title: "Knowledge assistant for a top-10 AU commercial law firm" dek: "RAG over twelve years of precedent. Citations on every answer. Deployed to 250 lawyers. Saved 62% of junior research time." sector: "Professional Services" client: "Top-10 AU commercial law firm" engagement: "Audit → Pilot → Practice Retainer" duration: "14 weeks" year: "2026" outcome: "62% reduction in research task time · 99.4% citation accuracy · 250 lawyers active" solution: "Custom RAG over 2.4M precedent paragraphs with jurisdiction-aware retrieval and paragraph-level citations." timeSaved: "29 minutes per research task · 62% faster" visual: "retrieval" cardFigure: "retrieval" timeMetric: "29 min" timeMetricLabel: "saved / task" costMetric: "A$0.04" costMetricLabel: "cost per query" speedMetric: "2.6×" speedMetricLabel: "faster research" publishedAt: "2026-04-08" keywords:
- legal AI Australia
- RAG law firm
- precedent retrieval AI
- law firm knowledge management
The problem
A top-10 Australian commercial law firm — 250 lawyers, 14 practice groups, twelve years of stored precedent — was watching a familiar pattern: junior associates spending 40% of their time on precedent search. The library was technically searchable. The searches technically worked. But the cognitive load of triangulating "is this the right precedent, for the right jurisdiction, at the right time, with the right adjacent reasoning" was eating billable hours and morale in equal measure.
The Managing Partner had seen two AI demos earlier in the year. Both ended with a polished interface that hallucinated case names. The firm’s internal risk committee — partner-led, intentionally adversarial — had vetoed both. Their position, written in the meeting minutes we later saw: "We will not deploy a system whose errors a regulator would describe as fabrication."
So the brief, when it came to us, was tighter than usual: build something the firm’s partners would actually trust. Not just the lawyers using it — the partners signing off on its existence.
What we did
The Audit took two weeks. The Pilot took ten. The Practice Retainer is now in month four. The architecture decisions we made in the first two weeks are the ones that did the heavy lifting.
Decision 1: Retrieval before generation. The model never invents a precedent. Every answer is a synthesis of retrieved passages with verbatim citations to source documents. If the retriever returns nothing relevant, the system says "no relevant precedent found" rather than generating one.
Decision 2: Citations as first-class output. Every answer surfaces 3–7 source documents with paragraph-level citations. Clicking a citation opens the underlying document at the cited paragraph. A junior lawyer cannot cite something the assistant told them without verifying the source — by design, the friction is low enough that they will.
Decision 3: Partner-managed risk surface. Partners can flag a citation as wrong. The flag is logged, reviewed weekly by the assistant’s named internal owner (a senior knowledge manager we trained), and feeds into retrieval-quality monitoring. The system is governed in the same way the firm governs its junior associates: by senior people who can spot a bad answer.
Decision 4: Jurisdiction-aware retrieval. Australian federal, NSW Supreme, Vic Supreme, and state district precedent are tagged separately. A search for tax-related precedent in NSW does not surface Queensland district court decisions unless explicitly broadened. This matched the firm’s actual work — most matters are state-bound — and dramatically cut noise.
The infrastructure: Azure OpenAI in the firm’s tenancy, AU-resident embeddings, Azure AI Search as the retrieval layer, custom React frontend embedded in the firm’s existing knowledge portal. No data left the firm’s Azure environment at any point.
The outcome — at four months in production
| Before | At 4 months | |
|---|---|---|
| Average research task time | 47 minutes | 18 minutes (~62% reduction) |
| Citation accuracy (spot-audited monthly) | n/a | 99.4% |
| Daily active lawyers | n/a | 198 of 250 (79%) |
| Partner-flagged errors per month | n/a | 3 (down from 11 in month 1) |
| Documents indexed | n/a | 2.4M paragraph-level chunks across 280K source documents |
The reclaimed time hasn’t gone back to the same lawyers. The firm has redeployed the capacity into matters the team was previously turning away — a more profitable use of the same headcount than billable-hour replacement would have been.
What we needed was an assistant our partners would actually trust. EasiraAI took governance seriously from week one. Citations, audit trails, the things a regulator would want — and the things a partner would want before signing off.
— Managing Partner, Top-10 AU commercial law firm
What we’d do differently
Index narrower, faster, first. We indexed all 280K documents in week three. In hindsight, indexing the top 20K by access-frequency in week one would have let us pressure-test retrieval quality earlier. The 260K we added later mostly served the long tail.
Roll out by practice group, not by seniority. We rolled out to "all juniors first" in month one. Practice groups whose juniors had different research patterns (litigation vs corporate) needed different retrieval tuning — which we discovered late.
What we didn’t do
We didn’t train a custom legal model. We didn’t fine-tune. We didn’t replace any of the firm’s case-management or document-management systems. We didn’t deploy any tool that takes legal-research actions without a human reviewing.
The most important architectural choice was the most boring one: aggressive retrieval, conservative generation. It’s also the choice that made the deployment defensible to the partners who had to sign off.
That’s usually the bar that matters.