Turning 3,000 client files into a firm-wide knowledge system
An Australian professional-migration law firm had two decades of expertise locked in folders and inboxes. Now it answers questions from its own history in seconds.
The business
A professional-migration law firm in Australia, with more than a decade of casework across employer-sponsored visas, skilled migration, and review-tribunal matters.
The problem
The firm's expertise was real but unreachable. Every visa application it had ever filed — the submissions that worked, the evidence sets that satisfied the Department, the arguments that survived a refusal — sat inside client folders and email threads, organised by client rather than by knowledge.
The practical cost showed up daily:
- Precedent hunting. A lawyer preparing a nomination would spend an hour digging through folders hoping a colleague remembered a similar case.
- Company memory in people's heads. The senior practitioners were the knowledge base. When they were unavailable, the firm slowed down.
- Patterns nobody could see. Refusals were handled one at a time, so the common threads across them — the ones that would change how the firm prepares a file — stayed invisible.
- Slow onboarding. A new agent facing an unfamiliar visa subclass had no path from "what is this?" to "here is how we do it."
What I built
A two-layer knowledge system over the firm's own history.
Layer one — the knowledge layer. Around forty atomic, densely linked pages covering visa subclasses, processes, legal concepts, evidence types, forms, and legislative authority. De-identified, so it is safe to share internally and functions as firm IP rather than confidential material.
Layer two — the matter layer. Every client file processed into a structured note: a case summary, a correspondence timeline, and a document register. Each matter links into the knowledge layer, so the concept pages accumulate real cases behind them.
The ingestion pipeline processed the firm's live file store — 3,091 client matters, 114,107 emails, and 139,388 documents, PDFs and images, including OCR over scanned material so that even photographed documents became searchable text.
On top of that sits a retrieval layer: deterministic full-text search for instant lookups, and an AI assistant connected through MCP that answers questions grounded in the firm's own matters, with citations back to the source file. Ask it how the firm has evidenced labour-market testing for a hard-to-fill role, and it answers from what this firm actually did — not from generic advice.
The result
- Precedent finding drops from an hour of folder-digging to a query. Open a concept, see every matter that touched it, filter to the closest match.
- Patterns become visible. Refusals can be queried across occupation, sponsor profile, and evidence method — surfacing compliance insights no individual file review would reveal.
- Deadline exposure is monitored, not remembered. Visa expiries are a field in the data, so at-risk matters surface before they lapse.
- New staff are productive in days, not weeks. Concept → worked example → the real matter behind it.
- The knowledge compounds. Every new matter links into the same concept hubs, so the system gets more valuable with each file rather than merely bigger.
Why it worked
Two design decisions did most of the work.
The two-layer split kept confidential client data separate from shareable firm knowledge — a hard requirement in a regulated practice, and the thing that makes the system safe to use daily.
And the system was built to sit on top of how the firm already works, reading the existing file store rather than asking anyone to adopt a new platform or re-file a decade of work. The best knowledge system is the one nobody has to remember to update.
Client details anonymised at the firm’s request.
Recognise this problem in your business?
Book a conversation