AI for lawyers supports drafting, document analysis and legal research, but never takes over final responsibility. The lawyer remains liable for every document that leaves the office, and since 2026 the Netherlands has had a clear example of what happens when that goes wrong: a reprimand for a lawyer who copied case law made up by AI into her written submissions.
The gain lies in time saved on routine work, the risk in output that sounds plausible but is factually wrong.
What can AI do for lawyers?
AI for lawyers works best for work that processes a lot of text and requires little independent legal judgement. A model speeds up three kinds of tasks, provided the lawyer checks the result personally.
Drafting first versions is the most obvious use. Within minutes, a model produces a rough version of an agreement, a memo or a letter to the opposing party, which the lawyer then refines for the specific case.
Document analysis is the second area. You can have long contracts and court documents summarised and searched for provisions or contradictions, and then go straight to the relevant passages.
Client communication is the third use. A model prepares answers to recurring questions or explains a legal position in accessible terms, after which the lawyer checks the tone and legal precision.
How do you use AI for legal research?
To use AI for legal research, let the model search and summarise, and verify every reference yourself in the official databases. In minutes, a model searches through what would take hours by hand, but gives no final view on the legal status of a judgment.
For searching and summarising long case files I use Claude myself, because in my experience it makes up fewer sources and figures for factual questions than ChatGPT. The difference from ChatGPT matters more the longer a file is, but it does not make checking unnecessary: making things up less is not the same as never making things up.
A useful distinction is between finding and verifying. The model produces candidates, such as judgments, statutory provisions and literature; the lawyer establishes whether they exist, are correct and apply.
What happens when AI makes up case law?
In the Netherlands, AI that makes up case law has now earned a lawyer a disciplinary measure. According to the Dutch legal news site Advocatie (30 July 2026), the Disciplinary Council in ‘s-Hertogenbosch reprimanded a lawyer who, in two tenancy cases, filed submissions containing eight references to case law taken from a legal AI tool (ECLI:NL:TADRSHE:2026:93).
One of those references was an ECLI number that does not exist; the other seven pointed to judgments that had nothing to do with the case, including divorce and criminal cases. The lawyer admitted she had not checked the sources. Besides the reprimand, the council ordered her to pay costs of 1,250 euros and took into account that the documents had been drafted in July 2025, when AI literacy in the profession was still developing. That mitigating circumstance carries less weight with every month that passes.
The mechanism behind it is called an AI hallucination: a model generates judgments, ECLI numbers and statutory provisions that sound convincing but do not exist. A recurring pattern I saw with ChatGPT was that it cited sources and quotations that turned out not to exist when checked. In the United States, lawyers were already sanctioned in 2023 after a court filing containing judgments made up by ChatGPT, in the case Mata v. Avianca.
What do bar associations recommend on AI? The Dutch example
The recommendations on AI from the Netherlands Bar (Nederlandse orde van advocaten, NOvA) make the five core values of the legal profession the test for any use of AI. The NOvA published them in December 2025, drawn up with its Digitalisation & AI project group, and they place responsibility squarely with the lawyer. If you practise in another country, check your own bar’s guidance alongside them.
The five core values are independence, partiality, expertise, integrity and confidentiality. For every task you leave to a model, you establish which core value is at stake: a summary of public case law does not touch confidentiality, sharing a client file does.
Three recommendations can be applied directly:
- Checking research output, where according to the NOvA “every reference, every footnote and every citation” is checked manually.
- Checking contracts when choosing a supplier, with attention to data ownership, liability and exit options.
- Transparency towards clients, by publishing the firm’s AI policy and informing clients about the use of AI.
For AI policy in the legal sector, that means a shift: the policy is no longer just an internal memo, but a document clients are entitled to know about. Dutch civil-law notaries received a comparable assessment framework with the KNB AI-weegschaal from the Royal Dutch Association of Civil-law Notaries (KNB).
How do you protect client data when using AI?
To protect client data when using AI, set down which data never goes into an AI tool, and with which account and under which agreements the rest may. The duty of confidentiality applies in full, including when data ends up in an external system.
The key point is what happens to the input. With free consumer versions of many tools, information you enter can be used to improve models, so confidential information slips out of the firm’s control. Business plans usually stipulate that input is not used for training; exactly which arrangement applies is set out in the supplier’s terms and data processing agreement.
Anonymise sensitive documents where possible before you enter them. Remove names, file numbers and identifying details, so the substantive question remains without the client being recognisable. Where an application processes personal data on a structural basis, it needs a DPIA, the risk assessment the GDPR requires for such processing.
How do you build in a four-eyes check for AI output?
To build in a four-eyes check for AI output, fix a set moment in the workflow at which a lawyer reviews the legal substance, the sources and the tone before a document goes to a client or a court. The tool does not determine the quality; the check behind it does.
- Legal substance
- Sources, checked in official databases
- Tone
Nothing goes out of the door without a second check.
From my years at major banks I bring the four-eyes principle into AI work: nothing goes out of the door without a second check. At a law firm, that means in practice that no AI output becomes a court document, submission or piece of advice unseen, and that every reference has been checked in an official database.
How thorough the check needs to be depends on the risk. An internal summary of public judgments needs less than a submission going to court. Since February 2025, Article 4 of the EU AI Act has also required a firm to take measures that support its staff’s AI literacy, worded since 27 July 2026 as an obligation of effort.
How do you introduce AI in a controlled way at a law firm?
To introduce AI in a controlled way at a law firm, start with one practice area and one task, and set down the way of working and the checking step there. One fully worked-out process delivers more than a broad rollout nobody owns.
Choose, for example, drafting a standard agreement within corporate law, and set down in an internal policy which tool is allowed for it, which data may go in and who checks the output. That policy also serves as the AI policy the NOvA recommends publishing to clients. In the AI Literacy Course you learn to work those steps out into a policy with checkpoints that match the core values.
AI for lawyers only becomes a shared way of working when everyone at the firm uses the same principles. With AI literacy for teams, everyone takes the same course, with their own account and a certificate in their own name.
Frequently asked questions about AI for lawyers
Besides disciplinary law and core values, AI for lawyers also involves the AI Act and the agreements you make with clients.
Is AI for lawyers a high-risk system under the AI Act?
AI for lawyers does not automatically fall under the high-risk systems of the AI Act. Annex III, point 8(a) covers AI systems intended to be used by or on behalf of a judicial authority to research and interpret facts and the law and to apply the law, or used in a similar way in alternative dispute resolution. The intended purpose of a tool determines whether it falls under this, so assess each application separately.
As a lawyer, do you need your client’s consent to use AI?
Not for each use separately. The NOvA recommendations mention client consent for AI use in the case file, but according to the NOvA’s explanatory notes you do not need to ask for separate consent for every use of AI. Transparency in advance is needed, through the engagement letter or a published AI policy, and in principle you respect a client’s objection. The NOvA also advises against entering confidential data or client data into free or public AI models, and recommends checking how suppliers handle data.
