Law firm cuts matter intake from days to hours

New-matter intake averaged ~36 hours of partner and admin time before a matter could be opened: conflict check, engagement letter draft, billing matter setup, ethical-wall configuration. Partners were declining marginal matters because the friction was too high.

What we built

• Intake agent ingests the partner’s initial email or call note, extracts party names and matter description, and runs a structured conflict query against the firm’s conflict database.
• Conflict-report agent drafts the conflict memo with citations to every prior matter or party hit — the agent briefs the conflicts partner, never decides.
• Engagement-letter draft agent grounded in the firm’s 12 approved templates; flags non-standard fee arrangements for partner review.
• Clio + iManage matter provisioning via MCP tools, gated on partner sign-off.
• Strict eval gating: conflict-report recall is the headline metric, held to ≥99% against a 200-conflict labelled set. Below 98% on rolling 30-day sampling auto-disables the agent and pages us.

Caveats and context

We over-engineered the entity-graph layer in v1 and rewrote in v1.1 using a simpler structured search inside the existing conflicts database. The senior conflicts partner was the review-queue bottleneck; we now produce a Friday-morning summary digest so she batches the work.

The 2025–2026 legal-AI hallucination wave (Sullivan & Cromwell, Gordon Rees twice sanctioned, 1,300+ documented incidents) drove our design: a human partner reviews every conflict output. “AI verifying AI” is not our verification standard.

Stack. AWS Bedrock (single-model Claude Sonnet for auditability) · Bedrock Knowledge Base · iManage + Clio REST · custom Bedrock guardrail · Langfuse self-hosted.

Median intake time fell from 32 hours to 4.5. Conflict-report recall held at 99.4% in production. Matters opened per quarter up 18%.