AI compliance for law firms: 5 pitfalls
The AI literacy obligation from the EU AI Act has applied since 2 February 2025[5], and the NOvA, the Dutch bar association, published its recommendations for AI in the legal profession at the end of 2025[6]. From 2026, the College van Toezicht, the supervisory board for the Dutch legal profession, is explicitly checking firms on safe and careful use of AI[9]. Most of the firms we speak to have something in place around AI: a paid Copilot subscription, a pilot with a legal tool, an internal guideline. In practice, though, they fall into the same five pitfalls.
We implement AI at law firms, notaries and accountancy practices in the Randstad[8]. In this article we describe the five pitfalls we come across most often, including a contrarian view on Microsoft 365 Copilot that goes against the prevailing marketing. For the broader context on AI implementation in professional services we refer you to our pillar article. Here we zoom in on compliance specifically.
Why AI compliance is no longer optional in 2026
The five core values of the legal profession, namely independence, partisanship, expertise, integrity and confidentiality, apply in full when AI is used. That is not a legal detail: it means a disciplinary complaint about an AI incident goes straight to your admission to the profession. The NOvA names three concrete obligations for firms:
A firm-wide AI policy: no ad hoc experiments per lawyer, but an adopted policy with use cases, prohibited applications and review procedures.
AI literacy for every member of staff: including trainees, paralegals and support staff, documented so you can prove it.
Transparency towards clients: clients need to know that AI is being used on their file, and how.
Any firm that does not have these three in order is empty-handed when an audit or a disciplinary complaint arrives. Here are the five pitfalls.
Pitfall 1: treating Microsoft 365 Copilot as "safe because it is Microsoft"
This is our most contrarian observation: Copilot in an EU tenant does not give you watertight legal compliance for legal work. We see firms rolling Copilot out broadly with the argument that "it sits inside our tenant, so the data stays with us". That is not correct.
Microsoft, Google and Amazon fall under the American CLOUD Act[4]. That law gives US authorities access, under certain conditions, to data processed by these suppliers, even when it physically sits on European servers. For client data in a civil case the risk is small. For criminal cases, sanctions files, M&A transactions with American interests or whistleblower cases it is not negligible.
What we advise law firms with sensitive files to do:
Limit Copilot to tasks that are not client-specific: internal documentation, HR, marketing, office organisation.
For work on client files, use a specialist legal tool (Harvey, Spellbook, Leya, Clearbrief, Kleos AI) with a hard contractual ban on reusing your input for training.
Put both data location and data sovereignty in the contract, not just "within the EU".
Copilot is an excellent productivity tool. It is not a compliant tool for client files, whatever the account manager tells you.
Pitfall 2: free ChatGPT for "just a quick check"
The free and consumer versions of ChatGPT, Claude, Gemini and Copilot use the data you enter to train the model further by default, unless that has been explicitly switched off in an Enterprise or Team subscription. That is not a theoretical risk.
We have come across files where an employee quickly dropped a draft contract into free ChatGPT to check the structure. Client name included, amounts included, specific clauses included. That is by definition a confidentiality incident within the meaning of article 11a of the Dutch Advocates Act, regardless of whether the data ever turns up in someone else's answer.
The concrete policy we implement at firms:
Free and consumer AI blocked technically on office devices, through MDM or a DNS filter.
Only business subscriptions with a data processing agreement and a no-training clause.
An internal owner per tool who reviews the contract every year.
Pitfall 3: an AI policy on paper with no operational enforcement
Writing an AI policy is easy. We see the document sitting at almost every firm above 15 FTE. What we do not see: enforcement.
AI compliance consists of three layers: the policy document, the technical enforcement (MDM, DLP, firewall rules) and the cultural layer (training, review moments, case discussions). On average, firms invest 80% of their time in layer 1 and 5% in layers 2 and 3 combined. An audit by the supervisory board looks at all three.
What works in practice:
A quarterly check in which a compliance officer reviews 5 random files for AI use.
A simple registration field in the matter file: "AI used? Yes/no. Which tool?"
A monthly 30-minute case discussion about AI incidents, both internal ones and ones from the news.
Without these three ingredients the policy document is a paper tiger.
Pitfall 4: ticking off AI literacy with a single e-learning
Many firms tick the AI literacy obligation from article 4 of the AI Act[2] with a generic 45-minute e-learning. Formally that is enough to have "done something", but in substance it is not enough to reach the goal: staff who recognise when AI is hallucinating, when bias is at play, and when confidentiality is at stake.
AI literacy is role-specific. An M&A partner reviewing contracts faces different risks from a trainee searching case law. Both face different risks from a secretary transcribing minutes. One generic e-learning does not cover that.
Our advice to firms:
Three tracks: legal substance (lawyers), administrative support (secretariat), and management (partners, board).
At least one live session a year using examples from your own firm, not just general theory.
Document per person which training they have followed, because that is what an inspector asks for.
Pitfall 5: "the lawyer remains ultimately responsible" as a universal safety net
This is the pitfall most firms do not spot themselves, because it comes wrapped up as reassurance: "AI is only a tool, the lawyer remains ultimately responsible, so the core values are safeguarded." Legally that is correct. In practice it is an illusion.
Research into AI and cognitive bias shows that people judge outputs that sound "well phrased" and "confident" less critically. That is exactly what generative AI produces. A lawyer reviewing an AI-generated memo at 17:30 on a Friday afternoon is a different reviewer from the same lawyer who would have written that memo herself on Tuesday morning.
The cases where this goes wrong are well known: lawyers in the US, the UK and by now in the Netherlands too who filed written submissions containing case law invented by AI[11]. In every single case they said: "I thought I had checked it."
What firms need to build in:
A mandatory source verification step for every AI-generated legal source (ECLI number, statutory article, reference to literature).
A rule that AI output never goes straight to a client or a court: there is always at least one independent human review in between.
Anonymous incident reporting with no disciplinary consequences for the first 12 months, so people dare to report.
What this means in practice for your firm in 2026
AI compliance is not a one-off project. It is an ongoing process that touches IT, HR, knowledge management, professional rules and client communication. For a firm of 20-50 FTE, setting it up takes roughly 6 to 10 weeks with an average investment of €15,000 to €30,000 for implementation, excluding licence costs for specialist tools. The alternative, reacting ad hoc to a disciplinary complaint or a negative inspection finding, is significantly more expensive and goes straight to your reputation.
We run AI compliance projects for law firms in Amsterdam, Rotterdam, The Hague and Utrecht. If your firm is expecting an inspection, or if you want to test yourself against the five pitfalls above, we will arrange a no-obligation 30-minute conversation in which we put your current situation next to the NOvA recommendations.
Sources
Artikel 11a ADVW (Geheimhoudingsplicht advocaat, medewerkers ...
Artikel 4: KI-Kompetenz | EU-Gesetz über künstliche Intelligenz
Wat betekent CLOUD-act voor jouw organisatie? | Yourcloudweb
AI in de advocatuur: wie nu niet beweegt, raakt straks achterop
College van toezicht zet in 2026 scherp in op risicogericht toezicht ...
Advocaten gebruiken AI in rechtszaken op de verkeerde manier
