AI for law firms: professional secrecy and bar guidance

What the rules ask

Professional secrecy is a core value of the legal profession in every EU member state. What a client shares in confidence stays confidential, including towards software suppliers. An AI service into which a colleague pastes a pleading is such a supplier.

Bar associations have worked this out for generative AI:

• The CCBE, the umbrella body of European bars, published a guide on lawyers’ use of generative AI in October 2025.
• The Dutch NOvA recommends in its 2025 AI recommendations: do not enter confidential or client data into public AI models, keep input and output within the firm’s environment, ask the client’s consent, and use only tools with a data processing agreement.
• The German BRAK published guidance in December 2024: using a tool that stores or passes on client data for model improvement breaches the duty of confidentiality (§ 43a(2) BRAO). Service providers may only access client secrets under § 43e BRAO, bound to confidentiality in writing.
• The French CNB published a practical guide in September 2024 that puts professional secrecy first and advises pseudonymising data before it goes into a generative AI.

Everywhere, the lawyer remains personally responsible for the advice and for checking what a model produces.

May a lawyer use ChatGPT?

For general tasks without client data, such as having a legal concept explained or rewriting a text without names, yes. For case work the free or Plus version does not fit the guidance: there is no data processing agreement, and the input leaves the firm.

A business version with a data processing agreement is a step forward, but the text still goes to an outside AI vendor, with a parent company outside the EU. Whether that fits “input and output within the firm’s environment” is a judgement each firm has to make and be able to explain to its bar and its clients. See also Is ChatGPT GDPR-compliant?

Where AI does help a firm

• summarising pleadings, correspondence and judgments;
• a first version of standard letters, notices and memos;
• searching and comparing your own templates, clauses and earlier opinions;
• building timelines and overviews from large files;
• due diligence: structuring large volumes of documents and flagging deviations, checked by a lawyer.

It is precisely this work that puts the most confidential information into the input. That is why the question where the model runs matters more than which model it is.

From separate tools to a firm-wide policy

Bar guidance does not ask for a ban, but for a firm-wide policy with oversight. In practice that means:

1. One approved AI environment instead of separate accounts per lawyer.
2. Rights per practice group, so one client’s files do not surface in another practice group’s environment.
3. Safeguards that stop client numbers, ID numbers and other identifying data where needed.
4. An audit log that shows who used which model and when, for your own oversight and for your bar.
5. Agreements with clients about AI use, recorded in the engagement letter.

werqly places a dedicated AI node in your firm’s server room for this. The models are open-weight and run only there: no outside model vendor receives the text. In werqly AI Control you set per role which models, sources and automations are allowed. More on our approach for firms is on AI for law firms and notaries.