Lathe.
← IndexEntry 002
AI agent / RAG buildLegal services · 2025

A precedent-search agent that gave a 40-lawyer firm its evenings back

of eval questions answered with correct citations

92%

median associate research time saved per week

11 hrs

total inference cost at firm-wide usage

£380/mo

The problem

§ 01

Archway, a 40-lawyer commercial firm, had three decades of precedent documents spread across a document management system that search had effectively given up on. Associates spent hours reconstructing work the firm had already done — and partners knew it, because they were billing the rework.

The firm had been pitched 'AI transformation' twice by larger consultancies. Both proposals started with a six-figure platform and ended without a measurable definition of success.

The build

§ 02

We started with the measurement, not the model: fifty real questions from partners, with the documents a good answer should cite. That evaluation set was the contract for the whole engagement.

The pipeline ingests the DMS nightly, chunks documents with their legal structure intact, and serves a retrieval agent that always cites its sources — answers without grounding are refused, not improvised. Claude does the reasoning; the firm's documents do the knowing.

We shipped to five partners first, watched the traces, fixed the retrieval misses, and only then rolled out firm-wide with usage analytics and a running-cost dashboard the managing partner can read.

Claude APIRAG pipelinepgvectorNext.jsEvaluation harnessDMS integration

Every other vendor led with the demo. Lathe led with how we'd know it was working. That's why this one actually stuck.

Managing Partner, Archway LLP

Want a result like this one?

Four short questions, then a call with the people who'd actually build it.

We reply within one UK business hour.