AI Legal Research Tools Compared: Lexis, Westlaw, Claude

By Jude Lee · · Comparison

Two lawyers comparing case law on a laptop next to printed reporters in a small firm library

The phrase covers three different technical things, and most vendor marketing blurs them deliberately.

Retrieval-grounded research tools run your question against a licensed, curated corpus of cases, statutes, and secondary sources, then have a language model summarize what it retrieved. As the vendors describe them in their own product documentation — see the current LexisNexis Lexis+ AI and Thomson Reuters CoCounsel product pages, which change often enough that you should read them at the time you buy — the model is constrained to the database the vendor licenses.

General AI assistants — Claude, ChatGPT, Gemini — are extremely strong at reading, synthesizing, and drafting from material you give them. Ask them to recall case law from memory and you are asking a language model to reproduce citations from training data, which is exactly how fabricated authority enters a brief. Give them the PDF of the actual opinion and ask for the holding, and they are a genuinely useful research assistant.

Agents go a step further: multi-step tools that search, pull documents, cross-check against your firm’s prior briefs, and draft a memo, taking actions along the way rather than answering one question at a time. That is where MCP (the Model Context Protocol, an open standard for giving an AI governed access to your systems) becomes relevant — it is how you let an assistant reach your document management system without copy-paste.

Where Lexis+ AI and Westlaw’s CoCounsel actually diverge

At the product level, three things decide this for a small firm, and none of them is “which model is smarter.”

Corpus scope. Each vendor’s AI answers draw on that vendor’s licensed library. If the treatise, practice guide, or state-specific form set your practice depends on lives only in one library, that settles the question before you evaluate any model behavior.

Citator depth. Shepard’s sits behind the Lexis products and KeyCite behind Westlaw. The meaningful difference is how deeply the citator is wired into the AI output — whether flags appear inline on every cited authority in a generated answer, or whether you have to click through and check each one yourself. Ask for a live demo on that exact point.

Whether drafting shares the seat. CoCounsel is positioned as an assistant with a bundle of tasks — research, document review, deposition preparation — alongside Westlaw research; Lexis has been packaging drafting features alongside Lexis+ AI as well. Tiers, bundling, and naming shift frequently, so as of 2026 the only reliable answer is the written quote you are given, not a comparison chart.

How the three approaches actually differ

Closed-corpus research platform

Strengths: answers are drawn from a licensed corpus; integrated citators tell you whether a case is still good law — general models have no equivalent; per-seat pricing is predictable; the vendor’s audit trail is defensible if a court asks how the research was done.

Weaknesses: per-seat cost is the largest line in many small-firm research budgets, and you are limited to the vendor’s corpus and workflow. Grounding also is not verification: Stanford RegLab and HAI researchers have benchmarked leading AI legal research tools and documented outputs containing unsupported or mischaracterized citations even on retrieval-grounded systems. Read that work rather than a marketing claim of accuracy, and test each tool on questions you already know the answer to.

General assistant + your own documents

Strengths: superb at synthesis, drafting, and reading long documents you supply — opinions, the other side’s brief, a 300-page administrative record. Far cheaper per user. Connect it to your document system via MCP and it can also draw on your own prior work product.

Weaknesses: no citator, no guarantee an authority exists, and no vendor-side audit trail. Confidentiality depends entirely on the plan and terms you signed. Unsupervised, it will produce a fluent memo built on authority that has been reversed — or invented.

The third option — a custom research agent — sits on top of both. It is a packaged workflow: search a licensed source through its API or a permitted interface, pull the retrieved opinions, compare them against your firm’s own briefs in your DMS, and draft a memo in your house format with a verification checklist attached. We’ve written about the underlying pattern in building a custom MCP server over your firm’s matter data, and the same architecture applies to a brief bank.

Be honest about when this is worth it. A custom agent pays off when you research the same shape of question repeatedly — the same motion in the same jurisdiction, dozens of times a year. It does not pay off for varied one-off research, where a seat license plus a careful lawyer beats anything you’d build.

The right question is not which tool is smartest. It is which tool makes verification cheap — because verification is the part you can never delegate.

The disadvantages people actually run into

The common objection — that research associates are the first legal role to be automated — points at the wrong risk. The realistic problems in a small firm are mundane:

On replacement: the durable value in research is judgment about which authority actually helps this client before this judge. Drafting and synthesis are compressing. Judgment isn’t.

Modeling the cost without inventing numbers

Don’t take anyone’s published savings figure, including ours. The three items below are input formulas you fill in, not measured findings:

hours × billing rate
Value of recovered research time — use your own realized rate, not standard rate
Input formula, not a benchmark
seats × annual license
Platform cost — get a written quote; list price is rarely the real price
Input formula, not a benchmark
cites × minutes to verify
Verification cost, which never goes to zero and belongs in every comparison
Input formula, not a benchmark

Run it for one real matter type. Say an associate spends N hours on a recurring research task; assume an AI-assisted draft cuts drafting time but adds a fixed verification block. Net recovered hours = (old hours) − (new drafting hours + verification hours). Multiply by realized rate, then subtract license cost. If the result is close to zero, the honest answer is that a seat license and a good research habit are enough. Our automation ROI walkthrough shows how to build that model properly, including the reallocation question — recovered hours only turn into revenue if there is billable work waiting to absorb them.

A four-week evaluation you can actually run

  1. Write five gold-standard questions

    Pick five research questions from closed matters where you already know the correct answer and the controlling authority. This is your test set. Never evaluate a tool on a question you can’t grade.

  2. Run all three approaches on the same set

    Trial the closed-corpus platform, run a general assistant with the actual source documents supplied, and have a lawyer do it manually. Score each on: correct authority found, hallucinated or mischaracterized authority produced, time to a verified answer. Set the kill criterion in advance — any fabricated citation on the gold-standard set from a paid tier ends that tool’s trial, full stop, regardless of how good the prose is.

  3. Cost the verification step honestly

    Time how long it takes to confirm every citation in each output. Include that in the comparison. A tool that produces a beautiful memo requiring 40 minutes of checking may lose to a slower tool with linked, citator-flagged sources.

  4. Write the standing instruction before you roll out

    Define, in writing, which tool is used for what, what may never be pasted into a general assistant, and who signs off. Package that instruction as a reusable skill so the process runs identically regardless of who is at the keyboard.

Which one fits your firm

Solo or two-lawyer firm, varied practice: a general assistant for reading and drafting, plus whatever research subscription you already have for authority and citator checks. Don’t build anything. This is the right answer more often than vendors admit.

Five to twenty lawyers, litigation-heavy, one or two jurisdictions: a closed-corpus seat for the people who research daily, a general assistant for everyone for document synthesis, and clear written rules dividing the two. Compare specific offerings against a custom option in our Harvey vs. CoCounsel vs. custom agent breakdown.

Firms with a genuine repeating research pattern and a real brief bank: a custom agent connecting your assistant to your DMS via MCP is worth scoping — after you’ve proven the workflow manually for a quarter.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides