AI at the accountancy firm: five situations in which PrivaSearch saves time

How does an accountancy firm use PrivaSearch?

An accountancy or bookkeeping firm uses PrivaSearch to search client files with an AI assistant without giving client data away. Documents are put in a project per client, personal data is replaced by pseudonyms, a colleague checks that, and the assistant then answers questions with the source attached.

That sounds abstract, so this article shows it through five made-up but recognisable situations. First, briefly, why it cannot simply be done with a chatbot.

Why not just use a chatbot?

An annual accounts file, a payroll administration or an income tax return is a pile of personal data. Payslips and tax returns contain BSNs (Dutch citizen service numbers), and strict rules apply to the use of the BSN: the Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, explains this on its page about the BSN. Anyone who pastes such a document into a chat window sends it in its entirety to the provider. The AP warns that this can lead to a data breach.

For accountants, professional regulation comes on top of that. In the VGBA of the NBA, the code of conduct of the Dutch professional body of accountants, confidentiality is one of the fundamental principles. The NBA and NOREA write in Guidance 2: AI applied (June 2026) that client and personal data may not simply be entered into public AI models and should remain in a secured, shielded environment. It also states that AI output on its own is never sufficient audit evidence.

The result at many firms: colleagues use AI secretly, or not at all. Neither is what you want. Our article on shadow AI is about the first of these.

What does the way of working look like?

PrivaSearch is software that lets your team search its own client files with Claude, ChatGPT or Mistral Vibe, after names, BSNs (Dutch citizen service numbers) and addresses have been replaced by pseudonyms and a colleague has checked that. At a firm, it looks like this.

1. You create a project per client and upload the documents, or you connect a SharePoint library read-only.
2. A language model within PrivaSearch and Dutch patterns find names, addresses, BSNs, IBANs, number plates and dates of birth and replace them with consistent pseudonyms.
3. A colleague sees the original and the pseudonymised text side by side, corrects what was missed and approves the document.
4. The document is then searchable, only for those who are a member of the project.
5. If you wish, you connect an AI assistant via a secured MCP connection. The assistant only sees released passages, never the originals.

Where an employee's name and BSN appeared on a payslip, the assistant reads "[PERSOON_B4A1]" and "[BSN_361B]". Because the pseudonyms are consistent, the assistant does recognise that three documents concern the same person.

Five situations from practice

1. The client's question about last year

A client calls: how did it work again with the loan from the director and major shareholder to the private limited company? The employee who handled the file last year no longer works there. Instead of leafing through the annual accounts, the minutes and the correspondence, a colleague asks in PrivaSearch: "What arrangements are there about the DGA's loan to the company?" The answer comes with the passages from the loan agreement and the notes to the annual accounts, with page references.

2. A new client with a moving box full of documents

A bookkeeping firm takes over a client from another firm and receives a folder with five years of documents. The question is which arrangements, schemes and ongoing obligations exist. Via PrivaSearch, an assistant can make a first overview: which contracts mention a term, which payment arrangements exist with the Dutch Tax Administration, which securities have been provided. The employee checks every point against the source.

3. Payroll and the collective labour agreement

For a payroll question, you often need to put the employment contract, the collective labour agreement and an earlier salary arrangement side by side. Employment contracts and payroll documents are exactly the kind of documents with BSNs, addresses and dates of birth. After pseudonymisation, the assistant can search those documents and point out which provision, for example on a year-end bonus or a notice period, appears to apply. The payroll administrator assesses it.

4. The tax file of a private individual

An income tax return sometimes contains healthcare costs or other information about health. That is special category data, with extra protection; more on this in our article on special category data. Detection finds names, BSNs and addresses, but is never complete. That is why the colleague in the review looks extra carefully at precisely such passages and adds what the model missed. If in doubt, leave the document out of the project.

5. Preparing an acquisition

In an acquisition or a due diligence, contracts, staff overviews and annual figures come together. An assistant can search the released passages for change-of-control provisions, non-compete clauses and warranties. Figures and amounts remain readable, because PrivaSearch replaces personal data and not amounts. Bear in mind that an acquisition price is not personal data, but it is confidential: such passages end up in the conversation with the assistant.

What does PrivaSearch do and not do?

Permissions work per project. A colleague who is not in a client's project sees nothing of that client, and administrators do not get automatic access. Every project gets an end date; after that, access is closed and copies, fragments and pseudonyms are cleared away, unless a legal hold applies.

What should you watch for when introducing it?

Connect only a business account. What PrivaSearch returns to the assistant reaches the provider of that assistant, with pseudonyms instead of names. So connect a business account with a data processing agreement; which subscription qualifies is set out in Which subscription for business use?.

Do not type names into your assistant. What you type into the assistant yourself goes directly to the provider, bypassing the pseudonymisation. Ask about "this client", not about the name.

Have employees complete the training. Without a valid AI literacy training, the workspace does not open and the connection with the assistant is refused as well. Why that training exists is explained in our article on AI literacy.

Document it. Include the use in your record of processing activities and your AI policy. How pseudonymisation relates to anonymisation, and why pseudonymised data remains personal data for your firm, is set out in Pseudonymise or anonymise?.

The way of working step by step is on the page How PrivaSearch works. If your firm also works for commercial clients with insurance, read how an insurance broker analyses policy conditions with AI. If you want to see it with a file of your own, book a demo.

Which questions work well?

The quality of an answer depends heavily on the question. A few lessons for anyone who starts working this way.

Ask about something that is in the documents, not for an opinion. "What arrangements about repayment are in the loan agreement?" produces passages you can reread. "Is this loan in order for tax purposes?" asks for an opinion that the assistant should not give and that appears in no document at all.

Limit the question to one project. Because permissions apply per project, the assistant only searches the file the question is about. That prevents an answer from mixing in passages from another client, and it keeps the source references clear.

Always ask for the source. An answer with the name of the document and the page can be checked in a few seconds. An answer without a source is a claim, and a claim is not audit evidence. If the assistant says something that is not in the passages, you know immediately that you should ignore that part.

And work in steps. Have the assistant first collect the relevant passages and only then make a summary. That way you see what it used and what it left aside.