Privilege is the product, and it does not survive an upload.
Discovery, redlining and trial preparation are the three jobs where AI earns its place in a firm — and all three require the model to read the client material a firm can least afford to send anywhere.
Where the hours go
Reading. Terabytes of case files, emails and transcripts, read by the people billing the highest rates, looking for the contradiction that matters. Contracts checked clause by clause against precedent. Cross-examination built the night before. Model Rule 1.6 does not bend for convenience: the disclosure is complete the moment a paralegal pastes a deposition excerpt into a consumer chatbot, and nothing in the firm’s records shows it happened.
Privileged e-discovery and deposition analysis
Search across every matter, pleading, exhibit and transcript the firm already holds, building timelines and surfacing contradictions, with every answer naming the document and page it came from.
Client evidence sent to a third-party model is a disclosure to a third party. On-premise inference keeps the Rule 1.6 analysis short: there is no recipient.
Contract redlining against your own precedent
M&A and vendor agreements compared against the firm’s historical positions, flagging the clauses that fall outside what the firm has accepted before.
A firm’s accumulated contract library is its commercial edge. Feeding it to a public model risks training someone else’s system on bespoke legal strategy the firm spent decades developing.
Trial preparation without a network
Cross-examination questions, evidence synthesis and motion drafting, running on the machine in the room.
Courthouse Wi-Fi is unreliable and often absent, and the hours when strategy changes fastest are the hours an attorney is sitting in one. A local system does not care whether there is a network.