Case studies

A single rental notice can be governed by four layers of law at once, state, city, housing authority, and federal voucher rules, and they don't always point the same direction. Getting from that tangle to 239 verified notice-language requirements meant building a reference that would rather refuse an answer than guess at one.

Nobody left at the firm understood the 15-year-old macro system that assembled its core documents, and the person who built it was long gone. Reverse-engineering roughly 20,000 lines of undocumented code fast enough to matter meant living by one rule: label every claim as inferred or verified, and log every reversal instead of quietly fixing it. That rule caught a wrong assumption a person, not a document, ended up correcting.

A law firm's AI adoption had been happening ad hoc, and attorneys carry real ethical duties around competence, confidentiality, and supervision that ad hoc doesn't satisfy. Replacing it meant scoring fifteen use cases against risk and the state bar's own ethics opinion, and being willing to hold some of them back, including one that looked ready until a transcription test came back unreliable.

Did we budget for this? was a question that kept getting answered from memory, against a general ledger export of roughly 25,000 transactions and several competing versions of the same detail sheet. The fix wasn't just a better workbook. It was a queryable record that could name the exact cell and the exact email thread behind an answer, and say plainly when it couldn't.

One staff member had built a 295-file court-rules reference library by hand over years, and nobody could say how much of it was still accurate. Checking is exactly what a large, informally maintained library resists, so the audit split 272 URLs across ten parallel agents just to get through it, and the real finding wasn't a broken link. It was that nothing could be verified automatically until each jurisdiction had a documented source of truth.