Internal answers in seconds, with citations — not guesses
Professional services firm, ~180 staff (anonymised) · Enterprise · Agent co-build · 8-week engagement
Client details anonymised for confidentiality. Figures are rounded and reflect the baseline agreed at the start of the engagement.
The situation
In a 180-person professional services firm, institutional knowledge lived in too many places: a policy folder nobody trusted to be current, an intranet last reorganised two office moves ago, long email threads and a handful of people who simply knew things. New joiners took months to find their feet; HR and operations answered the same questions weekly.
Staff had started asking public AI tools instead — fast, confident and frequently wrong about the firm’s actual policies. Leadership wanted the speed without the fiction.
What we did
We co-built an internal knowledge agent with the firm’s operations team. Discovery defined the question landscape — HR policy, expenses, engagement processes, IT, onboarding — and, critically, the approved source set: around forty documents nominated by their owners as the single current version of the truth.
The agent answers only from those sources and cites the document and section for every answer. If the sources do not cover a question, it says so and routes to the right owner instead of improvising. Permissions mirror the firm’s existing access model, so staff only receive answers from documents they are entitled to read. Two operations staff co-built alongside CedarPro and now own the agent: updating sources, reviewing the unanswered-questions log and refining behaviour.
The human role
Document owners remain accountable for content; the agent never edits or invents policy. Sensitive matters — grievances, personal HR cases, client-confidential questions — are deliberately out of scope and routed to people. The unanswered-questions log is reviewed weekly by a human, which is also how the source library improves.
Results
Around 70% of routine policy and process questions are now self-served against the baseline question audit, saving roughly six hours a week across operations and HR. New-joiner time-to-productivity has visibly improved — onboarding questions dropped from the top of the log to barely featuring. Because every answer carries a citation, trust in the agent has held: staff check sources, find them current and come back.
“The citation is the feature. People stopped asking whether the answer was right and started reading the source it pointed to.”
— Operations Director, professional services firm (anonymised)
What the client keeps
The agent and its governed source library, the permissions model, the unanswered-questions review routine, two trained internal owners, and a usage guide that makes the agent’s boundaries clear to every member of staff.
Facing something similar?
Tell us about your workflow and we’ll walk you through how this engagement would map to yours.