Retrieval-augmented generation breaks at scale because organizations treat it like an LLM feature rather than a platform discipline. Enterprises that succeed with RAG rely on a layered architecture.