Most AI governance programs end at sign-off. A use case is reviewed, risks are noted, and deployment proceeds. Weeks later, an agent invokes a tool that was never in scope — and no system blocks it.
The approval gap
Approval captures intent at a point in time. Agents operate continuously, models get updated, prompts drift, and new tools appear in agent frameworks. Without runtime enforcement, approved policy and live behavior diverge silently.
What runtime enforcement looks like
Every agent action passes through a policy evaluator. Allow — tool is in the approved manifest and within data scope. Deny — tool undeclared, data out of bounds, or policy version mismatch. Each decision is logged with timestamp, agent identity, policy version, and context for audit.
EU AI Act and logging
High-risk AI systems must enable automatic recording of events (logs) over the lifetime of the system. Runtime logs are not optional telemetry — they are evidence that oversight was operational, not theoretical. Regulators and internal auditors will ask: show me what the agent did, what policy applied, and who approved that policy.
Violations as feedback
Denied actions should alert GRC in real time — not surface in a quarterly review. Policy violations reveal gaps in intake, assessment, or testing. A denied hr_database.query call means either the agent was misconfigured or the approved manifest was incomplete. Both are governance failures worth fixing in process, not just in code.
Closing the loop
openRegal ties authorized deploys to policy tokens (e.g. policy v3). Runtime enforcer evaluates against that manifest. Engineering sees events in the runtime monitor; GRC sees alerts in the governance console. Governance extends from first deploy to every tool call — where risk actually materializes.