Presented by EDB
As enterprises dive deeper into granting AI agents more freedom and autonomy, a critical question arises during every architectural review: What prevents an agent from taking unauthorized actions?
These agents operate on your models, interact with your data within your infrastructure, and ultimately, the accountability for their actions falls on you. This responsibility cannot be fulfilled retroactively or through abstract policies that exist only on paper. Agents require rules that are context-specific, as they lack the ability to make judgments on their actions independently.
Consider a simple rule: Never open the car door. If taken literally, this rule would prevent an agent from entering or exiting the car altogether. However, when the context changes (e.g., the car has been in an accident, there’s a fire, someone is injured and needs to get out), the intended rule would be the opposite. Context is crucial in the moment. We task agents with performing intelligent actions, which necessitates intelligent rules.
The common approach is to implement guardrails around the agent: instructions, policies, and monitoring systems layered on top of the model. While these mechanisms are essential, they have a fundamental limitation: The effectiveness of a rule like “never open the car door” diminishes when the moment arrives to decide whether to open the door. Controls at the agent level are only as reliable as the predictability of the agent’s output, and autonomy is precisely what makes predicting that output challenging. Governance that relies on reviewing actions before they occur cannot keep up with systems that operate in milliseconds across numerous systems simultaneously.
Governance must be executable and enforced where agents conduct their operations: at the operational data layer, within the context, precisely at the moment of action.
The data layer serves as the enforcement point
Agents derive value from interacting with data. They query, retrieve, transform, and increasingly, act on data. A policy stating that an agent should not access a specific data class is meaningful only if the system can deny that access when the agent requests it. Similarly, the auditability requirement for AI is valuable only if the organization can reconstruct the agent’s actions, data interactions, user associations, and outcomes. When governance is integrated at the data layer, it remains effective regardless of how the agent was constructed or behaves, as the control is inherent to the database itself, not a promise made by the agent.
Agent behavior may be probabilistic, but governance cannot be
Enterprises should not rely on models choosing to adhere to policies. Policies must be enforced by the system, ensuring boundaries that agents cannot cross from the outset.
The necessary controls are commonly implemented at the data layer: role- and attribute-based access, row- and column-level security, data classification and masking, policy as code, and comprehensive audit trails.
Agents do not alter the mechanism itself but introduce a new entity that the mechanism must recognize. Identity management must treat the agent as a principal with its own identity and declared purpose at the session’s initiation.
This approach translates into nine controls, grouped into three imperatives:
Enforce it
-
Role- and attribute-based access control enforced at query time for agents and users
-
Dynamic column masking driven by the same policy pathway
-
Agent identity as a primary principal, with a declared purpose bound at session start and user actions preserved
See it and prove it
-
Data classification and tagging that shapes policy
-
Session-level audit logging capturing the agent’s actions, user interactions, and declared purpose
-
Lineage tracing across pipelines to trace results back to their origins
Unify and strengthen
-
Centralized, portable policy management
-
Encryption at rest and in transit
-
Consistent enforcement across on-premises, cloud, and air-gapped environments
“Declared purpose is the key differentiator. It becomes an attribute understood by the access layer, evaluated within the same policy framework as roles and security levels. The enforcement mechanism remains unchanged, but the agent’s purpose becomes part of the evaluation process and post-action record,” explains Priyanka Jain, VP of Product Management, Data & AI Governance at EDB.
Regardless of where you stand in your AI adoption journey, enforcing governance at the data layer enables faster progress. The necessary controls already exist in the database; the shift lies in agents needing to adhere to them.
A digital leash, not a locked door
The objective is not to hinder agents from performing valuable tasks but to define their boundaries, data access, change capabilities, escalation requirements, and event reconstruction processes if issues arise. Governed in this manner, agents are identified, scoped, monitored, and audited. This approach allows enterprises to embrace agents more readily, as security, risk, and leadership teams have confidence in the underlying operational model.
Open, sovereign, and enforceable at the source
Built on the open-source Postgres platform, this foundation empowers enterprises to control where their data resides, who can access it, and under what policies, without relinquishing governance to an external entity they cannot oversee. For regulated industries, the combination of data sovereignty and source-level enforcement is not optional but a prerequisite for deploying agents in production environments.
As agentic systems become more advanced and autonomous, it’s crucial to establish control at the data layer rather than slowing down progress. Enterprises that enforce governance at the data layer can confidently advance in AI initiatives, knowing that their data protection measures are not just theoretical but practical and effective.
EDB Postgres AI is an open, enterprise-grade sovereign data and AI platform that consolidates transactional, analytical, and AI workloads with governance enforced at the data layer. For further details, refer to EDB’s white paper Governing Agentic AI at Enterprise Speed.
Max Romanenko serves as the Chief Technology Officer at EDB.
Sponsored content is created by a company that either pays for the post or has a business relationship with VentureBeat, and it is always clearly indicated. For more information, contact sales@venturebeat.com.
