Every September, a new class of first-year associates comes through my research orientation, and every year at least one of them shows me a printout of a chatbot answer as if it were a finished memo. It usually looks great. It has headings, a confident conclusion, and three or four case citations. Sometimes one of those cases does not exist.
AI legal research is genuinely useful, and I teach it now instead of warning people away from it. But the associates who get the most out of it are the ones who treat the AI as a fast, well-read research partner whose every sentence still has to be checked against the source. The ones who get burned treat it as an oracle.
So rather than give you a list of principles, I am going to walk you through one research problem from the moment the assignment lands to the moment you can defend your answer to a partner. Same question, same tool, every prompt shown. You can lift the method and apply it to almost any question you are handed.
TL;DR: To use AI for legal research safely, pin down the jurisdiction, the date, and the exact question before you type anything; write a prompt that gives the AI your facts and asks for primary authority; then verify every citation, quotation, and effective date against the source before you rely on it. A tool that grounds its answers in real case law and statutes, such as LegesGPT with its verified citations and Deep Research mode, shortens the work, but the verification step is still yours.
The assignment we will follow
Here is the problem. It is a composite of the kind of question I see junior associates get in their first month, with the names changed and a few facts simplified.
From: Partner Re: Summit Ridge Medical Supply, departing sales manager
Our client, Summit Ridge Medical Supply, is a Denver distributor of surgical equipment. Its regional sales manager, Dana, resigned last Friday and starts Monday at a competitor in Colorado Springs. When Dana was promoted in March 2024, she signed a promotion letter. Paragraph 14 of that letter is a 12-month non-compete covering Colorado, with a Delaware choice-of-law clause. Her base salary is $105,000 and she earned about $40,000 in commissions last year. The client wants to send a cease-and-desist letter this week. Can they enforce the non-compete? Short answer by tomorrow.
Read that twice. A lot of research mistakes happen before the first search, because the researcher never decided what the question actually is.
Before you open an AI legal research tool, write three things down
I make associates fill out a three-line index card before they touch any database, AI or not. It takes two minutes and prevents most of the wrong answers I see.
1. The jurisdiction, stated precisely. Not "US law." Here it is Colorado, and specifically Colorado statutory law on restrictive covenants, with a possible conflict-of-laws issue because of the Delaware clause. Federal law is not the main event, although you would want to note that the FTC's attempted nationwide non-compete rule never took effect.
2. The relevant date. Non-compete law in many states has changed fast, and Colorado is a prime example. The date that matters is when Dana signed (March 2024), plus today's date, because some requirements apply both at signing and at enforcement. Write both down.
3. The question in one sentence, and what "done" looks like. "Is the non-compete in paragraph 14 enforceable against Dana under Colorado law, and does the client face any risk by trying to enforce it?" Done means: the controlling statute identified and read in its current form, each element applied to our facts, open issues flagged, and every authority checked.
Notice that the second half of that question (risk to the client) is not in the partner's email. It is there because a good researcher asks what could go wrong with the action the client wants to take, not only whether the client is right.
How to write a legal research prompt: before and after
Most bad AI answers I am shown came from bad prompts. The model filled the gaps the user left with its best guess, and its best guess was a generic, national, slightly out-of-date summary.
Here is the prompt most first-years type:
Before: Are non-competes enforceable in Colorado?
There is nothing wrong with that question for a curious layperson. For a research assignment it is nearly useless. It has no date, no facts, no indication of what kind of authority you want, and no request to show sources. You will get a paragraph that is probably accurate in broad strokes and tells you nothing about Dana.
Here is the prompt I would want to see:
After: I am researching Colorado law on employee non-compete agreements. Facts: an employee of a Denver medical-supply distributor signed a 12-month non-compete covering Colorado in March 2024, as paragraph 14 of a promotion letter. The letter has a Delaware choice-of-law clause. Her base salary is $105,000 and she earned about $40,000 in commissions last year. She resigned in October 2026 to join a competitor in Colorado. Our client (the employer) wants to send a cease-and-desist letter.
