Context
Knowledge existed across documents and chat, but finding the right answer required too much tribal memory.
Challenge
The prototype needed to respect data boundaries and show whether retrieval quality was good enough before a production build.
Approach
Selected a narrow set of documents and real user questions.
Built a retrieval-based prototype with source-aware answers.
Evaluated answers against expected responses and failure cases.
Outcome
Validated useful workflows and exposed content gaps.
Produced a realistic next-step plan for security, ingestion, and evaluation.
Avoided treating AI as magic search.