How Internal Auditors Can Audit AI Governance

An organization can have an AI policy, an AI steering committee, and quarterly board reporting and still be unable to answer three basic questions: How many AI systems are in use? Who is accountable for their decisions? What happens when a vendor changes a model?

That gap between declared governance and operating governance is where internal audit can provide meaningful assurance.

Auditing AI governance does not require every auditor to become a data scientist. Internal auditors can assess whether the organization has clear decision rights, accountable owners, risk criteria, lifecycle controls, monitoring, and evidence. Technical model validation may require specialists. Determining whether the organization knows what AI it has, who approves it, how risks are assessed, and how failures are detected is governance, risk, and control work.

The first question is not “Can I explain the algorithm?” It is “Can I trace accountability and evidence from the decision to use AI through deployment and ongoing monitoring?”

AI governance is a decision system, not a policy document

The IIA’s Artificial Intelligence Auditing Framework provides practical guidance for assessing AI strategy, governance maturity, risks, and controls. It is professional guidance, not a replacement for the Global Internal Audit Standards and not a new set of mandatory requirements.

The Global Internal Audit Standards provide the professional audit discipline underneath the engagement. Standard 9.1 requires the chief audit executive to understand the organization’s governance, risk management, and control processes. At the engagement level, Standard 13.2 requires internal auditors to understand the activity under review and assess relevant risks, while Standard 13.4 requires appropriate evaluation criteria. Findings and conclusions then need to be supported by relevant, reliable, and sufficient information.

This matters because a weak AI audit often starts with a generic list of “AI controls.” A stronger engagement starts one level earlier: Which AI-related decisions matter to organizational objectives, and what criteria define acceptable governance in this organization?

Give each framework a specific job

No single framework should automatically become a universal audit checklist. The most useful sources answer different questions.

Source	What it gives the auditor	Best use
IIA AI Auditing Framework	An internal-audit lens on AI strategy, governance, risks, and controls.	Shape objectives, questions, and coverage.
ISO/IEC 42001:2023	Management-system requirements covering context, leadership, policy, responsibilities, risk, operation, performance evaluation, and improvement.	Use as criteria when adopted, or as a structured benchmark where appropriate.
NIST AI RMF 1.0	Voluntary outcomes for governing, mapping, measuring, and managing AI risk across the lifecycle.	Assess whether risk management is continuous, contextual, and proportionate.
COSO	The familiar control-environment, risk-assessment, control-activities, information-and-communication, and monitoring lens.	Embed AI into existing internal control rather than building a parallel control universe.

ISO/IEC 42001 needs one important distinction. It is an AI management-system standard. Clause 9.2 requires the organization to conduct internal audits of the AI management system at planned intervals and to preserve objectivity and impartiality in the audit process. That requirement does not automatically mean the organization’s Internal Audit function, operating under the Global Internal Audit Standards, must perform those management-system audits. The organization can assign that program differently; Internal Audit should separately consider mandate, independence, competence, and whether reliance on that work is appropriate.

NIST describes the AI RMF as voluntary. Its Govern function cuts across Map, Measure, and Manage and emphasizes accountability, culture, documentation, and continuous risk management. NIST notes that AI RMF 1.0 is being revised, so audit teams should version-control the criteria they use.

COSO offers a different advantage: it lets auditors discuss AI in the language of business controls. Its February 2026 publication, Achieving Effective Internal Control Over Generative AI, translates the five components of the Internal Control—Integrated Framework into GenAI-oriented control practices. It is useful guidance for generative AI, not a universal requirement for every AI system.

Six questions turn governance into auditable evidence

You do not need to test every model to determine whether the governance system works. Select representative use cases and trace them from intake through monitoring.