- Identify the controlling Colorado statute and the version that applies to an agreement signed in March 2024.
- List each requirement for an enforceable non-compete under that statute, and apply each one to these facts.
- Explain whether the Delaware choice-of-law clause will be honored.
- Identify any penalties or liability the employer could face for attempting to enforce a void non-compete.
- Cite primary authority (statute sections and cases) for each point, with links, and tell me where the law is unsettled.
The differences are not cosmetic. Here is what each change does.
| What changed | Why it matters in the answer |
|---|---|
| Named the jurisdiction and area of law | Stops the model from blending in other states' rules or a national "general rule" |
| Gave the signing date and today's date | Forces the model to deal with which version of the statute governs |
| Supplied the real facts (pay, form of agreement, choice of law) | Lets the answer apply the law instead of reciting it |
| Broke the question into numbered sub-questions | Makes it harder for the model to skip an element, and easier for you to check each one |
| Asked about risk to the client | Surfaces issues the partner did not ask about but needs to know |
| Required primary authority, links, and unsettled points | Gives you something to verify and invites the model to admit uncertainty |
A few other habits that make prompts better. State your role ("I am an associate preparing a short answer for a partner") so the register fits. Ask for counterarguments in a separate prompt rather than the same one, because a single answer tends to pick a side and defend it. And never paste in client-identifying details you do not need; "Dana" and "Summit Ridge" add nothing to the legal analysis, which is why they are not in the prompt above.
First pass: what the AI answer looked like
For this walkthrough I am using LegesGPT, which answers legal questions with citations and source links you can click through, and searches case law and statutes rather than relying on whatever a general chatbot remembers from training. That choice matters for this particular question, for a reason you will see in a moment.
Run the "after" prompt through a tool grounded in sources and a good first answer follows the structure you asked for: statute, version, elements, choice of law, liability, open issues. In substance, it should tell you something like this:
- The controlling provision is Colo. Rev. Stat. § 8-2-113, which was rewritten by House Bill 22-1317, effective August 10, 2022. Because Dana signed in March 2024, the rewritten version governs.
- Under that version, most non-competes are void. The main exception requires that the worker earn at least the "highly compensated worker" threshold set each year by the Colorado Department of Labor and Employment, that the covenant be for the protection of trade secrets, and that it be no broader than reasonably necessary to protect the employer's legitimate interest in those trade secrets.
- The employer must give the worker a separate written notice of the non-compete, in clear and conspicuous terms, signed by the worker. For a current employee, that notice has to come at least 14 days before the covenant (or the raise or promotion that pays for it) takes effect.
- For a worker who primarily lived or worked in Colorado at the time of termination, Colorado law governs enforceability regardless of a contrary choice-of-law clause, and the employer cannot require the worker to litigate enforceability outside Colorado.
- An employer that presents or tries to enforce a non-compete that is void under the statute can be liable for actual damages, a $5,000 penalty per worker, and attorney fees.
- Open issue: whether this covenant is truly limited to protecting trade secrets, which depends on facts we do not have.
That is a strong first pass. It is also, at this stage, worth exactly nothing to the partner, because I have not checked any of it.
Here is the reason the tool choice matters. A general-purpose chatbot answering the bare "before" prompt from older training material can easily describe Colorado law as it stood before 2022: non-competes broadly void, with a list of exceptions that included one for "executive and management personnel." That exception no longer exists in that form. The answer was not invented; it was stale. Stale law is the quieter cousin of the hallucinated case, and in my experience it burns more people because nothing about it looks wrong.
Reading the answer like a skeptic
Before verifying, sort every claim in the answer into one of four buckets. I have my associates literally mark up the printout.
- Statutory text. Claims that should be traceable to specific language in § 8-2-113. These are the easiest to verify and the most important.
- Case law. Claims that a court has held something. Each needs the case pulled and the relevant passage read.
