Document intelligence pipeline for a professional services firm

The situation

Senior staff were spending a meaningful part of each week reading client documentation and transcribing fields into an internal system. It was expensive work done by people whose time was worth considerably more, and it was exactly the shape of problem where a language model earns its cost — unstructured input, reviewable output, a human currently doing it slowly.

What we built

An extraction pipeline that ingests documents, retrieves the relevant passages, and produces a structured draft record with every field traceable back to the text it came from. A reviewer confirms or corrects before anything is committed — that step was designed first and was never up for removal.

• Retrieval over the firm's own corpus, so answers are grounded in their material
• Every extracted field linked to its source passage for review
• An evaluation harness, so a prompt change can be measured rather than guessed at
• Human approval retained by design on every output

Why the review step stays

In professional services the cost of a confident error is asymmetric: one wrong field can outweigh a hundred correct ones saved. Keeping a person in the loop is not a lack of ambition, it is the thing that makes the system usable in a regulated context. The reasoning is set out in where LLMs earn their keep.

This kind of work sits under our AI and blockchain engineering practice, and usually starts with a short feasibility review rather than a build.