Governance question	Evidence you should expect	Possible audit test
1. What AI exists, and what matters?	Inventory, risk classification, purpose, owner, vendor, data, users, and lifecycle status.	Reconcile the inventory to procurement, SaaS, contracts, expenses, cloud, SSO, or API records to identify undeclared AI.
2. Who can decide, and who remains accountable?	Roles, authorities, committees, RACI, business owners, system owners, data owners, and monitoring owners.	Sample decisions and verify that approvals and escalations follow assigned authority.
3. What use is acceptable?	AI policy, risk appetite or tolerances, prohibited uses, exception process, and data-handling rules.	Test actual exceptions for timely approval at the required level and against defined conditions.
4. How does AI enter, change, and leave production?	Risk and impact assessments, approval criteria, testing, version/change records, and retirement controls.	Trace recent systems from request to production and test whether material changes trigger reassessment.
5. Is human oversight meaningful?	Reviewer responsibilities, competence, intervention authority, thresholds, logs, and escalation paths.	Observe the process, inspect overrides, and determine whether reviewers receive enough information to challenge AI outputs.
6. How does the organization know the system still works?	Metrics, thresholds, alerts, incidents, complaints, vendor changes, periodic reviews, and corrective actions.	Sample alerts or incidents and trace detection, escalation, investigation, remediation, and closure.

The value comes from traceability. A policy shows documented design. A use case that can be followed from inventory and approval to monitoring provides far stronger evidence that governance actually operates.

Test operation, not just documentation

A superficial governance audit asks whether an AI policy exists. A useful audit asks whether the policy changes behavior and decisions.

Suppose the policy prohibits employees from entering confidential information into public GenAI services. Internal audit can look beyond the document: Are approved enterprise tools available? Are technical restrictions configured? Do employees receive role-relevant training? Can the organization detect prohibited use? Is there an exception process, and are exceptions monitored?

Likewise, if risk assessment is required before AI deployment, sample recent implementations and determine when the assessment occurred, who approved it, which controls were identified, and whether those controls were in place before production.

A practical starting sample might include three to five use cases with different profiles: a low-criticality internal tool, a broadly used GenAI application, a third-party AI service, and—where present—a system influencing significant customer, employee, financial, or compliance decisions. For each one, follow the same chain:

confirm inventory entry and risk classification;
identify the accountable owner and decision authorities;
inspect risk/impact assessment and approval criteria;
test data, access, usage, oversight, and third-party controls;
review monitoring, changes, incidents, and corrective actions.

This is how a generalist auditor moves from governance statements to evidence of operating effectiveness without turning the engagement into model validation.

Knowing the specialist boundary is part of audit quality

Generalist internal auditors can assess strategy, ownership, inventory, policies, risk appetite, approvals, segregation of responsibilities, human oversight, third-party governance, monitoring, incidents, and documentation. They can also determine whether management has defined when independent technical validation is required.

The boundary appears when the audit conclusion depends on expertise the team does not possess. Statistical performance validation, advanced fairness testing, model robustness, adversarial testing, red teaming, deep review of data pipelines, or specialized cybersecurity testing may require data science, model risk, privacy, or security specialists.

Bringing in a specialist does not transfer Internal Audit’s accountability. The audit team still needs to frame the assurance question, define scope, assess the specialist’s competence and objectivity, integrate the specialist evidence with the rest of the engagement, and communicate a conclusion that answers the audit objective.

The most common failures are audit-design failures

Four mistakes quickly weaken an AI governance review:

Auditing one tool and calling it governance. A chatbot review says little about whether the organization governs its overall AI portfolio.
Turning one framework into a checklist without context. Criteria should reflect applicable obligations, organizational objectives, policies, risk appetite, and maturity.
Treating “human review” as a control by name alone. Human oversight only works when the reviewer has competence, information, time, intervention authority, and evidence of actual challenge.
Concluding from documents rather than operation. Policies, committees, and RACI charts support design; samples, decisions, logs, alerts, incidents, and remediation show whether governance functions in practice.

AI governance may sound new, but its audit logic is familiar: know what is being done, who is accountable, which risks are accepted, which controls are expected, and how failures will be detected.

Internal auditors do not need to master every algorithm to assess those questions. They need clear criteria, traceable evidence, professional skepticism, and the discipline to recognize when deeper technical testing requires specialist support.

The most useful first step is not to ask which model the company uses. Pick one real AI use case and determine whether you can trace it, without gaps, from the decision to use AI to evidence that it remains under control.