Enterprise AI platforms serving multiple clients face a hard constraint most AI security talks skip past: how do you let AI retrieve and reason over sensitive client data without any risk of cross-tenant leakage, especially when that data includes unfiled or confidential information? This talk covers the design of a confidential retrieval-augmented generation (RAG) system built for multi-tenant IP SaaS platforms, using trusted execution environments (TEEs) and policy-bound data access to enforce hard isolation boundaries at the infrastructure level, not just the application layer. Drawing on published research and production engineering experience, I'll cover practical tradeoffs between security guarantees and retrieval quality, how to design policy enforcement that survives model updates, and lessons learned building AI infrastructure for clients who cannot tolerate data exposure under any circumstances.