The situation
A 220 person engineering consultancy with fifteen years of completed projects. They sat across a document management system, a shared drive that predated it, and a lot of individual mailboxes.
Tender responses were the acute pain. Each one started with a senior engineer walking the floor asking who did something similar. Then they dug through folders named after clients who were since acquired twice.
The estimate we validated in discovery was a day and a half of senior time per response, on around forty responses a year.
The knowledge audit
Before building anything we spent two weeks on a knowledge audit. It produced two uncomfortable findings.
First, about a third of the archive was superseded. Old rate cards, old standards, drafts that had never been marked as drafts. Indexed naively, the system returns obsolete pricing with confidence.
Second, permissions were inconsistent. Several folders that were meant to be restricted were open to everyone, and nobody noticed because nobody was able to find them.
Both of these had to be resolved before retrieval was worth building. This is normal and it is why we no longer quote these programs without the audit in scope.
The build
Retrieval in place across the existing systems, with no migration. Every answer returns with citations that link back to the source document. Every source carries a freshness signal. A 2013 standard shows as a 2013 standard.
Permissions mirror the existing access model exactly. A person asking the brain sees precisely what they can find themselves. The security review took one meeting rather than a quarter.
We built exactly one agent: a tender skeleton assembler that pulls the closest three precedents and produces a starting draft with the sources marked. Everything else stayed as retrieval.
The operating model
Four knowledge domains, four named owners, one monthly meeting where they read the list of questions that got poor answers.
That meeting is the thing that keeps this alive. We saw the same technical build succeed in one company and rot in another. The difference is almost always whether someone reads the bad answers.
Result
Tender preparation dropped by around a day and a half per response. The unexpected benefit was in onboarding, where new starters stopped needing a senior person on standby for their first month.