AI Control Coverage

The Three Tiers of Coverage

When evaluating an AI agent's capabilities within your software stack, it is crucial to determine whether the platform's API supports visibility, access control, or action control.

Visibility

Visibility is the baseline. It means the system provides logs or audit trails of what the AI agent has read or accessed.

• What it provides: An understanding of what data is being ingested by the model.
• Limitation: It is passive. You can see what happened after the fact, but you cannot stop it in real time.

Access Control

Access control allows you to restrict the agent's connection to specific data sources or revoke its permission to use certain APIs entirely.

• What it provides: The ability to turn off connections, limit the agent to specific folders, or prevent it from reading certain classes of data.
• Limitation: It is often binary. You either give the agent access to a system or you don't.

Action Control

Action control is the most advanced tier. It allows a governance layer to intercept a request from the AI agent before the action is executed and apply rules or require human approval.

• What it provides: Granular, policy-based enforcement (e.g., "Allow emails to internal domains, block emails to external domains without approval").
• Limitation: It requires the underlying platform to support programmatic pauses or webhooks that can intercept and hold a transaction.

Reference Architecture

Mapping out these capabilities aligns with broader risk management practices. Organizations looking for authoritative models on structuring secure and manageable AI architectures can refer to the NIST AI Risk Management Framework, which emphasizes the necessity of mapping, measuring, and managing system interactions.