The problem
Fifteen years of institutional knowledge lived in SharePoint folders, a half-maintained wiki, and the heads of a few senior engineers. New hires took months to become self-sufficient, and the seniors spent a measurable part of every week answering the same questions: which template to use, what the travel policy says, how a past project solved a problem.
The approach
I built Atlas, an internal agent created in Microsoft Foundry, and kept its job description narrow: answer questions from company knowledge, cite the source, and say "I don't know" otherwise.
- Agent in Foundry — Atlas is defined in Microsoft Foundry's agent service with a deliberately strict system prompt: answers must come from retrieved company content and every answer must carry citations.
- Retrieval that respects permissions — documents are indexed into Azure AI Search (hybrid vector + keyword) with their Entra ID group ACLs. The ASP.NET Core backend trims every query by the signed-in user's groups, so Atlas can't leak an HR document to someone who couldn't open it in SharePoint.
- Tools, not guesses — beyond document retrieval, Atlas has one custom tool: a read-only API for the project registry, so questions like "who ran the Novi Sad plant project?" get a factual lookup instead of a hallucination.
- Evaluated before rollout — we collected 120 real questions from staff, wrote expected answers, and used them as an evaluation set for every prompt and retrieval change. Nothing shipped on vibes.
The front end is a plain React chat surface plus a Teams shortcut — nothing novel, which is exactly the point.
The result
After the first quarter, the seniors' "how do I…?" interruptions had dropped by roughly half, and new-hire onboarding surveys stopped listing "finding information" as the top frustration. Because every answer carries citations, staff verify instead of trusting blindly — and the answer logs now double as a heatmap of which documentation is missing or stale.