For accounting firms

AI at an accounting firm

AI can take a lot of research and writing off a firm's hands. But pasting client data into a chatbot can be a data breach. Here is what to know and arrange before you start.

In short

An accounting firm in the Netherlands may use AI, but client data cannot simply go into a chatbot.

Client data in a chatbot
Against the agreements: a data breach, in many cases reportable (AP).
Personal subscription
OpenAI and Anthropic may use input for training; you can turn that off, at Anthropic with exceptions.
BSN
Only use it where the law provides for it (art. 46 of the Dutch GDPR Implementation Act).
AI literacy
Applies to organisations deploying AI systems since 2 February 2025.
Read more

On 6 August 2024 the Dutch data protection authority, the Autoriteit Persoonsgegevens (AP), warned about data breaches caused by employees sharing personal data with an AI chatbot. If that happens against the agreements made, it is a data breach, and in many cases it must be reported. So arrange three things first: which tools your staff may use and which data may go in, a business agreement with the provider, and your client's permission if, as a processor, you have that provider process his data. Responsibility for the work stays with people: the business owner remains responsible for his dealings with the Dutch Belastingdienst, even when he authorises someone else.

What can AI do at an accounting firm?

AI is strong with text: summarising a long email, drafting a question for your client, explaining a rule in plain language or translating a letter from the Dutch Belastingdienst. That work can be done without client data if you leave out names and numbers.

Figures are different. A language model can copy an amount wrongly or invent a balance that looks convincing. So don't let AI calculate with figures you pasted in. An assistant connected to your accounting software reads the amounts from the software; you still need to check the explanation around them.

Separate chatbot or agent connected to your software

Separate chatbotAgent connected to your software
Where the figures come fromWhat you paste inFrom the books themselves
Which data it seesWhat you enterWhat you tick when connecting
Does data go to the AI service?YesYes, what the agent requests
BookingsNoneDepends on the software and your settings
Stopping accessNot applicable: what was pasted cannot be taken backRevoke the connection; what was already requested stays with the AI service
A connection makes the figures more reliable. You need the agreements with your client and the AI service either way.

May you put client data into a chatbot?

Not without agreements. The AP received notifications of data breaches caused by employees sharing personal data of, for example, customers with an AI chatbot. According to the AP, employees often use chatbots on their own initiative and against the agreements with their employer; if personal data were entered, that is a data breach. If the use is part of the organisation's policy, it is not a data breach, but according to the AP it is often not permitted by law.

The AP advises organisations to agree clearly whether employees may use chatbots and which data they may and may not enter. You can also arrange with the provider that it does not store the data entered. If things go wrong anyway, notifying the AP and the people affected is mandatory in many cases.

Business or personal subscription: what does the provider do with your input?

That differs per subscription. OpenAI writes that it may use content from services for individuals, such as ChatGPT, to train models, and that you can turn this off. It does not use content from its business offering, such as ChatGPT Business, ChatGPT Enterprise and the API Platform, for that.

Anthropic writes in its commercial terms that it does not train models on customer content from those services, and it has a Data Processing Addendum for them. On the consumer plans Free, Pro and Max, input may be used for training unless you turn that off; even then there is an exception if you give feedback or input is flagged for a safety review. Read your own provider's current terms before you start; they change.

What the provider does with your input

ProviderConsumer planBusiness offering
OpenAIServices for individuals, such as ChatGPT: content may be used for training; you can turn this offChatGPT Business, ChatGPT Enterprise and the API Platform: not used for training
AnthropicFree, Pro and Max: input may be used for training unless you turn that off (exception: feedback or a safety review)No training on customer content; Data Processing Addendum
Read your own provider's current terms before you start; they change.

Do you need your client's permission?

That depends on your role, and it is not always clear-cut: whether a firm is a processor or a controller itself differs per service. If you are a processor, there should be a processing agreement. Under Article 28(2) GDPR you may then only engage another processor, such as an AI service, with your client's prior specific or general written authorisation. You impose the same obligations on that processor and you remain liable to your client yourself (Article 28(4)).

So first check in your processing agreements whether you may add a sub-processor and how you notify your clients. Your role depends on what you actually do, and separate rules apply to transfers outside the EEA. If in doubt, get legal advice.

May a BSN go into an AI tool?

Be cautious here. Under Article 46 of the Dutch GDPR Implementation Act you use a statutory identification number, such as the BSN, only to carry out the law that prescribes it or for purposes set by law. So remove a BSN from anything you put into a chatbot. We found no statement by the AP or the Dutch Belastingdienst that allows a BSN in an AI tool.

