Technical leaders · Scientific & technical
Move AI agents to governed data, not data to agents
Running agents inside the governed data boundary can reduce duplication, egress cost, latency and observability gaps.
Moving enterprise data into a separate AI-agent stack often creates another copy of permissions, catalogs, storage and monitoring. As tools multiply, egress cost, multi-hop latency and policy inconsistencies grow with them.
The architectural principle
A data-native approach runs models, retrieval, agent memory and tools within the boundary where data access, quality, lineage and observability already exist. Policy can then be enforced before computation, while the answer’s trace stays connected to authoritative sources.
This is not a universal prescription. Vendor constraints, deployment location and the need for external models may require a hybrid design. Choose by measuring data movement, sensitivity, latency, auditability and total operating cost—not only the speed of a demonstration. The key question is whether the production architecture preserves the organization’s existing control plane as agents take on more steps and retain more state.