On-Prem Legal Investigation Agent

Challenge

The firm’s investigators read large packs of source material about a subject, such as news articles, court filings and corporate records, and write memos in a strict house style: every fact dated, every claim attributed to its source, nothing inferred. The first draft is slow, repetitive work that grows with the size of the pack. It is also high-stakes, because an invented fact or a misattributed source in a due-diligence memo is a professional liability, and confidential case material cannot go to a public AI service.

Our approach

For this client we built a self-hosted assistant that runs entirely in their own cloud. An investigator uploads the source pack to a case workspace and names the person or company under investigation. The system groups related material by topic, summarises each group about that subject only, and puts it in date order. Every paragraph opens with the month and year and names its source. The prompts forbid outside knowledge, even about well-known subjects, and fixed checks after the model strip placeholders, half-finished sentences and closing commentary. Investigators can make a summary longer or shorter, and ask follow-up questions that are answered only from the documents, or with “not mentioned”. We tested fine-tuning the model on anonymised memos and dropped it: it added cost and risk without better drafts than careful prompts and checks. Each summary is a draft for the investigator to verify, not a finished memo.