A knowledge agent partners actually open before billing.

01 — Client context

A magic-circle firm with a decade of precedent memos, deal documents, and partner notes scattered across DMS, SharePoint, and individual inboxes. Two prior internal AI pilots had failed — partners didn't trust the answers and stopped opening the tool within six weeks.

02 — Operational problem

Senior fee-earners needed answers grounded in the firm's own work product, with citations they could defend to a client and a confidentiality model that respected matter-level access controls. Generic chat over a vector store was a non-starter.

03 — System designed

A retrieval-grounded knowledge agent with matter-aware permissions, hybrid retrieval over normalised metadata, and a strict citation contract — every claim links to a specific paragraph in a specific document the user is entitled to read.

04 — Implementation shape

• Document model rebuilt around matter, jurisdiction, and document type before any retrieval was tuned
• Eval set co-authored with five partners across three practice areas
• Citation-or-refuse policy: the agent declines to answer if it cannot cite
• Quiet rollout to one practice group, then expanded only after 30 days of clean evals

05 — Guardrails, evals & governance

• Matter-level access enforced at retrieval time, not at the prompt
• Confidentiality redaction on draft outputs prior to any logging
• Quarterly partner-led review of refused queries and disputed citations

06 — Measurable result

• 12,000 partner and associate queries per month within 90 days of full rollout
• Zero reported confidentiality incidents across the first two quarters
• Average research task time down from 47 to 12 minutes on a sampled corpus