Introduction
DELTA is a public, practitioner-grade benchmark for Dutch legal AI, led by Legal Benchmarks with Zeno as Founding Research Partner. Its first release focuses on open-ended legal research. The benchmark and this companion survey address two related questions:
- How well do AI tools answer Dutch legal research questions?
- How do lawyers use those tools in their everyday legal work?
This report reflects the experiences of 115 legal professionals working in Dutch legal practice.
Who responded
- 11 practice areas represented: Corporate and M&A, technology and intellectual property, disputes, employment, real estate, finance, public law and other disciplines.
- Roles represented: Junior, mid-level and senior lawyers and partners, and legal operations, knowledge and innovation professionals.
- 19 organisations across the Dutch legal sector: Respondents came from organisations ranging from specialist boutiques and national practices to legal organisations with cross-border or international networks.
Key findings
- AI is now routine in Dutch legal research, especially among junior and mid-level lawyers. 63.5% of the legal professionals surveyed use AI for legal research every day.
- For most lawyers, using AI for legal work requires follow-up and correction. 67.8% said their first answer was usable half the time or less.
- Lawyers can work with an incomplete AI answer, but not one built on invented law. Incorrect or invented law was the error lawyers were least willing to accept.
- Lawyers are ready to outsource finding the law to AI, but not deciding what it means. 45.2% chose case law research as the first task to hand to AI.
Using AI for legal research has become a daily habit.
98.3% of lawyers now use AI at least once a week for legal research.
“How frequently do you use AI tools for legal research?”
- Daily
- 63.5%
- Several times a week
- 29.6%
- About once a week
- 5.2%
- A few times a month or less
- 1.7%
Each square represents one survey respondent.
Lawyers earlier in their careers use AI most often. Among the groups defined by years in practice, 73.0% of junior lawyers and 72.0% of mid-level lawyers use AI every day. Daily use falls to 45.0% among lawyers with eight or more years of experience.
The first answer is rarely the finished answer.
Most lawyers need to ask follow-up questions, correct the answer or add context before the result becomes usable.
“When using an AI tool, how often does your first prompt produce a usable result without revision or follow-up?”
- First answer not usually usable
- 67.8%
- First answer usually usable
- 32.2%
% of respondents
The first answer usually begins the research rather than completes it. Lawyers compare the result with their own understanding, challenge doubtful points and supply missing context. When the tool lacks important information, most lawyers want it to ask a question rather than make an assumption.
“The real risk is not simply that AI makes mistakes, but that lawyers handle its output too casually: adopting an AI-reviewed contract clause too quickly, failing to check a sentence because it sounds good, or treating a party’s argument in a judgment as the court’s decision.”
Invented law outranks every other failure.
Not every correction carries the same weight. Lawyers judge false law much more harshly than an omitted point.
“How severe is each of these AI failures in legal research?”
- Invented or incorrect law, facts or citations
- 82.6%
- Outdated, irrelevant or wrong-jurisdiction sources
- 72.2%
- Unsupported claims or no source
- 68.7%
- Missing issues, facts, arguments or authorities
- 53.9%
- Incorrect application or unsupported conclusion
- 41.7%
- Disorganised or overly long answer
- 20.0%
- Advice not tailored to the client's goals or risks
- 16.5%
Share of respondents who rated each failure as serious: 4 or 5 on a scale where 1 meant least severe and 5 meant most severe.
Invented law can undermine the whole answer; lawyers can still correct an omission. They check whether authorities exist, apply to the question and support the conclusion. A missing point leaves room for repair, but false law or an invented source makes the rest of the answer harder to trust.
“The error that made the greatest impression on me was an AI system confidently presenting a legal answer supported by a judgment that did not exist. The challenge is not simply that AI makes mistakes, but that some mistakes are difficult to recognise and presented with great conviction. To me, this highlights why human oversight and independent evaluation of legal AI are essential.”
Experienced lawyers set a higher quality bar for AI legal reasoning.
