The first production AI risk is usually the data path, not the model. Teams need to understand which source systems can be indexed, which records require permission filters, and how stale context will be detected.
A strong ingestion architecture separates parsing, chunking, metadata enrichment, embedding, indexing, and retrieval. That separation makes it easier to test quality and trace which source records influenced an answer.
Security should be part of retrieval design. Tenant, workspace, role, and record-level access rules must travel with context into the model payload.