At half past five a corporate lawyer has a thirty-page purchase agreement in front of her and a client who wants to know by nine tomorrow what it says. She pastes the text into a chatbot and has a summary within a minute. In the corridor a colleague mentions a contract analysis tool he is trying out after a conference. Nobody did anything wrong, and yet in both cases client information went to a vendor the firm has no agreement with.
That is what AI governance for a law firm is about. Not whether lawyers may use AI, because they already do. Which tools are in use, what the firm has decided per tool, whether the lawyer knows that decision at the moment the file sits in the text box, and whether the firm can show afterwards that the agreements were followed. What AI governance means in general is in what is AI governance. This piece is about what a law firm adds: the duty of confidentiality.
What do lawyers do with AI?
Three things, mostly. Summarising: a file, a stack of exhibits, a forty-page judgment. Drafting: the first version of a writ, a letter to the other side, a clause that needs tightening. And research: finding case law, exploring a legal question, testing an argument.
Around that sits a quieter layer. The firm's legal database now has an AI feature, the dictation software summarises, a transcription tool writes up the meeting with the client, a translation tool helps with a cross-border matter. Each processes text from case files, and most have never been decided on because nobody saw them as AI tools. How that layer comes about is in what is shadow AI.
| Task | What the input usually contains | The question for the firm |
|---|---|---|
| Summarising a file or exhibits | Party names, matter number, facts, positions | May this go into this tool, with this account? |
| Drafting a writ or agreement | Parties, amounts, the client's strategy | Does the strategy stay within the circle of confidentiality? |
| Searching case law | Often just a legal question, sometimes the facts of the matter | Do the facts give away who the client is? |
| Writing up a meeting | Everything the client said | Has any tool been assessed for this at all? |
Why does confidentiality weigh more than the GDPR?
For most organisations the question of AI and data is a GDPR question: personal data, legal basis, data processing agreement. For a law firm a duty sits on top that is older and reaches further. A Dutch lawyer (advocaat) has a statutory duty of confidentiality over everything entrusted in that capacity, with the right to refuse to testify about it as its counterpart. That right exists for the client, not the lawyer: a client must be able to speak freely.
Two things follow. The duty covers more than personal data: a client's negotiating strategy may contain no name at all and still falls under it. A data processing agreement settles the GDPR side, as a data processing agreement for AI tools explains, and leaves the professional duty untouched. And the duty generally extends to everyone working under the lawyer's responsibility: the secretariat, the paralegal, the trainee who writes the summary.
Supervision of lawyers in the Netherlands sits with the dekens, the local bar presidents, and the Netherlands Bar sets the professional rules. What it means when file information ends up in an AI tool differs per tool, account and situation. What holds in general: the firm must be able to explain the judgement it made, and "we did not know that tool was being used" is not a judgement. The Dutch Data Protection Authority, after reports of data breaches through AI chatbots, advises organisations to agree which tools are allowed and which data may go in, and you can only agree that about tools you know exist.
What does this look like at a firm of forty lawyers?
A firm has bought a business ChatGPT licence and allowed the AI feature of its legal database. The protocol is two pages and the partners have adopted it. Then the firm looks for the first time at which AI services are actually opened from the browser.
Next to the two allowed tools, Claude and Perplexity turn out to be in use, plus a transcription tool used for client meetings, the contract analysis tool from the conference, and a translation tool for a cross-border matter. Three of the five are recorded nowhere. The lawyer using the transcription tool had no reason to think it was not allowed.
This is not an incident. It shows that a protocol describes what the firm knew on the day it was adopted; use carried on afterwards. The real work starts here: choosing a status per tool. The transcription tool becomes "not allowed" until there is a data processing agreement and it is clear how recordings are stored. The translation tool becomes limited use, without party names. The contract analysis tool stays on "no decision yet", with only public model contracts going in until then. The four statuses are set out separately.
BeeSensible links a catalogue of 865 assessed AI tools to the use observed at the firm: only the domain of an opened tool, at most once per day per device, with no user id and no page content. An administrator sees which tools have no decision yet and records the decision in the AI Tools module. The overview:
