The problem
Associates were spending large blocks of billable capacity on research and first-draft assembly that drew heavily on work the firm had already done: precedents, briefs, and memos scattered across matter files. Public AI tools were a non-starter: privileged material can't leave the firm's control.
What we built
A research and drafting workflow grounded in the firm's own work product, deployed inside an architecture the firm controls, with citations back to source documents so attorneys verify rather than trust. Governance came first: which matters were in scope, who could query what, and what review applied before anything reached a client file.
First drafts, adversarially reviewed
The system also produced two litigation first drafts, each run through a multi-agent review before an attorney touched them: several frontier models independently attacked the draft from different reviewer personas, and the findings were synthesized — with agreement statistics across the models — into a prioritized revision list. The attorneys started from a draft that had already survived its harshest critics.
The result
Research time dropped by roughly 60%, and first-draft assembly moved from associate hours to a reviewed pipeline. As important: zero privileged material in public tools, and an audit trail for every answer. The firm's name stays out of this write-up for the same reason the architecture exists: their confidentiality obligations are the point.