What careful AI data work looks like in a legal setting
Litigation produces data nobody chose and everybody has to use anyway: account exports, message backups, audio recordings, spreadsheets with formats that never quite line up. Turning that into something an attorney could actually trust looked like a data problem going in. It turned out to be a discipline problem instead, and the discipline is what this case is really about.
Category
Data processing and compliance controls for legal and other regulated data work.
What I was trying to accomplish
Legal work produces awkward data: formats nobody chose, volumes nobody wants to read by hand, and a requirement that whatever comes out of the process holds up under scrutiny later. Over my time as operations lead I supported case teams with exactly that kind of work. The goal was never just to process the data. It was to produce something attorneys could review, trust, and use, with every step defensible if someone later asked how we got there.
The data challenges
The work itself was unglamorous by design. A parser for each format, instead of hoping one tool would handle all of them. Transcription instead of summarizing from memory. Results loaded into an index a person could sort, filter, and check, with every item traceable back to its source. Where a spreadsheet had to be produced, it was generated from the data rather than typed by hand, then re-checked by script for row counts, totals, and formulas before anyone saw it. On one project, that check caught a real formula error before delivery, which is the entire argument for having it.
The result
A repeatable pipeline with its own proof built in, not a one-time cleanup. Matter-specific rules lived in a written file so every session applied them the same way, instead of depending on what one person remembered. The default was flag, not delete, with a backup before touching any deliverable. Any change to a designation happened only on explicit instruction, with a backup and a dated note in the record. Anything requiring legal judgment went back to the attorneys as an open question rather than getting resolved by the tool. AI made the mechanical part faster. The controls are what made the output usable in legal work.
Technical details
Tools: Claude Cowork, Python, speech-to-text transcription, Excel, Word
Litigation teams sometimes receive large, messy data collections and need them turned into something attorneys can review: organized, searchable, screened for privilege, and ready to load into a review platform. On two occasions I supported case teams with this kind of work, with final responsiveness and privilege determinations remaining with the attorneys.
The first involved a large collection spanning several formats, including account exports, message backups, and audio files. Across a series of Cowork sessions, Claude wrote a parser for each data type, transcribed the audio, and built a multi-tab review index with search-term hits, a privilege log, and a coverage check against the full collection. The work ended in review-platform-ready production packages with validated load files. The second was a structured-data reconciliation. Before building the parser, I ran the plan through a multi-agent review in which several independent advisors critiqued it and a synthesis separated technical questions from decisions that belonged to counsel. Claude then built and verified a parser for very wide, irregular spreadsheet exports.
The controls mattered more than the code. Matter-specific rules, such as the privilege screening criteria, lived in a persistent file so every session applied them the same way. The working conventions were flag rather than delete, and back up before editing any deliverable. Scripted checks ran before every handoff, and one of them caught a formula error before delivery. Any change to a designation was made on explicit instruction from the case team, with a backup and an audit trail. Open questions went back to the attorneys rather than being resolved by the tool.
Pull up a chair.
Let's talk through where things stand, and where they could go.