Legal discovery is the clearest case I have worked on for where AI creates real value and where it absolutely must not be trusted. The retrieval problem is enormous and mechanical. The judgement problem is small and consequential.
Designing for defensibility
Any system used in litigation has to be explainable after the fact. A relevance score is not an answer to “why did you not produce this document.” The design therefore prioritised traceability — every surfaced result carrying its own provenance — over raw model performance.
Semantic search where it belongs
Semantic retrieval is genuinely transformative across four terabytes of unstructured material, because the vocabulary of a corpus rarely matches the vocabulary of the question. That is exactly the gap keyword search cannot close, and exactly where embedding-based retrieval earns its cost.
Where the line sits
The system narrows four terabytes to a reviewable set. It does not decide what matters. That boundary was set deliberately at the start of the engagement, and holding it is what made the tool adoptable by people whose professional licence is on the line.