AI for legal professionals mainly saves time on preparatory and text-based work, without shifting one iota of responsibility for the content. A language model reads through a contract, summarises a case file or drafts a letter, but the legal assessment remains human work.
Whether you are a lawyer, a notary or an in-house legal adviser, the question is not whether you use AI, but which check comes afterwards and what information you put in. This article covers both sides: where AI demonstrably strengthens your work, and where it gets you into trouble if you take the output at its word.
What AI changes in the work of legal professionals
What AI mainly changes in the work of legal professionals is the pace at which routine work gets done. A language model reads large amounts of text in seconds, recognises patterns in contracts and produces a first version. That affects the lawyer, the notary, the in-house counsel and the legal adviser equally, because much of their work consists of reading, analysing and writing.
Which work gets noticeably faster
The work that gets noticeably faster has three features in common. There is a lot of text, the outcome is a draft rather than a decision, and you hold the source material yourself. Think of cutting an eighty-page case file down to five key points, comparing a standard agreement with a version you have been offered, or rewriting a letter in plain language for a client.
Work without those features gains much less. Advice in which you weigh interests against each other, or a strategic choice in proceedings, cannot be outsourced to a model that predicts patterns.
What stays with the legal professional
What does not change is who is responsible for the outcome: the legal professional. A model that summarises a clause or flags a deadline delivers a suggestion, not a judgement.
The legal assessment, the balancing of interests and the final decision remain with you. That separation between preparatory work and responsibility is the starting point for any sensible use in legal practice, and it also determines how thorough your check needs to be for each type of task.
How do legal professionals use AI?
As a legal professional, you use AI for work that involves a lot of text and little discretion. The three applications below come up in almost every legal practice, from a sole practitioner to an in-house legal department of thirty people.
Analysing contracts and comparing clauses
Contract analysis works best when you give the model a basis for comparison rather than an open question. Supply your own standard agreement and ask which provisions deviate from it, which are missing and which are worded unusually broadly or narrowly.
Ask explicitly for factual deviations only, not interpretations. Without that instruction, a model fills the gaps with what is usual instead of what is actually there, and you end up reading advice about a contract that is not in front of you.
Preparing legal research without relying on the sources
You prepare legal research with AI by having it break down the legal question and suggest lines of enquiry, not by having it supply sources. A model can help you pin down which sub-questions are in play and which terms to search for in a legal database.
Always check the sources themselves in a reliable database. Treat every ruling, every article number and every citation a model mentions as unconfirmed until you have seen it yourself.
Summarising and rewriting legal texts
Summarising and rewriting legal texts is the application language models are best at, because you supply all the content yourself. Reducing a long case file to its core, drafting a letter or adapting a text to your client’s tone saves preparation time straight away.
A good prompt makes the difference here. You steer the output far more precisely by providing context and a clear role, and assigning a role in a prompt works best when that role ties in specifically with the legal task rather than with the profession in general.
Where AI goes wrong in legal work
In legal work, AI rarely goes wrong on grammar and almost always on facts. The common denominator is false certainty: a language model produces text that is correct in form, not necessarily in truth, and in legal work that distinction makes all the difference.
Made-up case law and article numbers
Made-up case law is the best-known risk: a model cites a ruling, an article number or a source that turns out not to exist when you check. Nothing makes it stand out: a made-up ECLI number has exactly the form you expect.
During the period when I worked intensively with ChatGPT, I came across that pattern consistently. The model cited studies and sources that sounded plausible and turned out not to exist when I checked. In my own work with Claude, I notice that it more often indicates when something is uncertain and is less quick to invent sources, but checking remains necessary even then.
Summaries that convincingly deviate from the source
Summaries that deviate from the source are the subtler risk, and precisely for that reason the more dangerous one. A summary can look fluent and complete and still say something different from the underlying document on a decisive point. If you only read the summary, you will never notice.
The practical remedy is mundane, and it works. Have the model refer, for each key point, to the passage it comes from, and for the points your conclusion rests on, check whether that passage really says what the summary claims.
What a legal professional never puts into an AI tool
As a legal professional, you never put client data and case documents unchecked into just any AI tool, because your duty of confidentiality and the GDPR continue to apply in full. Which tool you use is a secondary question; first you decide which data may leave your office at all.
Why the account type is decisive
The account type determines what happens to your input. Consumer versions may, under their terms, use the data you enter to improve the model, while business environments with data processing agreements are set up precisely to rule that out.
So before you start, check which account you are working with and whether your input is used for model training. An employee who pastes a deed into a personal account is doing something very different from the same action in a business environment, even though the screen looks identical.
Anonymising and putting data processing agreements in place
In many cases, anonymising removes most of the risk, because a contract analysis works just as well without names, addresses and details that can be traced back to a person. If you use AI structurally, arrange the right account type, put data processing agreements in place and agree which documents are always anonymised, because responsible AI use starts with those choices, not with the prompt.
Which AI tools suit legal professionals?
