Shipping search and RAG that operators trust in production is hard. This talk covers product and architecture lessons from building enterprise GenAI retrieval systems: tradeoffs, evaluation that catches silent failures, access control and tenancy, and governance so answers stay auditable. Examples draw from taking a conversational GenAI platform from pre-revenue to enterprise customers, and from working in regulated trading environments where wrong answers have real cost. Attendees leave with a short production checklist and failure modes to design against before the next pilot.