What AI legal research actually means in 2026
Legal research is the highest-stakes use of AI inside a law firm. Get it wrong and your brief cites cases that do not exist. Get it right and a solo attorney does the research that used to take a junior associate a full day, in about an hour, with better synthesis at the end.
The category has split into two very different kinds of tool. On one side, citation-grounded research products from the legal data vendors (Lexis, Thomson Reuters, Clio). They search a real case database, verify the citations, and ground the AI summary in actual law. On the other side, general-purpose models like ChatGPT, Claude, and Gemini, which can read and summarize cases you give them but cannot reliably look up case law on their own. Both have a place. They do different jobs.
This guide is for the lawyer running a 1- to 25-person practice who wants to use AI for research without ending up in the next Mata v. Avianca. We’ll walk the actual workflow, name the tools that fit a small-firm budget, and end with the questions a partner should ask before letting the firm rely on any of this. For broader context on AI in the practice, see the pillar guide on AI for law firms.
The five-step legal research workflow
The output of legal research is a memo (or a brief section, or a quick advice email). The job of an AI legal research tool is not to write the memo for you. It is to compress the steps before the memo: framing the question, finding the cases, verifying them, and synthesizing the holdings.
1. Frame the question precisely
The single biggest determinant of whether AI research saves you time is how tightly you frame the question. “What does the law say about non-compete agreements” returns noise. “Whether a six-month, fifty-mile non-compete signed by a salaried sales representative is enforceable in Florida after the 2024 amendment to Fla. Stat. § 542.335” returns case law you can actually use.
I run this framing step in a general model (ChatGPT or Claude) before touching the legal database. It costs nothing and saves database query time later.
Act as a senior legal researcher. I am preparing a memo on the following question: [ROUGH QUESTION IN PLAIN ENGLISH] The jurisdiction is [STATE OR FEDERAL CIRCUIT]. The relevant facts are: [3-5 RELEVANT FACTS] Reformulate this as a precise legal question or a set of sub-questions a court would actually decide. Identify the controlling statute or rule if there is one. List the elements I will need to address.
The output is the question you should be searching for, broken into sub-issues. Paste that into your citation-grounded research tool in the next step.
2. Run the citation-grounded search
This is where you switch tools. Take the framed question into a research product that searches a real case database. The three small-firm-relevant options in 2026:
- Lexis+ with Protégé (the product formerly known as Lexis+ AI; renamed in February 2026)
- Westlaw Precision AI
- CoCounsel (Thomson Reuters; absorbed Casetext in 2023)
Pricing tiers and exact features change quarterly, so call the vendor for current numbers. As of mid-2026, all three sit in roughly the $100 to $300 per-attorney per-month range for the AI tier on top of a base subscription. If your firm already has a Lexis or Westlaw subscription, the AI add-on is the cheaper path. If you have neither, CoCounsel sells direct without requiring the full Westlaw stack.
Ask the tool a focused question and let it return both a synthesized answer and the underlying cases. Do not accept the synthesized answer as your research output. Treat it as a reading list.
3. Verify every citation before you read it
Even a citation-grounded tool can return a case that exists but does not stand for the proposition the AI summary attaches to it. The verify step takes five minutes and protects you from the kind of sanction that ended careers in 2023.
For each case the tool surfaces:
- Confirm the case exists by pulling it up directly in the database (not by clicking the AI summary’s link).
- Read the headnote or the procedural posture section to confirm the holding matches what the AI told you.
- Shepardize or KeyCite the case for negative treatment.
If a case does not check out, drop it. Do not let it through to the next step.
4. Synthesize the holdings into a draft memo
Once you have a verified short list of cases, the synthesis step is where general-purpose models earn their place. Paste the verified cases (full text, not summaries) into ChatGPT, Claude, or Gemini, with a memo prompt:
Act as a legal research associate. Draft a research memo on the following question: [QUESTION FROM STEP 1] Use ONLY the cases I provide below. Do not introduce new citations. If a case I provide does not address the question, say so. For each case, identify: (1) the holding relevant to the question, (2) the procedural posture, (3) any limiting language. Then synthesize the cases into a position our firm can take, noting any conflicts in the authority. Output structure: 1. Question presented 2. Short answer 3. Discussion (organized by sub-issue, not by case) 4. Conclusion Cases: [PASTE CASE 1 FULL TEXT] [PASTE CASE 2 FULL TEXT]
The model drafts the memo. You edit it. Keep your judgment in the loop, especially on the short answer.
Confidentiality note before you paste. Free-tier ChatGPT, Claude, and Gemini may use inputs to train future models. For privileged research that involves any client identifier, use enterprise tiers like ChatGPT Team, Claude for Work, or Microsoft Copilot for Business, where data is excluded from training. For pure case-law synthesis with no client facts, the free tiers are fine.