Two kinds of tools suit legal work, and the choice between them depends on the type of work, not on the marketing around it. A general-purpose language model covers most of it; a specialised legal tool pays off in a limited number of situations.
When a general-purpose language model is enough
A general-purpose language model such as ChatGPT or Claude is enough for anything where you supply the source material. That covers contract analysis against your own standard, summarising, rewriting, drafting a letter and drawing up a checklist from a case file you provide yourself.
For those tasks, a specialised tool mainly charges you for functionality you do not use. So start with what you already have and see whether you hit a limit before you sign up for a licence.
When a specialised legal tool pays off
A specialised legal tool pays off when your work centres on source research rather than text processing. Tools trained on case law and legislation, with a link to a legal database, solve exactly the problem where general models break down: supplying citations.
Weigh what the tool demonstrably solves against what it costs per user per year. For a firm where research accounts for only a limited share of revenue, that calculation often works out less favourably than the brochure suggests.
What professional rules and the AI Act ask of legal professionals
Professional rules and the AI Act require legal professionals to keep responsibility for AI output unambiguously with the human. Disciplinary law makes no distinction between a mistake you make yourself and a mistake you copy from a model without checking.
Disciplinary law and the rules of conduct for lawyers
Under disciplinary law and the rules of conduct, a made-up ruling in a court document is squarely your responsibility. In the Netherlands, the framework for lawyers is the NOvA Rules of Conduct of the Netherlands Bar, and on top of general due care, AI for lawyers calls for specific agreements on confidentiality and the use of case files.
The KNB AI scales for Dutch notaries
For notaries in the Netherlands, the AI-weegschaal (AI scales) of the Royal Dutch Association of Civil-law Notaries (KNB) is the starting point for deciding which use is responsible for each type of deed. AI in notarial practice touches directly on the duty of care in Article 17 of the Dutch Notaries Act (Wna) and on the notary’s gatekeeper role, which demands more rigorous checks than general legal work.
What Article 4 of the AI Act asks since its revision
Article 4 of the AI Act has applied since 2 February 2025 and originally required organisations to take measures to ensure, to the best of their ability, a sufficient level of AI literacy among their staff. Article 99(4) attaches no fine to Article 4; member states set the penalties themselves under Article 99(1), and the Dutch implementing act that arranges supervision is not yet in force. The Digital Omnibus rewrote that article with effect from 27 July 2026: since then, organisations take measures to support the development of AI literacy, without having to guarantee a fixed level for each individual employee.
What that means for your firm depends on how employees use the output, but the legal basis applies to almost every legal practice that works with AI. The obligation lies with the firm, not with the individual legal professional, and what exactly AI literacy involves determines which measures fit. Measures that support knowledge of the possibilities and risks are therefore no longer an optional choice.
Setting up AI responsibly in a legal practice
Setting up AI responsibly in a legal practice means moving from loose prompt tips to a shared way of working. An individual legal professional who prompts skilfully saves time for themselves, but only when the whole firm works to the same agreements does the result become verifiable and transferable.
The internal AI policy and the checkpoints for each type of task
An internal AI policy sets out which tools are permitted, which data may go into them and which check applies to each type of task. The four-eyes principle from my years at major banks translates directly here into checkpoints at firm level: nothing with legal consequences goes out without a second reading.
Scale that check to what is at stake. An internal note has a different risk profile from a court document or client advice, and a policy that does not make that distinction gets ignored in practice because it is equally burdensome everywhere.
Starting small with one to three processes
Starting small with one to three processes works more reliably than a firm-wide strategy that takes months to prepare. Choose processes with high frequency and limited variation, because that is where time savings show up fastest and where a mistake is easiest to catch.
Appoint one person internally to keep track of it all and update the agreements.
If you want AI in your firm to become a fixed way of working instead of a series of loose experiments, the AI Literacy Course for the whole team is a logical first step. Together with your team you build a prompt library, an internal AI policy and the checkpoints your profession requires, with responsibility where it belongs: with the legal professional.
Frequently asked questions about AI for legal professionals
AI for legal professionals is most reliable when you supply the source texts yourself, and for public sources there are fixed places to find them.
Can you have AI summarise published rulings from rechtspraak.nl?
Yes, you can have AI summarise published rulings from rechtspraak.nl, because the Dutch judiciary publishes them in pseudonymised form. You then supply the full text yourself, so the model works with the real ruling instead of inventing one. For larger numbers, the Dutch judiciary offers a free public web service that delivers rulings and ECLI data in XML. Compare every summary with the ruling before you use it.
How do you give an AI tool the right version of a statute?
You give an AI tool the right version of a Dutch statute by retrieving the text yourself from wetten.overheid.nl and pasting it into your instruction. On that site you can look up legislation as it applies today or as it applied on an earlier date, which matters for a case from the past. Then ask the model to work only with that text. Note that, according to its own disclaimer, wetten.overheid.nl is not an official publication.
