How to Use AI for Legal Research: A Practical Guide for In-House Legal Teams

At a glance

• Use legal-specific tools for legal research: Even those get things wrong: a Stanford study found leading AI legal research tools hallucinated between 17% and 33% of the time.
• Name the jurisdiction and the date in every question.
• Read every answer in three layers: what the source says, what the tool concludes and what it assumed.
• Open every citation before it goes into a memo, board paper or filing.
• Tie research to your own records. For in-house teams, the answer depends on the law and on your articles, minutes and entity data.
• AI output is legal research, not legal advice: A qualified lawyer reviews and signs off.

What AI for legal research actually does

AI for legal research is the use of AI tools to find, read and summarize legislation, case law and legal commentary in response to a plain-language question. Good tools retrieve published sources first, then write an answer that cites them, so you can check each conclusion against the text it relies on.

That retrieval step is one distinction between a research tool and a general-purpose chatbot. A general model answers from patterns it learned in training. A legal research tool searches a defined body of law and then writes from what it found (a technique usually called retrieval-augmented generation, or RAG).

On easy questions, the difference may be small. On harder questions, the difference can be significant, especially when the answer depends on an obscure regulation, a recent appellate decision or a rule that changed last year.

Why a general chatbot is the wrong place to start

Consumer chatbots are very good at sounding like a lawyer. That’s the problem.

In Mata v. Avianca (S.D.N.Y., 2023), lawyers were sanctioned after filing a brief that cited cases ChatGPT had invented. They had submitted citations nobody had verified and stood by them after the court started asking questions.

Legal-specific tools do better. They still aren’t infallible. A preregistered Stanford RegLab study found that AI research tools from two major legal publishers each hallucinated between 17% and 33% of the time, even though they performed better than a general-purpose chatbot.

Then there’s confidentiality. In February 2026, Judge Rakoff of the Southern District of New York ruled in United States v. Heppner that a defendant’s exchanges with the consumer version of an AI chatbot were protected by neither attorney-client privilege nor the work product doctrine. A key factor was the platform’s privacy policy, which allowed the provider to use and disclose user inputs.

Heppner involved a client acting without his lawyer’s direction, and the analysis may differ for counsel using a tool under enterprise terms. The practical lesson still holds. If your team pastes a draft litigation strategy into a consumer tool whose terms let the provider use or share inputs, assume a court may treat that as disclosure to a third party.

How to use AI for legal research in 6 steps

This is the workflow we’d hand to a new in-house lawyer on day one. It works with any serious AI legal research tool.

1. Frame the question with jurisdiction, date and facts

Vague questions produce confident, vague answers. Compare these two prompts:

• Weak: “Are non-competes enforceable?”
• Strong: “Under French law, is a 2-year post-employment non-compete for a sales director enforceable without financial compensation? Cite the relevant legislation and Cour de cassation decisions from the last 10 years.”

Name the jurisdiction every time. If your tool asks you to confirm it before searching, that’s a good sign.

2. Pick the right source for the question

Two very different questions look identical in a chat box. What does the law say, and what do our own documents say?

The first needs published legislation, case law and doctrine. The second needs your contracts, minutes and policies. If you only need to find a clause, summarize a board pack or translate a document, an AI assistant that works directly on your own documents, such as Ask Lini (formerly DiliChat), gets you there faster. Save legal research for questions the law has to answer.

3. Read the answer in three layers

A usable answer separates three things: what a source says, what the tool concludes from it and what it assumed to get there.

Assumptions are where research goes wrong quietly. “Assuming the company is a société anonyme” is harmless if true. If the entity is actually an SAS, whose governance rules are set largely by its own articles, the whole analysis may shift.

If your tool blends all three layers into one confident paragraph, you’ll have to pull them apart yourself.

4. Verify every citation before it leaves your desk

Treat each citation as unconfirmed until you’ve done four things:

• Open the source. Read the passage the tool relies on, not just the citation..
• Check the date. Has the statute been amended, or the decision been overturned or limited since?
• Check the fit. Does the holding match your facts, or only your keywords?
• Log it. Note who checked which citation and when.

