2,400 engagements indexed. Analyst onboarding 8 weeks → 2 weeks. £1.2M/yr in analyst time recovered.

Olympia Advisory is a 40-person strategy consulting firm with offices in London and New York. The firm had run 2,400 client engagements over 14 years: market entries, competitive strategy, M&A advisory, growth strategy, operating model design, regulatory positioning.

Every engagement produced a board deck, a financial model, a research dossier, and partner-written analysis. All of it sat in SharePoint organised by client folder and year. Analysts trying to find prior work on, say, payments market entry in Southeast Asia could spend a full day searching — and might find 3 of the 7 relevant precedents because they couldn't be located by sector or methodology, only by client.

The compounding effect of 14 years of work was lost. Partners answered the same precedent questions over and over because there was no other path. New analysts spent the first 8 weeks of their tenure essentially learning the firm's history by osmosis — sitting next to seniors, asking what work had been done before, building a personal mental index.

The firm had tried generic enterprise search (Microsoft, Glean) on the SharePoint corpus. Hit rates were mediocre because the questions analysts ask aren't keyword-shaped — they're about analogous situations, comparable frameworks, what a partner had concluded last time. Generic search couldn't surface those connections.

Every document in the 2,400-engagement corpus was processed through structured extraction: PDF, DOCX, PPTX, XLSX. Each engagement was tagged by sector, geography, deal type, methodology, framework, and partner sign-off. Custom embeddings were trained on the firm's own taxonomy rather than off-the-shelf semantic similarity.

Analysts query via a web UI or a Slack interface. A question like 'what did we conclude when a US life insurer asked us about real-time underwriting in 2022?' returns a citation-grounded answer in 90 seconds — with the original deck pages, the relevant model tabs, and the partner's concluding recommendation surfaced as primary evidence.

Citations are mandatory. Every answer the system produces is grounded in a specific document, page, or model cell. No hallucinated synthesis. If the firm hasn't done analogous work, the system says so — it doesn't paper over the gap with a plausible-sounding generic answer.

Partners review the access controls deal-by-deal. Some engagements are NDA-sensitive and don't enter the corpus. Others are usable for retrieval but not for direct quoting. A small number are public reference material and citable verbatim. The classification is partner-driven, audit-logged, and revocable.

Onboarding analysts now spend their first two weeks using the system to learn the firm's body of work the same way a senior would — by exploring precedents in the questions that come up on their first engagement. The 8-week shadowing period collapsed.

Query-to-citation hit rate measured at 87% over the first 4,000 analyst queries: the system surfaced relevant precedents that the analyst confirmed were useful. The remaining 13% split between genuine gaps (firm hadn't done analogous work) and questions where the analyst preferred to ask a partner directly. Both are legitimate — and the gap data is itself valuable, surfacing where the firm might commission new IP work.

Analyst onboarding compressed from 8 weeks to 2 weeks. The 6 weeks recovered per new hire — across 18 hires in the year — adds up to £180k in faster productivity. Bigger than that: senior partners stopped being the firm's enterprise search.

Annualised analyst time recovered: £1.2M. The savings show up as more engagements delivered per analyst, less senior partner time spent answering precedent questions, and faster proposal turnaround (proposals now lean heavily on the precedent retrieval system to assemble relevant prior credentials).

The system has also flagged 4 instances where Olympia's prior work on a client situation directly disagreed with a junior analyst's draft recommendation — surfaced before the partner review, not after. The institutional memory is finally producing the warning signals that justify its existence.

Week 1–3: Discovery, taxonomy audit, partner classification protocols. Week 4–6: Corpus ingestion (2,400 engagements, ~150,000 source documents), embeddings, retrieval system. Week 7–8: Claude synthesis layer, citation pipeline, web UI. Week 9: Slack integration, MCP server. Week 10: Partner access control, audit logging. Week 11: Pilot rollout to 4 analysts. Week 12: Firm-wide rollout, onboarding-flow integration.