RAG-Ready Patterns for Data Platforms

Chapter 6. Scenario-Based Access Control for RAG Systems

Building an enterprise-grade AI system requires more than high-quality data. While accuracy is foundational, the true measure of a production-ready application is its ability to operate safely within the boundaries of the organization that deploys it. In practice, the primary challenge is rarely determining whether data is correct. The more difficult question is whether that data is appropriate to use for a particular user, in a particular environment, and for a particular intent.

Traditional access control models were designed for deterministic systems where a user opens a specific file, report, or database table. Retrieval-augmented generation (RAG) changes that model entirely. Instead of retrieving a single predefined asset, a RAG system composes context dynamically across many data products at query-time. A single response may draw from metrics, entities, documents, and intermediate metadata before producing a final answer.

This shift creates a fundamental tension inside enterprise platforms. AI systems require broad access to organizational data to generate useful responses. At the same time, security and compliance teams are responsible for enforcing strict least-privilege policies that limit how sensitive information is accessed and shared.

Without a mechanism to reconcile these competing requirements, organizations face an uncomfortable trade-off. They can restrict the system so tightly that it becomes ineffective, ...