Twenty years of files, turned into answers a partner can stand behind.

The client is a regional professional-services firm, advisory and audit, whose edge is institutional knowledge: how a similar problem was solved three years ago, which precedent applies, what a given regulator expects. That knowledge lived in the memories of a few senior people and across tens of thousands of documents spread over matters, drives, and inboxes.

So every inbound request started with a hunt. A junior would ask around, dig through old engagements, and assemble a thin first draft; a senior would find it wanting and redo the search themselves. The people whose judgement the firm sells were spending their days retrieving things the firm already knew.

As the firm grew, the hunt got slower, not faster. More history meant more to search, and more ways to miss the one precedent that mattered.

The presenting complaint was that first drafts took too long. The value driver underneath it was senior utilisation: the firm made money when its experts advised, and lost it when they searched. The binding constraint was trust in retrieval. A junior could find something, but a senior could not rely on it without checking the source, so the search got done twice.

That is why an off-the-shelf assistant would not have moved the number. A tool that produces fluent answers with no traceable source adds a third search, because now someone has to verify the machine. The leverage was in retrieval a senior could trust on sight, every claim tied to the document it came from.

So we built for grounding first and fluency second, and drew a hard line at advice. The platform finds, cites, and drafts. It does not counsel the client.

The platform ingests the firm's engagement history and resolves it: which documents belong to which matter, which client, which entity, so a search is scoped correctly and one client's material never leaks into another's answer. On top of that sits retrieval built for grounding: every passage an answer rests on is pulled, ranked, and attached, so a claim can be traced to the exact document it came from.

An agent layer then drafts the first-pass response, a memo, a scope, a client note, grounded in the retrieved material and constrained so it cannot assert what it cannot cite. A maker-checker step tests each draft against its own sources before a person sees it, and anything thin or unsupported is flagged rather than smoothed over.

The result is not a chatbot on the firm's files. It is a governed retrieval-and-drafting system with the firm's own knowledge as its ground truth and a citation behind every sentence.

Every client-facing claim carries a citation to the source it came from, so a senior reviews a grounded draft rather than a fluent guess. What the platform cannot support with a source, it flags rather than asserts.

Client material is walled: retrieval is scoped by matter and client, so one engagement's documents can never surface in another's answer. A person approves anything that goes to a client, and every retrieval and draft is logged.

A first client-ready draft now takes about forty minutes, down from most of a working day, because the search behind it is trusted the first time. Across the senior team, the best part of a day a week was freed from retrieval and redrafting and returned to advisory work the firm actually bills for.

Nine in ten inbound requests are now resolved without escalating to a partner, because the first draft is good enough to build on. Capacity rose without a single new hire, and the firm's knowledge compounds, as every answered request is itself grounded, reviewed, and available to the next.

The time we gave back was senior time, the most valuable and most constrained thing the firm has, about a fifth of each senior's week, and it did not disappear into slack. Redeployed to billable advisory it is worth an estimated AED 48,000 a month per senior, and it went straight into the two things that grow a professional-services firm: fee-earning work, and the new business the team was previously too busy to chase.

Faster, better-grounded answers also win more of the work the firm pitches for, because in professional services the firm that responds first and best usually wins: on this team the pitch win rate moved from roughly 25% to 33%. And because every answered request is captured and grounded, the firm's institutional knowledge compounds, so each engagement makes the next one faster. Taken together, the platform lifted revenue from the same team by an estimated 20%, with no new hires.

That is the leverage: more revenue per senior, more clients served without more people, and a knowledge base that widens the gap on competitors with every matter.

The platform finds, grounds, and drafts. It does not decide what a client should do. Judgement, the read on a regulator, the call on a risk, the relationship itself, was kept firmly with the firm's people. We built the thing that hands a senior a trustworthy starting point in minutes, precisely so the senior spends the saved hours on the part only they can do.