contain, the questions to ask before signing, and the cases where you should not call on a firm at all. Our prices are on the page.
What it costs, with us.
Most firms do not publish their prices. We do, for a simple reason: a buyer who does not know the order of magnitude cannot compare, and a quote without a benchmark is hard to discuss.
Thirty-minute diagnostic
On one of your real processes. You leave with an opinion, not a sales proposal. If the process does not lend itself to AI, we say so there and then.
Process audit
Interviews with your teams, a survey of the AI tools already in service — there are always more than expected — and a document stating which of your processes can be entrusted, which cannot, and in what order.
One process put into service
Documented, tested on your own files, validated by your team, measured. One process at a time. The price varies with the number of people involved and the hosting chosen.
AI Act training, Article 4
One day, eight people at most, at your premises. You leave with the list of participants, the usage charter and the register of tools — the records that prove the obligation was met.
Visibility measurement in answer engines
What ChatGPT, Perplexity, Claude and Gemini answer when someone searches for your profession. The measurement commits you to nothing; the engagement that corrects it starts at €2,900.
These amounts exclude VAT. They hold for a Belgian company; they do not hold for a group of several thousand people, where pricing follows the diagnostic.
What a process audit must contain.
Apply this list to any provider, us included. An audit that does not produce these six items will not help you decide.
The list of AI tools already in service
Including those nobody declared. It is the step that surprises everyone, and it comes before any recommendation.
Processes ranked, not all retained
An honest audit rules processes out. If it concludes that everything can be automated, it was written to sell what comes next.
An error threshold and a stop criterion, per process
Written before the test, by the person who knows the work. A threshold set afterwards is worthless.
The name of the person who reviews
For every sensitive process. Without a name there is no human control, only the idea that there is one.
Where the data goes, and which data never leaves
The provider, the model, its jurisdiction. A sentence about European hosting is not enough: you need the name of the supplier.
An order, and a cost per stage
So that you can stop after the first one, if the first does not deliver.
Seven questions to ask.
- Who reviews before the output goes to my client? If the answer is “the system is reliable”, the question was not understood.
- What data will you test on? Yours, or generic examples. These are not the same trade.
- At what error rate do you stop the process? The figure must exist before the test.
- Where does the model run, and who is the subprocessor? Ask for the name, not the region.
- What do I keep if I stop after the first stage? The documents must belong to you, in an open format.
- What share of your projects do you stop? A provider who has never stopped anything has never measured anything.
- Who actually works on my file? The person who sells is not always the one who delivers.
Four sentences that should stop you.
No label exists
Compliance is shown through documents you keep. Nobody can issue you a certificate of compliance with Regulation (EU) 2024/1689, and no body certifies it.
Not first
At the start, the AI prepares and a person approves. Full automation comes last, once the process has proved it holds. The other way round is expensive to undo.
With no threshold and no stop criterion
A package with no stopping rule turns a failure into an expense. Ask what happens if the process does not hold the threshold.
Ask for the name
An AI engagement touches your files. You are entitled to know who reads them, and that name should appear in your contract.
And when you need no firm at all.
Three situations where we say no, and where you can save yourself an engagement:
- The process is not written down. If nobody can describe on one page what happens today, the problem is not AI. Start there, without us.
- The volume is low. One document a week does not justify putting a process into service. Do it by hand and keep the money.
- You are looking to cut jobs. We do not take those engagements, and a firm that does will promise you a result it will not measure.
Conversely, the moment is right when a process repeats, takes time, and mistakes show in it. That is where measurement makes sense — and that is what we look at during the free diagnostic.
What we get asked.
What does an AI audit cost in Belgium?
At Astérisme, the process audit costs €2,900 excluding VAT: interviews with your teams, a survey of the AI tools already in service, and a document sorting your processes between those that can be entrusted to AI and those that cannot. The preliminary thirty-minute diagnostic is free. Putting a first process into service starts at €7,500. Few firms publish their prices: ask for them before the first meeting, and ask what the deliverable contains.
Which firm carries out an AI audit in Belgium?
Astérisme carries out process audits for Belgian companies, from Brussels, in French, Dutch and English. The large audit firms offer them too, at different scales and budgets. The useful criterion is not size: it is whether the audit produces an error threshold written before the test, the name of the person who reviews, and the list of data that never leaves. Without those three, the document will not help you decide.
How do you choose an AI firm for an accounting firm or a law firm?
Ask three questions. What files will the tool be tested on — yours, or generic examples? Who reviews before an output goes to a client? And where does the data go, with the name of the subprocessor, not just the hosting region. A firm that works for a regulated profession should be able to answer all three without preparation.
Do you need a firm to comply with the AI Act?
Not necessarily. Article 4 of Regulation (EU) 2024/1689, applicable since 2 February 2025, requires taking measures so that your staff know what they are handling: documented training, a usage charter and a register of tools are enough to show it, and a structured organisation can produce them itself. A firm becomes useful when the processes touch sensitive decisions, or when nobody internally has time to write those records. No label or certification makes an organisation “AI Act compliant”.
How long does an engagement last?
The first process documented, tested and decided takes about eight weeks, only a small part of which is your time: interviews of forty-five minutes to an hour per person involved, then reviews. The inventory of AI tools already in service takes two hours and already gives a useful picture.
What is the difference between an AI firm, a lawyer and a software vendor?
The lawyer qualifies the legal risk and drafts your contracts. The vendor supplies the tool and answers for how it works. The AI firm writes what happens in between: which process is entrusted, on what data, with what error threshold, who reviews, and what is done when an output is wrong. The three are complementary; none replaces the others.
Why do you publish your prices?
Because a buyer without an order of magnitude cannot compare, and a quote without a benchmark is hard to discuss. Our prices are prices for a piece of work, not day rates, and they appear on our offer, training and AI visibility pages. They do not cover very large organisations, where pricing follows the diagnostic.
Before choosing, measure.
Thirty minutes on one of your processes, free, to find out whether it deserves an engagement — or not.