AI for lawyers: practical, confidential, and compliant

What AI means for legal practice

In 2023, a New York lawyer stood before a judge to explain why his brief cited six cases that did not exist. He had asked ChatGPT to find case law and submitted the answer without reading it. That is the story every lawyer hears the moment AI comes up, and the caution it teaches is fair.

But it is not the whole story. A contract analysis that takes a junior 4 hours, AI does in 10 minutes. A due diligence of 200 documents goes from a week to a day. A targeted case-law question yields an annotated shortlist within minutes. That is not marketing copy, it is what happens once you set it up properly.

The problem is that "set it up properly" sets a much higher bar in legal work than in marketing or finance. Confidentiality, accuracy, and liability are not minor concerns, they are the foundations. Bar association rules, professional secrecy, professional liability: a hallucination in a pleading can have disciplinary consequences. A wrong assumption in advice can cost millions. This article shows where AI already works in a law firm, where it is genuinely dangerous, and how to start without repeating the New York mistake.

Where AI is already being used

Contract analysis and due diligence

This is the clearest win. Kira Systems searches hundreds of contracts for specific clauses, risks, and deviations from standard terms. For an M&A process with 50+ contracts, that is the difference between a week of junior hours and a few hours of review. The human reads the signal, not the whole stack. Alternatives: Luminance, eBrevia, and LeXial for smaller firms.

Legal research

ROSS Intelligence searches case law and legislation in natural language. Instead of stitching search strings into case-law databases or EUR-Lex, you ask the question the way you would ask a colleague. For the Netherlands, Lexis+ AI, vLex, and since 2024 also Westlaw work with integrated AI research. The non-negotiable: a tool that cites from a validated database, not a chatbot guessing from memory.

Document generation

LawGeex generates and reviews legal documents based on pre-set criteria. Standard contracts, NDAs, terms and conditions, summons, deeds: AI produces the first version, the lawyer refines and signs off. For SME firms, Claude or ChatGPT with a well-developed prompt and an in-house template library often does the same job, at a fraction of the cost of enterprise tools. Start there.

Predictive analysis

Based on historical rulings, AI predicts the likely outcome of a case. Not as a replacement for legal judgment, but as a data point alongside your own read. Particularly useful for settle-or-litigate calls or estimating timelines. Tools: Lex Machina (mostly US), Premonition. For the Netherlands this is less developed, simply because our case-law dataset is smaller.

Standard correspondence and client management

The dullest category and at the same time the easiest place to start tomorrow. AI-drafted letters to opposing counsel, fee statements with explanations, intake questions for new clients, dossier summaries for periodic updates. Low risk, immediate time saved, no tooling investment. If you have no idea where to start, start here.

The challenges

Confidentiality

Attorney-client privilege is sacred and legally protected. Every AI tool must be GDPR-compliant and data must never be used to further train the model. Cloud AI with a data processing agreement works fine for most activities. For the heaviest matters (M&A, criminal law, international arbitrations) on-premise or an EU-private deployment is a serious option to consider. Discussed per matter, not the default route.

Hallucinations and factual errors

The biggest risk in legal work, and the reason the New York lawyer caught sanctions. AI confidently cites incorrect case law, invents statutes, and misreads clauses, all in sentences that look authoritative. Generic ChatGPT as a research tool for pleadings is a bad idea. Full stop.

The fix is threefold. One: work with retrieval systems that only cite from validated source documents (case-law databases, EUR-Lex, your own contract library), not from "training data". Two: every AI claim links back to its source so you can always verify. Three: for pleadings and client advice, heavy human review is the floor, not the ceiling. If you can't commit to that, don't use AI for that task.

Transparency

When AI makes a recommendation, the lawyer must be able to explain why. Black-box algorithms are unacceptable in a courtroom and in advice to clients. Ask every vendor explicitly how the output is constructed and which sources have been consulted. For AI Act compliance this is also a legal requirement for high-risk applications, and most legal work falls in that bucket.

Liability

If AI misses a clause in a contract review, who is liable? The vendor? The firm? The individual lawyer? Under most jurisdictions the lawyer remains ultimately responsible, so protocols for human supervision are not optional. Align your professional liability insurance with AI use as well. Most policies cover AI errors only if demonstrable human review has taken place, and "I trusted the tool" is not a defence.

Bias in historical data

Legal data carries historical biases: decisions that were normal in the past but read as discriminatory today. AI systems trained on this reproduce those biases, especially in predictive systems for case outcomes or sentencing. Work with systems that are transparent about training data and methods, and treat outputs as signal, not verdict.

Getting started with AI safely: step plan for law firms

Month 1: analysis and goals. Which processes cost the most time right now: contract review, standard correspondence, case-law research? Which are high-risk from a liability standpoint and which are low-risk? Train your team in AI fundamentals, and then specifically the legal application, not the generic MOOC.

Months 2-3: pilot on a low-risk process. Don't start with pleadings. Start with standard correspondence or contract templates. Tooling: Claude or ChatGPT (business version with a data processing agreement) plus a developed template library. Workflow: AI drafts, lawyer checks. Measure: time per document, error rate in the first version, client satisfaction.

Months 4-6: scale and expand. Only after the first pilots are running do you move to contract review or case-law research. Here you set up a retrieval system against your own contract library or a validated case-law database. Build protocols for human supervision and document everything you do.

Month 7+: AI Act compliance and confidentiality upgrade. Document all AI applications for AI Act compliance: which AI you use, for which processes, which data it processes, how supervision is arranged, how clients can object. For the most confidential work, investigate on-premise options.

How a typical engagement unfolds

An engagement runs in phases. First a short analysis to find the impact points in your firm. Then a focused pilot on one tool (a contract scanner or due-diligence assistant, for example) to test if it fits your way of working. If it doesn't, we stop and try something else. If it does, scaling follows. You evaluate at each phase whether to continue, so the firm gradually adapts and gives feedback. For matters with strict confidentiality requirements, on-premise or an EU-private deployment is looked at case by case.

The future

AI makes legal services more accessible, faster, and more affordable. Not by replacing lawyers (which is neither possible nor permitted), but by freeing them from routine work so they can focus on what truly adds value: thinking, negotiating, advising, litigating.

The firms that invest in AI now are building a lead that will be hard to close. Not just in efficiency, but in client satisfaction (faster answers, better predictability of timelines) and in talent (juniors don't want to do only document review). The New York lawyer's story isn't a reason to avoid AI, it's a reason to organise it seriously.

Want to know which processes in your firm are most suitable for AI? Take the free AI Scan and receive a personalised analysis in your inbox within 1 minute.