- Agency material. Here, the compensation threshold, which comes from the state labor department, not from the statute itself.
- Analysis or inference. The AI's application of the law to Dana's facts. This is the part you should trust least and rewrite most, because it is your professional judgment the partner is paying for.
Doing this for our answer shows something useful: almost everything load-bearing was statutory or agency material. That tells me where to spend my verification time. It also tells me that the one genuinely open question (the trade-secret purpose) will need facts from the client, not more searching.
The verification checklist
This is the table I hand out at orientation. It works for any AI legal research answer, and the right-hand column shows what each check turned up on the Summit Ridge problem.
| Check | What you are looking for | How to do it | What it means for Summit Ridge |
|---|---|---|---|
| Every citation exists | The case or statute section is real and the cite is correct | Click the source link, or pull the authority in an official or commercial database | Each statute subsection should resolve to real text; any case cite gets pulled |
| The source says what the answer claims | Pin cites, holdings, and paraphrases match the actual text | Read the cited passage itself, not the AI summary of it | The notice requirement is stricter than "written notice": it must be a separate document, signed by the worker |
| Quotations are verbatim | No smoothed-over or merged quotes | Compare word for word | Any quoted statutory phrase gets compared to the enacted text before it goes in the memo |
| Right version, right date | The text in force when the agreement was signed (and now) | Check the session law and effective date; check for later amendments | Confirmed the 2022 rewrite applies; noted 2025 amendments (SB 25-083) dealt with health care workers and sale-of-business deals, not our facts |
| Numbers are current | Thresholds, deadlines, penalty amounts | Go to the issuing agency's own publication for the relevant year | Threshold changes every January; confirmed the figure for 2024 (signing) and 2026 (enforcement) |
| Still good law | No reversal, overruling, or negative treatment | Run every case through a citator | Not needed for the statute; would be required for any case relied on |
| Nothing obvious is missing | Exceptions, defenses, related provisions | Read the whole statute section, not just the subsections cited | Found the non-solicitation provisions, which have a lower pay threshold and might be a fallback |
| Analysis fits the facts | The AI did not assume facts you never gave it | Reread the application against the assignment | Watch for any sentence assuming notice was given; we have no basis for that |
Two of those rows are worth slowing down on.
The "numbers are current" row is where most real-world AI research mistakes I see actually live. The Colorado threshold was $101,250 when the 2022 rewrite took effect, and the labor department has raised it each January since; for 2026 it is $130,014. A model or a blog post that is a year or two old will quote the wrong figure with full confidence. Always go to the agency's own published order for the year you need. In our problem, the threshold must be met both when the covenant is signed and when it is enforced, so you need two numbers, not one.
The "analysis fits the facts" row is where AI answers most often overstep. Our facts say the non-compete was paragraph 14 of a promotion letter. Nothing says a separate notice was given. An answer that applies the law smoothly can paper over that gap without saying so. In the memo, it becomes the most important open question.
Following up: the second and third prompts
Good AI legal research is a conversation, not a single query. After verifying the first pass, run follow-ups, each narrow.
Prompt two, on the facts we were missing:
Under C.R.S. § 8-2-113 as amended in 2022, does a non-compete clause included as a numbered paragraph inside a promotion letter satisfy the notice requirement, or must the notice be a separate document? Quote the statutory language and cite any Colorado cases or agency guidance applying it.
What you want back is the separate-document language and the signature requirement, quoted, plus any authority on clauses buried inside other documents. Read the statutory text yourself to confirm. That absence of contrary authority is itself a finding, but it is a weaker one than a holding, so the memo says "the statute's plain language" rather than "courts have held."
Prompt three, on compensation:
For purposes of the highly compensated worker threshold in C.R.S. § 8-2-113, how is "annualized cash compensation" defined, and do commissions count toward it? Distinguish statutory text from secondary-source interpretation.
This is a good example of asking the tool to label its own sources. The statutory definition refers to gross salary, wage, fee, or other compensation; several practice guides read that to include commissions and nondiscretionary bonuses. With $105,000 in base pay and about $40,000 in commissions, Dana is over the 2024 and 2026 thresholds if commissions count, and under them if they do not. That makes the definition load-bearing, so it goes in the memo with the source of each reading identified.
Prompt four, a broader sweep in Deep Research mode:
Find Colorado decisions since August 2022 applying the trade-secret requirement in C.R.S. § 8-2-113(2)(b) to sales employees, and summarize what employers had to show. Note any federal decisions applying Colorado law.
This is where a longer, multi-step search earns its keep. LegesGPT's Deep Research mode runs a more thorough pass across case law and statutes and returns a structured report with sources, which is what you want when the question is "what is out there" rather than "what does section X say." Treat the output as a reading list, not a conclusion: every case on it gets pulled, read, and run through a citator before it goes near the memo.
One more prompt I recommend on almost every assignment, run last:
Assume you represent Dana. What are the strongest arguments that this non-compete is unenforceable and that the employer faces liability for trying to enforce it?
Asking the tool to argue the other side is the fastest way to find the hole in your own analysis.
Where the answer landed
After the checklist and the follow-ups, here is roughly what I would expect a careful associate to send the partner. It is short, because the partner asked for a short answer, and every sentence in it traces to a source the associate actually read.
Short answer: Probably not, and sending a cease-and-desist letter carries real risk.
Dana's non-compete is governed by C.R.S. § 8-2-113 as rewritten in 2022, and the Delaware choice-of-law clause will not change that because Dana lived and worked in Colorado. The statute voids non-competes unless several conditions are met. Two look problematic. First, the statute requires a separate, signed notice of the non-compete given at least 14 days before the promotion took effect; the covenant appears only as paragraph 14 of the promotion letter, and we have no indication a separate notice was given. Second, the covenant must be for the protection of trade secrets and no broader than necessary; we need facts from the client on what confidential information Dana had. Whether Dana meets the compensation threshold depends on whether commissions count as "annualized cash compensation," which several practice guides read the definition to include.
If the covenant is void, the statute exposes an employer that tries to enforce it to actual damages, a $5,000 penalty, and attorney fees. I recommend we confirm the notice question with the client before any letter goes out, and consider whether a customer non-solicitation covenant (which has a lower pay threshold) is a better route.
Look at what changed between the AI's first pass and this. The structure is similar. The substance is not: the notice gap is now the lead issue, the compensation question is flagged honestly, and the recommendation protects the client from a step that could cost it money. That is the value a researcher adds, and AI does not remove the need for it.
How the method changes for other kinds of questions
The Summit Ridge problem is mostly statutory, which makes verification relatively clean. The same five moves (define, prompt, sort, verify, follow up) apply elsewhere, with different emphasis.
Common-law questions (for example, the elements of tortious interference in a particular state) put almost all the weight on case law. Expect to spend most of your time pulling cases, reading the holdings rather than the headnotes, and running a citator on each one. Ask the tool for the leading case and the most recent appellate case separately, since they are often not the same.
Procedural questions (deadlines, filing requirements, local rules) are where stale information is most dangerous and where the answer is often in a court's local rules or standing orders, which AI tools may cover unevenly. Verify against the court's own website every time.
Multi-state surveys are where AI legal research tools save the most time and where errors multiply. Ask for each state separately, or at least ask for a table with a citation per state, and spot-check a sample before you trust the pattern.
Habits I drill into every first-year
Most of what goes wrong with AI legal research is not exotic. It is the same handful of habits, repeated.
Pasting an answer into a work product before reading the sources. The 2023 Colorado discipline case People v. Crabill involved a lawyer who filed a motion citing cases from ChatGPT he had not read. Some were fictitious. He was suspended. The lesson is not "never use AI"; it is "never cite what you have not read."
Asking one giant question. Break the problem into sub-questions, as in the "after" prompt. Each one is easier to answer well and easier to check.
