Every platform now promises the same thing: point an agent at your files and it will answer questions. It will, too. Confidently, pleasantly, and just often enough correctly to be dangerous.

This session is about the distance between that demo and a system a thousand accountants use to look up answers they act on. The distance is not one missing feature. It is a different discipline at every layer: content that has to be prepared before it ever reaches an index, retrieval that plans queries instead of matching keywords, answers that stay inside what was actually found, citations that link to sources because trust is earned per sentence, and evaluation that turns "it feels better" into a number you can defend.

I walk through that whole stack as built on Azure AI Search and a custom agent, in production, in German, for users who did not lower their standards just because AI was involved. Along the way: the shortcuts that looked fine and failed quietly, the real cost per question, and the eval harness that settled every argument.

For engineers and architects who have a working RAG demo and a nagging feeling about it.

Point an agent at your files and it answers. Confidently. Sometimes even correctly. This talk covers everything between that demo and production RAG: content pipelines, agentic retrieval, grounding, citations, evals, and real costs, built on Azure AI Search for a thousand users.