The largest difference between experience groups concerned what happened after AI found the relevant law. More experienced lawyers were more likely to treat incorrect application or an unsupported conclusion as a serious failure.
Percentage of junior versus senior lawyers who rated weak legal reasoning as a serious failure
- Junior0–3 years
- 21.6%
- Senior8+ years
- 60.0%
Rated 4 or 5 on a scale where 1 meant least severe and 5 meant most severe.
More experienced lawyers were more than twice as likely to view weak legal reasoning as a serious failure. 21.6% of junior lawyers held this view, rising to 52.0% of mid-level lawyers and 60.0% of senior lawyers.
The 38.4 percentage-point gap between junior and senior lawyers shows that source accuracy is only the starting point. As experience increases, lawyers place greater weight on whether an answer applies the law correctly and supports its conclusion.
“An AI tool should not replace a lawyer’s critical thinking, but complement, sharpen and challenge it.”
Lawyers would hand over the search to AI, not the judgment.
When asked what they would hand over first, lawyers concentrated on research, review and first drafts: work that finds, organises or prepares material before they apply it to the legal question.
“If AI could reliably take over one part of legal research tomorrow, what would you hand over first?”
- Case law research
- 45.2%
- Document and file review
- 21.7%
- Drafting and preparing written work
- 18.3%
Each square represents one of 115 respondents.
Lawyers most want AI to handle case law research, document review and first drafts. These tasks reduce the time spent finding and organising material. Lawyers still expect to decide how that material applies to the question and the client.
“The most serious error I encounter is cognitive outsourcing: handing over reasoning, problem-solving, or decision-making to AI. That failure usually sits with how the tool is used. A good AI tool strengthens the lawyer’s critical thinking and helps them take responsibility, rather than handing that responsibility over to AI.”
Looking ahead
When asked how their work might change, lawyers looked beyond today’s tools and imagined legal practice in 2041.
The strongest predictions were about speed and how legal work is organised. Lawyers expected research, drafting and document review to require less time and manual work. They also anticipated changes to billing, client expectations, training and everyday working methods.
In their own words, how legal professionals see practice in 2041.
“Legal headcount will drop, even though both the volume of deliverables and the number of new roles and their representatives will grow. What will remain is the need for critical thinking and attention to detail while working at scale.”
Maxim SvyatovLegal Operations, Knowledge or Innovation Professional, Loyens & Loeff
Key takeaways for legal teams
Many lawyers already use AI in daily legal research. Legal teams now need to make the path from an AI answer to verified legal work faster, clearer and more reliable.
Auditability and verification should become core criteria when evaluating products.
As tools become more capable and are trusted with more legal work, legal teams need to understand how an answer was produced and confirm it quickly. A tool that is easy to verify reduces the hidden work between a fast answer and a usable one.
Human review should follow the risk pattern revealed by the survey.
Review should begin with the legal foundation. Are the law and authorities real, cited accurately and capable of supporting the conclusion? It should then examine whether the research is complete and applicable. Has the answer missed an important issue, relied on an outdated source or used law from the wrong jurisdiction? This sequence distinguishes a missing point that can be added from a flaw that makes the whole answer unreliable.
AI training should teach lawyers an iterative research process.
Because important errors may not be obvious in the first answer, lawyers should challenge the interpretation, ask what may be missing and test key conclusions in a new conversation or another AI tool. Differences between the answers show where closer review is needed, but even agreement should be confirmed against the underlying authorities.
“Enthusiasm gets you your early adopters and structure gets you the firm. The people who love this will find the time and pass on what they learn. The rest will follow once the time is made available to them, which means paying for the learning with hours counted as billable, the same as any other client work. That is the point at which a firm starts to move.”
Acknowledgements
We thank everyone who completed the survey and shared their practical experience of using AI for legal research.
We also thank the participating firms, members of the legal technology community and individual contributors who reviewed the research and shared their expertise.
Contributing organisations
