Professional secrecy and generative AI in Belgium: lawyers and notaries

In a law firm in Brussels, a notary's office in Namur or an accounting firm in Antwerp, the days look alike: pleadings to review, deeds to prepare, tax returns to check, often in French, Dutch and English. Generative AI assistants promise to lighten all of this. Many staff already use them.

But these professions have one thing in common: they live on secrecy. A client who shares a divorce, an estate or the company accounts does not expect that information to pass through an AI provider's servers. Here is how to organise the use of these tools in practice without weakening that bond of trust.

Professional secrecy now sits in Article 352

For more than a century, Belgian professional secrecy rested on Article 458 of the old Penal Code. The new Penal Code, created by two laws of 29 February 2024 and in force since 1 September 2026, carries the principle over into its Article 352. The substance is unchanged: people who hold secrets by virtue of their status or profession may not reveal them, except where the law provides. A breach is punished with a level 2 penalty, which includes prison terms of six months to three years.

Lawyers, notaries, and external accountants and tax advisers are all covered. For the latter, the law governing the profession expressly refers to the secrecy provision of the Penal Code. On top of this come the GDPR, enforced in Belgium by the Data Protection Authority (APD/GBA), and each professional order's own ethics rules.

What the bars have already said

In January 2025, the Orde van Vlaamse Balies and the Ordre des barreaux francophones et germanophone published guidelines on the use of artificial intelligence by lawyers. They are recommendations, not binding rules, but the message is clear. The lawyer remains responsible for the output. The lawyer must read the tool's terms of use. And information covered by secrecy should not be entered into an AI tool unless there is absolute certainty that it runs in a closed environment with sufficient safeguards. Personal data should be pseudonymised.

Notaries and accountants may not have such a specific text, but the logic is the same: without guarantees on how data is handled, secrecy must come before convenience.

An often misunderstood point: negligence

Under criminal law, breaching secrecy requires a deliberate disclosure. A staff member who pastes a file into a chatbot without thinking will not necessarily be prosecuted. But that protects neither the firm nor the client. Civil liability still applies, even without intent, and the professional order can impose disciplinary sanctions. Then there is the GDPR: a personal data breach may have to be notified to the Data Protection Authority.

In other words, saying "it wasn't deliberate" settles nothing. What counts is having taken reasonable steps to prevent the incident.

Five working rules for the firm

A useful usage policy fits on one page. Here is a starting point most firms and offices can adapt:

1. Classify the tools. Separate consumer tools, banned for client matters, from tools approved after reviewing the contract, data location and retention terms.

2. Ban identifiers in clear text. Names of parties, national register numbers, company numbers, IBANs and addresses must never be sent as they are.

3. Pseudonymise by default. Replace identifying data with consistent fictitious values, as the bars recommend.

4. Always review. No AI output goes to a client or a court registry without review by a professional.

5. Log incidents. Keep a record of blocked or corrected submissions so you can answer the order, the regulator or a client.

For the implementation side, our 90-day rollout playbook sets out a realistic timeline.

Why training alone is not enough

Firms often hold one awareness session and stop there. The trouble is that leaks happen under pressure: a filing deadline, an urgent translation, a client on the phone. That is exactly when attention slips. We explain this in our article on why AI security training does not work on its own.

A technical control on the workstation backs up training. It spots a national register number, a client name or an IBAN at the moment of sending, then replaces those elements or blocks the request. With format preserving anonymization, the text stays usable by the AI without containing the real data.

The link with governance

In our first article on Belgium, we looked at the board's role in generative AI and NIS2. For a firm or a notary's office, the question sits with the partners: who approves the tools, who follows up incidents, who answers when a client asks how their data is handled? Naming one person in charge, even part time, keeps the subject from falling between the cracks. That person also keeps the list of approved tools up to date and prepares a clear answer for clients who ask.

In summary

Professional secrecy has not changed in nature with the new Penal Code, and it applies in full to AI tools. The bars already advise against entering information covered by secrecy without solid guarantees. A short policy, systematic pseudonymisation and a control at the moment of sending let lawyers, notaries and accountants benefit from AI without betraying their clients.

ZeusLock detects and anonymizes client data in prompts and files before they are sent to AI assistants, and keeps a record of every incident. Request a demo for your firm.