ABA Formal Opinion 512 is clear that the duty of competence covers your use of generative AI. You need a reasonable understanding of what the tool can and can’t do, and you can’t hand your professional judgment to it.

The same opinion notes that a tool designed for legal work may need less independent checking once your experience with it gives you a reasonable basis to rely on it. Less isn’t zero.

5. Add your company’s facts

This is where in-house research parts ways with law-firm research. The answer to “can our board meet by videoconference and still reach quorum?” depends on the statute. It also depends on your articles of association, your board’s internal rules and how you’ve run past meetings.

If your research tool can read your board minutes, entity records or matter files (within the access rights you already have), you can get an answer grounded in your company’s records. If it can’t, someone copies facts in by hand. That adds error risk and creates confidentiality exposure every time.

6. Turn the result into a memo you’d sign

Structure the output for the person who reads it 18 months from now, when an auditor or a new General Counsel asks why the decision was taken:

1. Question and jurisdiction
2. Short answer
3. Sources, with pinpoint citations
4. Analysis
5. Assumptions and open points
6. Reviewer, date and sign-off

Keep the disclaimer. Until a qualified lawyer has reviewed it, the memo is legal research and analysis, not legal advice.

Where AI legal research pays off for in-house teams

Time saved is the obvious benefit and the least interesting one. The real gain is making decisions without waiting for research to finish.

Situation	Research question	What changes for your team
Board meeting next week	What quorum and majority rules apply to a videoconference board meeting in France?	The corporate secretary confirms the rule against the company’s own meeting history before the vote, so the resolution is less exposed to challenge afterwards.
New director at a Belgian subsidiary	What filing obligations does the appointment trigger?	The entity team sees the obligation, its source and the current entity record side by side.
Commercial dispute in France	What does recent case law say about abruptly ending an established commercial relationship?	The litigation lead briefs outside counsel with the case law already gathered, so external hours go to strategy.

Notice what the three examples have in common. Each answer only becomes useful once it’s tied to a specific meeting, entity or matter.

How to test AI legal research tools before you rely on them

There’s no single best AI for legal research. The best tool is the one that gets your questions right, in your jurisdictions, under your data rules. Run these 5 tests during any trial:

1. Re-run 5 questions you’ve already answered: Compare the output with the memo outside counsel wrote.
2. Ask about a jurisdiction outside your home market: Does coverage hold up?
3. Set a trap: Ask about a repealed provision or an overturned decision and see if the tool catches it.
4. Check every citation in 2 answers by hand: Count the errors.
5. Read the data terms: Who can see your prompts, how long are they kept and are they used for training?

These tests help you assess whether a tool’s answers are reliable enough for your workflow. They don’t cover the full buying decision. For coverage, pricing, security reviews and vendor questions, see our guide to “choosing legal research software for an in-house team”.

Set your team’s rules before anyone starts

Some of your lawyers may already be using AI for research, approved or not. A one-page policy closes most of the gaps:

• An approved tools list, with the type of data allowed in each
• No confidential or privileged material in consumer tools
• A mandatory citation check before AI-assisted research is relied on
• A named reviewer for every memo
• A record of which tool produced which research
• Short training on writing precise questions and spotting assumptions

How Lini handles legal research in DiliTrust

Legal Research and Analysis with Lini is available in the Board Portal, Legal Entity Management and Matter Management modules. Lini searches published legislation, case law and legal doctrine across more than 230 jurisdictions. It asks you to confirm the jurisdiction when a question is ambiguous, checks every citation and lists the sources behind each answer.

What matters for review is how the answer is built. Lini separates what comes from a source, its own analysis and any assumptions it made, and lists the sources behind each point. The reviewing lawyer sees straight away which points are supported by a source and which need a second look. You can then turn the result into a memo or summary, and Lini reminds you that it supports professional judgment rather than replacing legal advice.

If you want to know how Lini is built and where your data is processed, read how Lini works across the platform.

Frequently asked questions