Cutting Contract Review Time by 74%

The firm runs 200 or more commercial contracts a month through junior associates for the first review pass. Each contract takes three to four hours of manual work: flagging non-standard clauses, summarising key terms, and routing risk to a partner. Independent benchmarks put even a routine manual contract review near 92 minutes, and put routine review at 60 to 80% of a legal team's time, so at this firm's volume the first pass alone consumed a large, fixed slice of the associate base. Partners were paying associate hours (mid-market associate billing runs 300 to 625 dollars an hour, and far higher at the top of the market) for work whose consistency they did not trust. The commercial stake was direct: slower turnaround on client contracts, uneven flagging between associates, and senior time spent re-checking junior output.

The firm had already tried two off-the-shelf legal-AI tools and rejected both. One was too US-centric for a global practice. The other produced clause citations that did not exist or did not say what the tool claimed. That failure is the industry norm, not bad luck: the Stanford RegLab and HAI study of leading legal-AI research tools found Lexis+ AI and Ask Practical Law AI produced incorrect information more than 17% of the time, and Westlaw AI-Assisted Research more than 34%, where a hallucination includes a real citation that does not support the stated claim. A tool that misgrounds one flag in three is worse than no tool, because a partner cannot tell which third is wrong without redoing the review. That is why the prior attempts failed and why grounding, not speed, was the problem to solve.

We spent the first two days reading 50 of the firm's contracts to learn their clause taxonomy and risk criteria from their own paper, not a generic template. The system splits each contract, embeds and indexes it against a curated library of standard commercial clauses, and classifies and flags against that library. Every flag carries the retrieved clause it matched, so an associate sees the basis, not just a verdict. High-risk clauses are routed to a human reviewer rather than auto-accepted. The interface lives inside the firm's existing document-management system: associates see flagged sections with citations, partners get a daily digest of the high-risk contracts.

Each contract is segmented and embedded, then retrieved against a pgvector index of the curated clause library, so classification is grounded in a real reference clause rather than free generation. GPT-5.5 runs the classification and flagging pass; Claude Opus 4.8/Sonnet 4.6 and Gemini 3.5 Flash cross-check on the harder clause types. The grounding contract is strict: a flag only stands if it traces to a retrieved clause, which is what holds the fabricated-citation rate at zero against the 17 to 34% industry bar. The human boundary is explicit and non-negotiable: high-risk clauses are always human-reviewed and never auto-accepted, so the system carries volume and a lawyer carries judgement. Built on Next.js, PostgreSQL, pgvector, and AWS Lambda, integrated into the existing document-management system.

SECOND-ORDER EFFECTS