The best AI legal research tools split into three tiers: dedicated research platforms (CoCounsel, Lexis+ AI, Westlaw Precision) built on real case databases, litigation analytics (Lex Machina, Trellis) that predict behavior, and general AI (Claude, ChatGPT, NotebookLM) for orientation. Most small firms should start at the cheapest tier and upgrade only when they hit its ceiling weekly.
Before I rank a single tool, one warning, because it matters more than anything else on this page.
Lawyers have been sanctioned in federal court for filing briefs with citations that AI invented. Mata v. Avianca in 2023 was the famous one, and courts have kept catching new instances since. The cases looked real. The quotes looked real. None of them existed. Judges do not care that the AI sounded confident.
So here is the rule at my firms, and it applies to every tool below, at every price point: any AI research output gets verified against the actual source before it touches a filing. You pull the case. You read the passage. Non-negotiable. The tools differ in how often they’re wrong, not in whether verification is required.
With that settled, here’s how the market actually breaks down. Three tiers, three different jobs.
| Tier | Tools | The job | Typical cost |
|---|---|---|---|
| Dedicated research AI | CoCounsel, Lexis+ AI, Westlaw Precision | Find and check the law | Hundreds per seat per month |
| Litigation analytics | Lex Machina, Trellis | Predict judges, opposing counsel, outcomes | Firm subscription, varies |
| General AI | Claude, ChatGPT, NotebookLM | Orientation and drafting | Free to $20 per month |
Tier 1: dedicated legal research AI
These are the tools built on real case databases, which is what separates them from a chatbot guessing.
CoCounsel is Thomson Reuters’ offering, built on the foundation Casetext developed before the $650 million acquisition. It’s connected to Westlaw content, so when it answers a research question it’s pulling from an actual, maintained body of law. It handles research memos, deposition prep, and document review, and it’s the most polished of the group.
Lexis+ AI is LexisNexis’ answer, now fronted by their Protégé assistant. Its strongest card is Shepard’s integration: the tool checks whether the cases it hands you are still good law as part of the workflow. That closes one of the more embarrassing failure modes, citing a case that got reversed.
Westlaw Precision AI is Thomson Reuters’ other product, embedded in Westlaw itself. If your firm already lives in Westlaw, it’s the shortest path to AI-assisted research without changing anything about how your team works.
All three cost real money: hundreds of dollars per month, per seat, usually on top of the underlying research subscription. For a litigation-heavy firm doing novel research weekly, they pay for themselves. And all three still hallucinate occasionally, because the retrieval layer can grab the wrong passage or the model can misread what it retrieved. Grounded in real databases means fewer errors, not zero. Verify anyway.
Tier 2: litigation analytics
Different job entirely. These tools don’t answer “what’s the law.” They answer “what will this judge do.”
Lex Machina, also owned by LexisNexis, mines federal dockets for patterns: how often a specific judge grants summary judgment, how long her cases take to resolve, what opposing counsel’s settlement history looks like, how a particular defendant behaves in litigation. When you’re deciding whether to file in one district or another, or whether the defense firm across the table actually tries cases, this is data you cannot get by reading opinions.
Trellis does the same for state courts, which matters because most litigation is state litigation and federal analytics won’t help you in front of a county judge. Coverage varies by state, so check yours before subscribing.
Neither replaces Tier 1. A firm with heavy motion practice might want both. A firm that settles everything pre-suit might want neither.
Tier 3: general AI for preliminary research
This is where Claude and ChatGPT live, and where lawyers get in trouble by misunderstanding the job.
The job is orientation. You inherit a matter in an area you haven’t touched since law school. Twenty minutes with Claude gets you the framework: the governing statutes, the major doctrines, the terms of art you’d need to even run a competent Westlaw search. It’s also excellent for structuring an argument, stress-testing your theory, and translating a doctrine into plain English for a client letter.
What it is not: a citation source. Never cite a case from a general AI chat. This is where the sanctioned lawyers went wrong. The models generate case names and reporter cites that look exactly like real ones because that’s what the format looks like. Some are real. Some aren’t. You can’t tell by looking.
The cost is free to $20 a month, which is why this tier is where every firm should start. For questions that need current web sources with citations attached, Perplexity sits in this tier too.
One more tool belongs here: NotebookLM. Upload the documents yourself, the statute, the contract, the deposition transcripts, and it answers questions grounded only in what you gave it, with pinpoint references back to the source. For research inside a defined document set, it’s the safest AI research experience available, because there’s no open-ended database for it to hallucinate from.
The budget path for small firms
Here’s the combination I’d run at a two-lawyer firm before spending a dollar on Tier 1:
General AI for orientation and drafting, at $20 a month. Free case law from CourtListener and Google Scholar, which between them cover most published opinions you’ll need. Your state bar’s research resources, which many lawyers forget they already pay for through dues. And the verification discipline from the top of this page, applied every time.
That stack, used daily, beats a $500-a-month subscription used twice a quarter. I’ve watched lawyers buy the premium tool first, let it sit, and conclude AI research doesn’t work. The tool was fine. The habits weren’t there. Build the habits cheap, then upgrade when you’re actually hitting the ceiling.
How my firms do it
CPLG handles lemon law in all 50 states, and lemon law is statute-heavy. Song-Beverly in California, Magnuson-Moss federally, and each state’s own lemon statute. The law doesn’t shift much month to month, which changes the math on tools.
So instead of paying for Tier 1 seats across 62 staff, we loaded the actual statute text into Claude Projects and built prompt libraries around it. A team member asks how a repair attempt is counted under a specific state’s statute, and the answer comes grounded in the statutory text we provided, not the model’s memory. That covers most day-to-day research questions for a fraction of the cost.
The litigation team still verifies everything against primary sources before filing, and for genuinely novel questions they go to the traditional databases. The AI didn’t replace research. It removed the repetitive 80% so the humans spend their hours on the hard 20%.
The bottom line
- Verify against the primary source at every tier and every price
- Buy Tier 1 when novel research is a weekly reality, not an aspiration
- Analytics answer “what will this judge do,” not “what is the law”
- Start at $20 a month; where that fits in the wider adoption order is in the roadmap
What to do today
Pick one real research question from a current matter. Run it through Claude or ChatGPT for the framework, then verify what it tells you on CourtListener or Google Scholar. Time the whole thing. Compare it to what that question normally costs you.
If the cheap stack keeps up with your caseload, you just saved the subscription money. If you hit its limits inside a week, you’re the firm the Tier 1 tools were built for, and now you’ll actually use what you buy.