Accepting silence as an answer. If you ask about penalties and the answer does not mention any, that is not proof there are none. Read the whole statute section.
Ignoring dates. Every answer about statutory law should prompt the question: which version, effective when, amended since?
Trusting the application more than the law. The AI is often right about the rule and wrong about how it applies to your facts, usually because it filled a factual gap with an assumption. Your job is to find those assumptions.
Leaving client details in prompts. Strip names and identifying facts unless they matter legally, and know what your tool does with your inputs before you use it on client work.
Choosing a tool for this kind of work
For the walkthrough I used LegesGPT because it does the things this method depends on: answers with verified citations and source links, case law and statute search in the same place, and a Deep Research mode for the broader sweeps. It is web-based and starts at $19.99 a month, with a $1 three-day trial, which makes it practical for solo lawyers, small firms, and students who do not have an enterprise research subscription. If your firm already pays for Westlaw or Lexis, their AI layers are worth testing on the same problem; if not, a tool like this can cover most day-to-day research questions.
Whatever you pick, test it the way this guide does: give it a question you already know the answer to, including a recent statutory change, and see whether it catches the change and cites the source. If you want to compare options, start with our side-by-side look at AI legal research tools, and the guide to legal research databases covers the traditional platforms you will still verify against. Firms looking to move off an incumbent can look at the Westlaw alternatives we have reviewed, and students building these habits early should see our list of AI tools for law students. For the bigger picture of where research fits among other legal AI uses, see our overview of AI for legal work.
The short version, for your index card
The tools have changed my job less than people expect. I still teach first-years to define the question, find primary authority, read it, and confirm it is current. AI makes the finding faster, sometimes dramatically so. It does nothing for the reading and the confirming, and those are the steps that keep your name off a sanctions order.
Write the jurisdiction, the date, and the question on a card. Prompt with facts and ask for sources. Sort the answer by type of claim. Run the checklist. Follow up narrowly, then argue the other side. If you do that every time, AI legal research stops being a risk and becomes the most useful thing on your desk.
Frequently Asked Questions
How do I use AI for legal research without citing fake cases?
Treat every citation the AI gives you as a lead, not an authority. Pull each case or statute from the source, read the passage the answer relies on, and run cases through a citator before they go into anything you file or send. Tools that link each claim to its source make this faster, but they do not remove the step.
Is AI legal research accurate enough for client work?
It is accurate enough to speed up the finding part of research, especially when the tool grounds its answers in real case law and statutes. It is not accurate enough to skip verification. The most common errors are outdated law, figures that change every year, and confident application of the law to facts you never provided.
What should a good legal research prompt include?
The jurisdiction, the relevant dates, the key facts, the specific sub-questions you need answered, and a request for primary authority with links and a note of where the law is unsettled. Leave out client names and identifying details that do not affect the legal analysis.
Do I still need Westlaw or Lexis if I use an AI legal research tool?
For many day-to-day questions, a tool that answers with verified citations and searches case law and statutes can cover what a solo lawyer or small firm needs. For heavy litigation work you still need a reliable way to check whether cases remain good law, so many lawyers pair an AI tool with a citator or traditional database.
How do I check whether an AI answer reflects the current version of a statute?
Look up the session law that last amended the section and its effective date, then confirm which version applied on the date that matters to your problem, such as when a contract was signed. For numbers set by agencies, like salary thresholds or filing fees, go to the agency's own publication for the year you need.
Is it ethical for lawyers to use AI for legal research?
Yes, as long as you meet the same duties you always have: competence, confidentiality, and candor to the court. That means understanding what the tool does with your inputs, keeping client confidences out of prompts unless the tool is approved for them, and personally verifying every authority you rely on.
What is Deep Research mode useful for in legal research?
It suits open-ended questions such as finding every decision applying a statute to a certain kind of employee, where a quick answer would miss things. LegesGPT's Deep Research mode runs a longer, multi-step search across case law and statutes and returns a structured report with sources, which works best as a reading list you then verify.
