AI and client contact at a bookkeeping or financial services practice

A bookkeeping practice is not a webshop. If you use AI for client contact — answering email, chasing missing paperwork over WhatsApp, a phone assistant that takes the first question — rules apply that simply do not exist in other industries. The most important one is the Dutch Money Laundering and Terrorist Financing (Prevention) Act, the Wwft. On top of that you carry a bookkeeping obligation. And you work with figures that are not yours, but belong to your client and to the people behind that client.

That you need a legal basis, sign a processing agreement with your supplier and make the assistant recognisable as AI applies here exactly as it does everywhere else, and it is covered elsewhere in this knowledge base. This article is only about what comes on top of that in a financial practice.

Why is a financial practice a different legal case?

Accountants, tax advisers, bookkeeping practices and many financial advisers are an institution under the Wwft. That word matters, because the law places the obligations on the institution itself. Not on a system, not on a tool, not on a supplier. You can outsource work, but the duty stays with the practice. A vendor claiming its software is Wwft-proof does not take that duty off your hands — at best it hands you a tool for doing it better yourself.

Three things stack up here. First the Wwft: working out who your client really is, and reporting what looks unusual. Second the bookkeeping and retention obligation: contact with your client is often part of a set of records itself, and sometimes part of two. Third the nature of the data: a ledger holds not only your client's figures, but also those of their staff, their customers and their suppliers. Each of those three draws a different line around what an AI assistant may do.

Why is client due diligence not a job you hand to AI?

Article 3 of the Wwft is short and clear: an institution carries out client due diligence in order to prevent money laundering and terrorist financing. It does not say a review must merely exist; it says the institution performs it. That review is also risk-based. You establish who the client is, who ultimately sits behind them, what the purpose and intended nature of the relationship is, and whether what you see fits that picture. When in doubt you do more, not less.

That is a judgement, and one you must be able to explain to a supervisor later. A language model does not deliver a judgement, it delivers a likely answer. It does not know that the director's nephew suddenly became a partner last year, it does not notice that a story about where the start-up capital came from is a little too smooth, and it cannot account for its own reasoning in a way a supervisor can use. Anyone who lets an assistant finish the review has no risk assessment, only a summary.

What does work is the preparation. An assistant can explain which documents you need and why. It can keep track of what is still missing and chase it politely. It can ask the client for an extract, book the introductory meeting, or bring an open question back to the surface. That saves real work. But the conclusion — accept, apply enhanced due diligence, or decline — is set by a person, with a name and a date attached.

Watch out for the reverse trap as well. Do not let your assistant take on new clients by itself, even when the conversation drifts that way naturally. The decision to enter a business relationship is exactly the moment the Wwft requires you to look first. A chat window saying "welcome as a client" before anyone has checked anything reverses that order.

May an AI assistant take in identity documents?

Article 11 of the Wwft covers verification: where the client is a natural person, their identity is verified. Verifying is more than asking someone's name. It is establishing that the person is who they say they are, on the basis of a document fit for that purpose.

A chat channel is a poor place for that, for three reasons at once. The first is the channel itself. An identity document carries a photo, a document number and the Dutch citizen service number. That is the most sensitive thing a client owns, and it ends up in a conversation history that usually sits in more places than you think: with the channel provider, in the AI model's logs, in the search index of your own office system. The second reason is the check. A model can describe a photo, but it cannot establish whether a document is genuine. An edited scan looks just as convincing in a chat window as a real one. The third reason is evidence. If someone later asks how you established identity, you want to show a recorded action by a person, not a chat log in which an assistant replied "thanks, received".

The practical fix is simple and cheap. Teach the assistant that identification never runs through the chat. It names the route your practice does use — in person, in a secure environment, or through a service built for it — and steers the conversation there. If a client still sends a photo of their passport unprompted, there should be a fixed reply for that and a procedure to get the file out of the channel.

What happens if your assistant accidentally mentions a report?

Article 16 of the Wwft requires an institution to report an executed or intended unusual transaction to the Financial Intelligence Unit without delay, as soon as its unusual character has become known. The Wwft also contains a tipping-off ban: you keep quiet about such a report, including towards the client (article 23 Wwft). That is not a matter of etiquette but the heart of the system. A client who knows they have been reported can erase traces.

This is where the biggest AI risk in a financial practice sits, and it is a risk you meet nowhere else. An AI assistant is built to be helpful and to draw on everything it can read. If the file contains a note saying "reported", and the assistant reads that file, there is a real chance the information comes out at some point. Not necessarily word for word. It can be subtle too: an answer that suddenly turns formal, a reference to "an ongoing investigation", or an assistant refusing to say something in a way that says everything.

The measure is structural, not textual. Make sure everything connected to the Wwft report stays outside the sources the assistant can consult: a separate file, a separate folder, a separate system, with its own access rules. In the client system the assistant does read, put no flags, status labels or loose notes from which anyone could infer the report. A field named "Wwft status" is already too much if the assistant can reach it. And test it: deliberately ask the assistant the questions a suspicious client would ask, and see what it does.

The reverse holds as well. Do not expect an AI assistant to spot unusual transactions for you. Judging whether something is unusual depends on indicators and on what you know about this client and this file. That weighing stays human work, and the reporting duty stays with the practice.

Which tasks can you automate and which not?

It helps to look task by task instead of talking about "AI in practice" in general. The table below is not advice about your own practice, but it shows what kind of reason draws each line.

Client contact tasks in a financial practice, with the legal reason why automating is or is not possible

Task	Automate?	Legal or factual reason
Booking the annual accounts meeting	Yes	No statutory role, no substantive judgement.
Chasing missing receipts and bank entries	Yes	Preparation for the bookkeeping duty; a person assesses what arrives.
Answering questions on deadlines and procedures	Yes	General information, not advice tailored to a person.
Carrying out and closing client due diligence	No	Article 3 Wwft places the review with the institution; it is a risk judgement.
Establishing and verifying identity	No	Article 11 Wwft requires verification against a document.
Judging whether a transaction is unusual	No	Article 16 Wwft; the weighing and the report belong to the institution.
Saying anything about a report that was filed	Never	Tipping-off ban, article 23 Wwft; the assistant should not even be able to read it.
Taking a tax position for this client	Draft yes, send no	The advice remains the practice's; liability and professional rules.
Retaining business chat conversations	Yes, and often mandatory	Article 52 AWR: data carriers must be kept for seven years.
Automatically accepting or refusing a client	No	The decision on the business relationship belongs to the institution, after review.

Does a WhatsApp chat with a client belong to your records?

This is the point practices miss most often. In its guidance on automated bookkeeping, the Dutch tax authority notes that communication tools used mainly in private — a private email address, WhatsApp — can form part of the records if they are also used for business. So it is not the channel that decides, it is the content.

Translate that to an AI assistant. If the assistant uses WhatsApp to settle with a client which receipt belongs to which entry, which amount is still missing or why an invoice was corrected, that conversation is not "just a chat". It is a document in which rights and obligations are recorded, and so it can be part of the records. Article 52, paragraph 4 of the Dutch General Tax Act provides that data carriers must be kept for seven years, unless tax law says otherwise.

There is a clash there you have to resolve in advance. Many chat and AI platforms default to a short retention period, or delete conversations after a few months because that looks tidy from a privacy angle. For a financial practice that can mean your records are missing documents. The opposite is no answer either: keeping everything forever is not a solution, because the retention duty is a duty and not a licence. What you need is a deliberate choice per type of conversation.

• Conversations about documents, amounts, corrections and arrangements on the file: keep them, and keep them findable during an inspection.
• Pure logistics — moving an appointment, an opening time — need not be part of the records and can have a shorter period.
• Everything the assistant itself produces in intermediate steps and prompt logs is not records, but may well contain personal data. Set a separate period for that.

What changes because you work with other people's figures?

A practice rarely keeps only its own books. The Dutch tax authority points out that when an entrepreneur has their records kept wholly or partly by third parties, the data those third parties hold about the business must be retained as well — and that the retention duty continues after the business has ended.

For you as a practice that means something uncomfortable: your correspondence about your client's file can be part of your client's records. If that client leaves or stops trading, the duty does not disappear. An AI assistant holding conversations on behalf of the practice therefore produces material that is not only yours. That has two consequences.

The first is that you never simply wipe a file because a client walked out. Agree how you hand over at the end of the relationship and what stays behind, and make sure the chat history is part of that agreement.

The second concerns what sits inside the file. A payroll administration holds citizen service numbers, salaries and sometimes absence data of people who never made any choice about your software. They are not your clients; they work for your client. The same goes for your client's debtors. The sensible line is a sharp one: let the assistant read the contact file — who the contact person is, which documents are outstanding, what has been agreed — and not the ledger itself. That is also simply less work than sorting out afterwards which fields were sensitive.

If your supplier or the model sits outside Europe, the question of international transfer comes on top. That applies everywhere, but it bites harder here, because the data is not yours and your client had no say in it.

Where is the line between information and advice?

An assistant that answers quickly is soon seen by clients as the practice itself. That is exactly what you want in tone and exactly what you do not want in substance. "We take care of the income tax return and send you a list of documents beforehand" is information. "You would be better off putting that car on the business" is advice, and that advice ends up in your practice's name, with all the liability that carries.

Draw the line in the assistant's instructions, not in the hope that it will go well. In practice a short list of topics where it always refers on works well: anything involving a position, an estimate, a deadline in an individual case, or a choice between two tax routes. And let it say so properly, with a concrete next step: who calls back, and when. A referral without a follow-up feels to a client like a closed door.

The same applies in reverse. If a client mentions something in the chat that deserves attention — a cash payment they are casual about, a loan from an unknown party — that has to reach a person. So make sure the assistant does not only give answers but can also pass signals on. That is the one place where AI genuinely moves the Wwft work forward: not by judging, but by letting nothing slip.

How do you record all this without turning it into a project?

This does not need a quarter of your time. What you need is three pages setting out what you have decided, so you can show it and adjust it when something changes.

1. Write down which topics the assistant handles and where it always stops. Include client acceptance, identification, anything around reports, and individual tax advice.
2. Put on paper which sources the assistant may read, and which systems deliberately stay out of reach. The Wwft file belongs on that second list.
3. Decide per type of conversation how long you keep it, and test that against the seven years in article 52 AWR for the part that counts as records.
4. State in the engagement letter which route identification takes, so a client does not drop a passport into the chat out of convenience.
5. Agree what happens to the conversation history when the relationship ends.
6. Test twice a year with awkward questions, and keep what came out. That is your evidence that you take this seriously.

The thread running through all of it is the same. In a financial practice, AI may prepare, request, remind and phrase things nicely. The judgement, the identification and the report stay with a person who can sign for them by name. Hold on to that separation and automation actually brings calm here: the assistant does the chasing, and you keep time for the work the law appoints you to do by name.