One distinction determines whether AI is useful or dangerous in your firm: the difference between language and numbers. AI is remarkably good at the first and fundamentally unreliable at the second.
Why AI is bad at arithmetic
A language model predicts the most likely next piece of text. That is a stunningly powerful principle for language — and a poor principle for arithmetic. Faced with “21% of 1,847.50”, the model does not produce the result of a calculation, but the number that most looks like a plausible answer.
Most of the time it is right. Sometimes not. And the difference is invisible from the outside, because a wrong answer arrives with exactly as much confidence as a right one. In a sector where a discrepancy of a few euros makes a return wobble, “usually right” is not a usable standard.
The rule I apply myself: anything that is an amount, a percentage or a balance is calculated in code that always gives the same answer. AI may read and phrase, not calculate and not decide.
What AI is genuinely good at
Reading documents that look different every time
This is the strongest application in an accounting firm. Classic OCR software needs to know where on the page an amount sits. A language model understands that “payable before 15/09” is a due date, whether it sits in a box at the bottom right for one supplier or at the top for another.
Important: the model reads the amount from the invoice, it does not calculate it. That is precisely the permitted use.
Recognising messy descriptions
“PAYM INV 2024-0871 + 0872 LESS CN” on a bank statement. A human reads straight away that this concerns two invoices with a credit note in between. Fixed rules stumble over this; a language model does not. The amounts themselves are then simply recalculated.
Writing a first version of a text
The commentary on a monthly report, a chasing email, a summary of what stands out in the figures. AI delivers a usable starting point that you polish in two minutes rather than write in fifteen. As long as you read it back, the risk is low.
Searching your own documents
“Which file had that clause about the company car?” AI is excellent at retrieving passages based on meaning rather than exact keywords.
Where it gets genuinely dangerous
Having tax or legal questions answered
A language model gives a fluent and convincing answer to a question about deductibility or a VAT treatment. That answer rests on what was in its training data — potentially outdated regulation, potentially Dutch rather than Belgian. To a layperson that difference is invisible.
As an aid for finding a lead that you then verify at source yourself: fine. As an answer that goes to a client: no.
Pasting client data into a public chatbot
The moment you paste an invoice, a set of annual accounts or a list of client names into a free consumer tool, that information leaves your control. For a profession bound by professional secrecy, that is not a theoretical risk. Use an environment where you know the data stays put, or do not use it.
Posting automatically without intervention
Technically it is possible. But a wrong posting nobody has seen often only surfaces months later — and then you have to correct not just the error but also work out how many other entries were made the same way. The time saved by skipping the check is lost a hundredfold at the first incident.
The practical trade-off
A usable rule of thumb: the easier a mistake is to fix afterwards, the more freedom you can give AI.
- Low risk: a draft text, a search, a proposal you review anyway. Here AI is a clean time saving.
- Medium risk: extracting data from documents. Do it, but build in checks that recalculate the outcome.
- High risk: something that goes straight to a client or to the tax authorities without a human step. AI does not belong here.
What this means for your firm
The question “do you use AI?” is really the wrong question. The right questions are: what for exactly, what happens when it goes wrong, and who looks at it before it goes out.
A supplier who has no clear answer to that — or who presents AI as something that always and everywhere works better — is selling you a story rather than a solution.
That is how I build it too: AI where it helps, with you in final control. See what that delivers concretely, or request a free workflow scan.