stack

Legal research isn't one tool anymore — it's a stack. The firms billing the most hours on research are the ones still using single-platform workflows from 2019. The modern approach layers AI-native search on top of traditional databases, then runs everything through a verification step that catches hallucinations before they hit a brief.

Here's what actually changed: Westlaw AI-Assisted Research and Lexis+ AI both launched retrieval-augmented generation in 2024, meaning they pull from their verified databases instead of hallucinating citations. But they're not the only game. vLex's Vincent AI and Claude are reshaping how solos and mid-size firms approach research — often at a fraction of the cost. The question isn't whether to use AI for research. It's which combination gets you to confident, citable answers fastest.

The Modern Legal Research Stack: Query to Citation in 4 Steps

Step 1: Query formulation. Claude or GPT-4 to brainstorm search terms, identify relevant statutes, and frame the legal question. This takes 2 minutes instead of 15. Step 2: Primary source search. Run your refined query through Westlaw AI-Assisted Research ($125-200/user/month) or Lexis+ AI (similar pricing). For international or secondary sources, vLex Vincent AI ($99/month for solos) covers 130+ jurisdictions. Step 3: Verification. Every AI-surfaced case gets checked against the source database. Westlaw and Lexis do this internally — their AI cites to their own verified corpus. For Claude-generated research, you verify manually or use Casetext's CoCounsel. Step 4: Citation formatting. Automated Bluebook formatting through your platform or a tool like Clearbrief. Total workflow: 45 minutes for research that used to take 3-4 hours.

Westlaw AI vs Lexis+ AI vs vLex Vincent: Head-to-Head

Westlaw AI-Assisted Research is the most conservative — Thomson Reuters built it to minimize hallucination risk, grounding every response in WestSearch results. It's the best choice for litigation where citation accuracy is non-negotiable. Weakness: it's expensive and sometimes over-cautious, returning broad results that need human narrowing.

Lexis+ AI is more conversational and better at synthesizing across practice areas. It handles multi-jurisdictional questions well and integrates with Practical Guidance for transactional work. Weakness: the AI layer sometimes surfaces older, less relevant precedent first.

vLex Vincent AI is the dark horse. At roughly half the cost of the Big Two, it covers international law better than either and handles civil law jurisdictions that Westlaw barely touches. For immigration, international trade, or cross-border work, it's arguably superior. Weakness: U.S. case law depth still doesn't match Westlaw's.

Claude (direct use) isn't a legal database — it's a reasoning engine. Use it for issue spotting, argument construction, and statutory interpretation. Never for citation. The winning stack: Claude for thinking + Westlaw or Lexis for sourcing.

Where AI Research Actually Fails (And How to Catch It)

AI legal research fails in three predictable ways. First: fabricated citations. General-purpose LLMs (ChatGPT, Claude without database access) will generate plausible-looking case names that don't exist. The Mata v. Avianca disaster in 2023 proved this publicly. Fix: never cite anything from a general LLM without verifying in a primary source database.

Second: outdated law. AI models have training cutoffs. Claude's knowledge has a cutoff date, and even Westlaw AI can lag on very recent decisions by days. Fix: always check the date range of your AI results and run a recency filter.

Third: jurisdiction confusion. AI models sometimes blend rules from different jurisdictions in a single response, especially on procedural questions where state rules diverge significantly. Fix: always specify jurisdiction in your query and verify each cited rule applies to your court.

Firms that build verification into their workflow — making it a required step, not optional — report zero citation errors from AI-assisted research.

Cost Comparison: What Research Actually Costs Per Matter

Traditional approach: Associate spends 4 hours at $350/hour = $1,400 in billable time per research memo. Client gets billed, associate gets experience, but the economics are brutal for smaller matters.

AI-assisted approach: Same associate spends 45 minutes with AI tools = $262.50 in billable time plus $150-300/month in tool costs (amortized across matters). That's an 80% reduction in per-matter research cost.

For solos: vLex Vincent at $99/month or Casetext at $150/month replaces what used to require a $30,000+/year Westlaw subscription for basic research needs. You won't get the same depth on niche federal regulatory questions, but for 80% of research tasks in general practice, it's sufficient.

The real savings aren't in tool costs — they're in time. Managing partners report associates handle 2-3x more matters when AI research is part of the workflow. That's revenue multiplication, not just cost reduction.

Building Your Firm's Research Protocol

Don't let every attorney freelance their AI research approach. Build a protocol. Step 1: Choose your primary database (Westlaw or Lexis) and your AI supplement (Claude, vLex, or CoCounsel). Step 2: Create a verification checklist — every AI-surfaced case must be confirmed in the primary database before citation. Step 3: Require jurisdiction specification in every query. Step 4: Mandate date-range checks on all results.

Document the protocol. Train on it quarterly. The firms getting burned by AI research aren't the ones using it — they're the ones using it without guardrails. A 30-minute training session prevents a malpractice claim. The ABA's Formal Opinion 512 (2024) makes clear that lawyers must understand the AI tools they use. Having a written protocol demonstrates that competence.

The Bottom Line: Westlaw AI-Assisted Research for litigation-heavy firms that need bulletproof citations. vLex Vincent AI for solos and international practices watching costs. Claude as your thinking layer regardless — it's the best legal reasoning engine available, just don't cite from it directly.