Do your staff need to know about AI?

Yes. Article 4 of the European AI Act requires organisations that use AI systems to work on AI literacy. This has applied since 2 February 2025; according to the European Commission, supervision by national authorities has run since 2 August 2026. The text was amended in 2026: you take measures to support the AI literacy of your staff, and no specific level per person is required.

In practice that means explaining which tools are allowed, what must not go in and that every answer gets checked. If you are a member of the Dutch association NOAB, also look at its conduct rules on confidentiality (Article 6) and the AI Governance course, which NOAB has designated as compulsory for 2026; NOAB can tell you who at your firm takes it. Accountants also have the duty of confidentiality in the VGBA.

Who is responsible when AI makes a mistake?

Not the AI. The Dutch Belastingdienst is clear about authorising someone: you are responsible for your own tax affairs, even if you have authorised someone else. It says nothing about AI there, but the line is the same: the business owner remains answerable for a mistake in a return. What your firm has to answer for towards the client follows from your engagement and your terms.

So let nothing AI produces reach a client or the Dutch Belastingdienst unseen. Record who checks what.

How to arrange AI at your firm

  1. Choose your toolsWhich AI services are allowed, and on which subscription.
  2. Agree what may go inWhich data may and may not. A BSN does not belong there.
  3. Arrange the agreementWith the provider, and read whether it stores input or trains on it.
  4. Look at your clientsMay you add a sub-processor under your processing agreement?
  5. Explain it to your staffIncluding what to do when something goes wrong.
  6. Keep checkingNo answer goes out unseen.
Sources: Autoriteit Persoonsgegevens, 6 August 2024; GDPR Article 28; European Commission on Article 4 of the AI Act, read on 31 October 2026.

How does your own AI agent work in booxx?

In booxx you connect your own assistant, such as Claude or ChatGPT, through MCP to the client you have open. You choose which parts it may read. With 'All clients I can access' you ask questions about your whole portfolio, such as which clients are behind on VAT; such a connection can only read.

If you have the agent prepare work for one client, bookings are by default proposals a person approves. What the agent requests goes to the AI service you choose, so the agreements in this guide apply here too. What the agent requests or tries for a client is in that client's history, call by call, with its name.

Try free for 90 days90 days free, then €19 a month · no payment method needed · cancel any month

Updated October 31, 2026. General explanation of the Dutch rules, not tax advice — for your own situation, decide together with your accountant or the Dutch Tax Administration.

Frequently asked questions

Will AI replace bookkeepers?

Nobody knows. AI already takes over research and writing, but a business owner remains liable for his tax obligations and wants someone who knows the figures and explains the choices. So we expect the work to shift towards checking and advising.

May I use ChatGPT with my clients' data?

Only with agreements. The AP calls it a data breach when an employee puts personal data into a chatbot against the agreements. Even with a policy, the AP says it is often not permitted by law. So use a business agreement, agree which data may go in and check your processing agreements.

Do I have to report a data breach caused by a chatbot?

According to the AP, notifying the AP and the people affected is mandatory in many cases if an employee leaks personal data through a chatbot against the agreements. If you are a processor, you inform your client without undue delay (Article 33(2) GDPR). Unless a risk is unlikely, the client reports to the AP, where possible within 72 hours of becoming aware. If you are the controller, you report.

Is a business AI subscription enough?

It is a start, not the answer. A business subscription settles what the provider does with your input. You also need agreements with your staff, and your client's permission if, as a processor, you have the AI service process his data.

Do my staff have to take an AI course?

The AI Act does not prescribe a course. You take measures to support your staff's AI literacy, which can be a short explanation of what is and isn't allowed. NOAB has designated the AI Governance course as compulsory for 2026; ask NOAB who at your firm takes it.

Can I use an AI agent for all my clients in booxx?

Yes, for reading. A staff member chooses 'All clients I can access' when connecting and asks questions about the whole portfolio. Bookings are prepared per client, and the agent can never do more than the staff member.

GuidesAccounting software that speaks your language

booxx.nlYour books, in your own language

Try it first with your own firm's books

Want to see how the package handles the books first? Start a trial with your own firm's administration. You then see the package the way your client sees it, and on your dashboard you request the firm environment with the worklist: you enter how many clients you have and in which languages, and we set it up.

90 days free, then €19 a month · no payment method needed · cancel any month

From €9 per managed administration per month, excluding VAT. No fixed firm fee, no price per staff member.