The challenge
The firm had a problem that nobody talks about openly: their junior associates were spending most of their time reading. Not analysing — reading, looking for the same recurring clauses in the same recurring contract types, tagging them, and writing summaries that the partner would then re-read at twice the billable rate.
The work was necessary. The bottleneck was that it consumed the associates the firm had hired specifically to grow into senior practitioners. Every hour spent finding indemnity clauses was an hour not spent learning to negotiate them.
Off-the-shelf legal AI tools existed. The firm had piloted two. Both failed for the same reason: the firm’s playbook — the partners’ particular take on what counts as risky, acceptable, or non-negotiable — wasn’t in the tool. The tool flagged clauses the partners didn’t care about and missed clauses they did.
Our approach
We didn’t build “an AI lawyer”. We built a structured assistant that applies the firm’s playbook to incoming contracts.
Phase 1 — Codify the playbook (4 weeks). We sat with two senior partners and a managing associate and extracted, in structured form, what the firm actually checks: 64 clause categories, with severity tiers, the firm’s preferred language, and the typical fallback positions. This became the source of truth — versioned, owned by the partners, editable.
Phase 2 — Pipeline (5 weeks). Document ingestion (PDF, DOCX, scanned), structural parsing, clause classification using a fine-tuned model with retrieval against the playbook, and a UI that shows each flagged clause next to the relevant playbook entry, with proposed redlines.
Phase 3 — Lawyer-in-the-loop training (3 weeks). Every flag includes a thumbs-up / thumbs-down. Disagreements feed back into the classifier and into the playbook itself when a partner decides the firm’s position should evolve. Crucially, the system never suggests sending a redline to the counterparty — every output is for an internal reviewer.
Phase 4 — Audit and rollout (2 weeks). A retrospective on 200 historical contracts: how would the system have performed? Where did it disagree with the partners, and was the partner or the system right? Three of those disagreements led to playbook updates.
The outcome
First-pass contract review time dropped 70%. A 60-page commercial agreement that used to take an associate four hours to mark up now takes about an hour, and the partner’s review starts from a far better base.
Two effects we did not anticipate:
- Onboarding accelerated. New associates now learn the firm’s positions by reviewing the system’s flags and the playbook entries side by side. The senior partners called this “the most useful training material we’ve ever produced — and we didn’t write it for training.”
- The playbook itself improved. Codifying it forced disagreements between partners to surface and resolve. The firm now has a single, versioned source of truth on positions that previously lived in three partners’ heads.
Capacity unlocked: roughly 320 junior-associate hours per month redirected from reading to negotiation, drafting, and client work. At the firm’s effective rate, that’s around €120k of annualised capacity — without hiring.