How to Run an AI Governance Internal Audit: Step-by-Step

What an AI Governance Internal Audit Verifies

An AI governance internal audit is a structured review that tests whether an organization's governance policies for AI agents are actually enforced in production, not just documented. This distinguishes it from a policy review, which examines what the organization says it does, and from continuous monitoring, which observes runtime activity on an ongoing basis without necessarily testing conformance against a defined standard.

ISO/IEC 42001 requires organizations operating an AI management system to conduct internal audits at planned intervals to verify conformance to both internal requirements and the standard itself. For enterprises deploying AI agents, this means confirming that agent identity, permission grants, tool-call activity, and policy enforcement points can be evidenced, not assumed. The audit's core question is whether governance controls hold up against logs and configurations, or whether a gap exists between stated policy and operational reality.

Mapping Policy Documents to Runtime Enforcement

A recurring failure pattern in AI governance audits is treating a policy document as evidence of control. Auditors need to identify where enforcement is actually implemented, whether at the application layer, a runtime gateway, or the model runtime itself, and confirm that this is distinct from where policy is merely written down.

In architectures using the Model Context Protocol, agents connect to tools and data sources through defined client-server boundaries, and each MCP server introduces its own permission scope that must be reconciled against the organization's documented governance policy. Multi-agent or multi-tool systems require checking permission boundaries at each integration point rather than only at the top-level agent.

Auditors should be able to correlate a single action back to the identity that initiated it, the permission grant that authorized it, and the tool-call record that logged it. Runtime governance platforms, including Trussed AI's runtime policy enforcement and audit logging capabilities, are built to produce this correlated evidence at the point of enforcement, which is what an auditor ultimately needs to test control effectiveness rather than control design.

Common Control Gaps Found During Audits

These gaps appear frequently when policy is compared with production evidence:

• Permission drift: agents accumulate broader access over time than their documented least-privilege policy specifies
• Staging leakage: test environment configurations or credentials remain reachable from production runtime paths
• Incomplete tool-call logs: logs capture that a call occurred but omit parameters or outcome needed for testing
• Non-human identity gaps: agent identities lack the same audit trail rigor applied to human users
• Unverified MCP permission boundaries: server-side tool access scopes are not reconciled against governance policy
• Design-time-only enforcement: a policy exists but no runtime control point actually applies it

Governance Reporting and Audit Independence

ISO/IEC 42001 supports the principle that the internal audit function should remain independent from the teams that build or operate the AI systems under review, which reduces the risk of findings being shaped by the same assumptions that produced the original control design.

Audit frequency and depth should be aligned to the risk tier of the agent or system, following the risk-based orientation of NIST's AI RMF and its Generative AI Profile, rather than applying the same audit cadence across low- and high-risk deployments uniformly.

Where an organization's AI systems fall within the scope of regulation such as the EU AI Act, audit findings should be explicitly mapped to the relevant documentation, logging, or risk-management obligation, since regulatory review will expect evidence of that mapping rather than a general statement of compliance. Audit results and open remediation items should be reported to executive or board-level oversight on a regular basis, consistent with how management-system audits are handled under ISO/IEC 42001, so that governance gaps are visible above the operational team responsible for closing them.