5. Pressure-test against opposing authority
The final step is the one most lawyers skip and most AI tools will not do unless asked. Run a separate query for cases that cut the other way:
I have drafted the memo above taking the position that [STATE THE POSITION]. Now act as opposing counsel. Identify the strongest cases and the strongest legal arguments AGAINST this position. For each, give the citation, the holding, and a one-paragraph response we should be ready to make.
Run this in your citation-grounded tool first (so the citations are real), then use a general model to draft the response paragraphs. This step turns a one-sided research memo into something you can actually walk into court with.
Tools to consider
Three tool categories cover most of what a small firm needs. The right answer depends on what you already pay for.
For citation-grounded research
- Lexis+ with Protégé. AI tier on top of the Lexis case database. Strong for jurisdictions where Lexis has deep coverage. Built-in citation verification. Pick this if you already pay for Lexis.
- Westlaw Precision AI. Thomson Reuters’ equivalent. AI on top of an existing Westlaw subscription. Pick this if you already pay for Westlaw.
- CoCounsel. Thomson Reuters product that absorbed Casetext in 2023. Sells direct, more memo-focused, useful if you do not have a full Westlaw subscription.
For framing, synthesis, and pressure-testing
- ChatGPT (Plus, Team, or Enterprise). The general-purpose default. Team and Enterprise tiers exclude inputs from training and are appropriate for client-adjacent work.
- Claude (Pro or for Work). Strong on long-document handling. Good when you are pasting in 80 pages of cases at a time.
- Gemini. Useful if your firm runs on Google Workspace and you want the research to live in Docs.
What I would not pay for as a small firm: Harvey. It is a strong product and it is built for the AmLaw 200 budget. The same workflow above runs on tools that cost a tenth as much.
What to watch for
Three risks come up in every small-firm engagement on this topic.
Hallucinated citations. The mistake that produced Mata v. Avianca is still possible in 2026, especially if you let a general-purpose model do raw case lookup. The Step 3 verification check is non-negotiable. The follow-on cases since Mata have produced fines, referrals to the bar, and one judge’s standing order requiring a certification that no AI was used in the brief. Do not become an entry on that list.
Out-of-jurisdiction or out-of-date law. AI tools can hand you a federal case for a state law question, or a 2018 case in an area where the rule changed in 2024. Your jurisdiction filter and your date filter matter as much as the AI prompt does. Check both.
Confidentiality. ABA Formal Opinion 512 (July 2024) reminded the bar that Model Rule 1.6 still applies to AI tools. Free-tier general models may train on your inputs. Use enterprise tiers for anything client-identifiable. When in doubt, redact before you paste, and confirm your malpractice carrier knows what you are using.
When to fall back to traditional research
AI legal research tools are good. They are not better than a careful Lexis or Westlaw query plus a hand-read of the top three cases for: novel issues with thin case law, very recent precedent the model may not have, and any question where the procedural history matters more than the holding. For those situations, the AI step adds noise. Run the search yourself and read the cases the old-fashioned way.
FAQ
Can I use ChatGPT alone for legal research?
Not for finding cases. ChatGPT does not have a verified case database; it can confidently invent citations. It is fine for framing the question (Step 1), synthesizing cases you have already verified (Step 4), and pressure-testing your position (Step 5). For raw case lookup, use a citation-grounded tool. The deeper take on this question lives in ChatGPT for legal research.
How much does an AI legal research subscription actually cost?
As of mid-2026, Lexis+ with Protégé and Westlaw Precision AI run roughly $100 to $300 per attorney per month for the AI tier on top of the base research subscription. CoCounsel sells separately and lands in a similar range. Pricing changes; call the vendor and ask for current per-seat numbers and any small-firm discount.
Will my firm’s bar discipline me for using AI?
Bar opinions are jurisdiction-by-jurisdiction. ABA Formal Opinion 512 sets the federal-baseline guidance: AI use is allowed, with duties of competence (Rule 1.1), confidentiality (Rule 1.6), supervision (Rules 5.1 and 5.3), and reasonable fee disclosure (Rule 1.5). State bars have layered on additional guidance. Confirm with your state bar and your malpractice carrier before you build a workflow that depends on AI output.
Is the AI version of my existing subscription worth it?
If you are already paying for Lexis or Westlaw, the AI tier is usually the right entry point. You get the new capability without standing up another vendor relationship. If you are starting from zero, weigh CoCounsel against the other two; it sells direct and does not require the full Westlaw stack.
What about deposition summaries and other research-adjacent work?
The same general workflow applies, with different tools. For depositions specifically, see AI deposition summary; for a head-to-head ranking of research products, see best AI legal research tools.
Related on Business AI Workflows
- AI for law firms: the pillar guide for solo and small practices.
- ChatGPT for legal research: when general-purpose models work and when they do not.
- AI deposition summary: workflow for compressing transcripts into usable summaries.
- Best AI legal research tools: head-to-head comparison of the small-firm options.
