Implement AI Agent Governance at the Data Layer

The movement toward giving AI agents autonomy to plan, judge, and execute without human approval is spreading across enterprises. In response to this situation, EDB, an enterprise database company, advocates for implementing AI governance—the mechanisms of control—at the data layer. The question of what actually stops an agent when it attempts unauthorized actions is emerging as a central concern in system design.

Traditionally, AI control has been implemented by layering instructions, policies, and monitoring mechanisms on top of the agent layer. However, this approach has structural limitations. Agent outputs are probabilistic—fundamentally unpredictable. Autonomy itself is the property that makes outputs difficult to foresee, and governance that audits actions before they occur cannot keep pace with agents operating at millisecond speeds across multiple systems. EDB illustrates this problem using the example of car doors—rules that apply in opposite ways depending on context. A rule established in advance may seem valid until the moment judgment is required, but context changes can make the opposite action the correct one.

Against this backdrop, EDB advocates implementing governance 'in executable form, at the data layer where agents actually perform work, and in real-time.' Agents create value through querying, retrieving, transforming, and manipulating data. A policy prohibiting access to specific data is meaningful only when the system can deny the request at the exact moment an agent asks for access. Auditability works similarly—there must be mechanisms to reconstruct what the agent did, which data it accessed, which user it served, and what occurred.

The greatest advantage of placing governance at the data layer is that control functions consistently regardless of how agents are built or how they behave. This is because control exists as an intrinsic property of the database itself, not as an agent 'promise.' EDB explains that the access control mechanisms—such as role-based and attribute-based access controls—that enterprises already operate at the data layer provide the foundation for realizing this. Rather than entrusting agents to follow policies, system-side design must enforce policies.

The significance of this discussion is not trivial for enterprises advancing AI adoption. As AI agent autonomy increases, the conventional approach of giving prior instructions and monitoring whether agents behave as expected becomes less effective. Particularly in sectors such as finance, healthcare, and law, where strict regulations govern data handling, the risk of depending on governance external to the agent is relatively high.

The concept EDB presents—'governance implementation at the data layer'—offers guidance in ongoing discussions about AI agent design. As enterprises seek to expand the scope of autonomy granted to agents, the design decision of 'at which layer to ensure control' will increasingly become an important evaluation criterion from security and compliance perspectives, alongside the question of what to delegate to agents.