The DELTA Benchmark

5 min read · from the paper

The work between an answer and usable advice

Among 115 legal professionals surveyed for the DELTA report, 63.5% used AI daily for legal research. Yet 67.8% said the first answer was usable without revision or follow-up only half the time or less. Frequent use and dependable first answers are different things.

That gap is the starting point for The DELTA Benchmark: Measuring Legal Taste in Artificial Intelligence. The report brings together an introduction to a practitioner-led benchmark for Dutch legal AI and the Dutch Legal AI Adoption Survey. One examines how to evaluate model answers. The other asks legal professionals what using AI is like in practice.

Read together, they make a practical case for evaluating the work that remains after an answer appears: checking authorities, testing assumptions, finding omissions and deciding whether the reasoning fits the matter.

What does “legal taste” mean?

A legally plausible answer can still leave the decisive issue buried in the fourth paragraph. It can quote the right authority without explaining why it matters, or present a conclusion confidently where the facts call for a qualification.

DELTA uses “legal taste” to describe observable professional judgement in the answer: recognising what matters, setting priorities, handling uncertainty and producing work a lawyer can use. The report translates that idea into three evaluation categories.

• Substance: are the conclusions legally defensible, are the material rules and issues addressed, and has the assignment been completed?
• Citation: are the necessary authorities cited accurately, and do they support the propositions attached to them?
• Form: is the answer clear, proportionate, appropriately qualified and organised around what the reader needs?

These categories give reviewers a way to explain why an answer succeeds or fails. Correctness remains essential, but a useful evaluation also examines the reasoning and the way it is presented for professional review.

Built around Dutch legal work

The report describes a benchmark built around legal research assignments from Dutch practice. Its first public tasks were selected from a growing collection of more than 200 assignments. Practising lawyers helped review the questions and evaluation criteria and scored model answers. The wider project involved 12 partner firms.

The distinction between the collection and the public release matters: the report does not say that all 200-plus assignments were released. Nor should the survey percentages be read as performance scores for a particular model or product. They describe respondents’ own experiences across their use of legal AI.

Legal Benchmarks leads the research and controls the methodology, analysis and published conclusions. Zeno participates as Founding Research Partner, contributing technical experience in legal AI, evaluation design and benchmark development. That division of responsibility is part of how the report presents the collaboration.

Daily use still involves follow-up

The survey covers 115 legal professionals across 11 practice areas and 19 organisations. Respondents include junior, mid-level and senior lawyers, partners, and people working in legal operations, knowledge and innovation.

Only 32.2% reported that their first prompt usually produced a usable result without revision or follow-up. For the remaining 67.8%, that happened half the time or less. These are self-reported experiences within the surveyed group, rather than an estimate of every Dutch lawyer’s experience.

The practical consequence is that the first answer often starts another round of work. A lawyer supplies missing context, narrows a question, challenges an interpretation or asks for the authority behind a conclusion. The report argues for making that process easier to inspect and verify, rather than treating the speed of the first response as the whole measure of usefulness.

The errors lawyers are least willing to accept

When asked to rate the severity of failure modes, 82.6% of respondents rated invented or incorrect law, facts or citations at four or five on a five-point scale. Outdated, irrelevant or wrong-jurisdiction sources followed at 72.2%, and unsupported statements or missing sources at 68.7%.

Those percentages measure how seriously respondents viewed the errors, not how often the errors occurred. They point to a review priority: establish that the legal foundation is sound before polishing the answer.

Experience also affected what respondents noticed. Weak reasoning or unsupported conclusions were rated as serious by 60% of senior respondents with eight or more years of experience, compared with 21.6% of juniors with up to three years. Checking that a source exists is one part of review. Assessing whether the conclusion follows from it calls on another part of a lawyer’s judgement.

Delegation begins with groundwork

Asked which task they would delegate first, 45.2% chose case-law research, followed by document review at 21.7% and drafting at 18.3%. These preferences help explain where legal teams see an immediate role for AI preparation.

The report keeps the professional’s role in view. Retrieved authorities, reviewed documents and draft text still need to be assessed in the context of the matter. The lawyer remains responsible for deciding what the work means and how it should inform advice, strategy or a next step.

A more useful way to evaluate legal AI

The report closes with three recommendations for legal teams: make auditability a core evaluation criterion, organise human review around the risks revealed by the survey, and teach an iterative research process.

In practice, that means starting with the authorities. Are they real, accurately cited and capable of supporting the conclusion? Then examine completeness and applicability: what has been omitted, what may be outdated, and what depends on the jurisdiction or the facts? Follow-up questions and comparisons between answers can reveal where closer review is needed. Agreement between tools still needs to be checked against the underlying sources.

DELTA offers a vocabulary for that assessment. Substance, citation and form turn a general impression of quality into questions that can be discussed, tested and revisited. The useful question for a legal team is how much well-founded, reviewable work an answer provides, and what it still asks the lawyer to resolve.

About the report

The DELTA Benchmark: Measuring Legal Taste in Artificial Intelligence is an English digital report by Legal Benchmarks, with Zeno as Founding Research Partner. The August 2026 edition contains 30 PDF pages, including the benchmark introduction, survey findings, practitioner perspectives and references. Download the report to examine the findings in their original context.