What Boards Should Ask Internal Audit About AI

A board can receive a detailed presentation on artificial intelligence and still miss the answer it actually needs: Where is the organization relying on AI, and how much independent confidence exists that the resulting risks are being governed effectively?

Internal Audit can help answer that question, but it should not become the owner of AI governance. Its value is to provide an independent view of whether the organization understands its exposure, whether controls are working, whether assurance coverage is sufficient, and whether the audit function itself can credibly assess a fast-changing risk landscape.

In June 2026, The IIA added support tools specifically aimed at this conversation, including the Board AI Governance Questionnaire, the Audit Committee: Internal Audit AI Coverage Questionnaire, and the Management AI Risk and Control Questionnaire. These are diagnostic tools and examples of good practice, not new mandatory requirements. The professional foundation remains the Global Internal Audit Standards and each function's board-approved mandate.

Ask Internal Audit for assurance, not ownership

The first issue is governance. Management remains responsible for deciding how AI will be used, establishing policies, operating controls, maintaining inventories, managing vendors, and responding to incidents. Internal Audit, as the third line, provides independent assurance and advice without taking on management responsibility.

That boundary changes the quality of board questions. Asking, “Do we have a good AI policy?” can produce a description of a document. Asking, “What independent evidence do we have that the policy is operating and that exceptions are identified?” moves the discussion toward actual effectiveness.

The Global Internal Audit Standards reinforce this expectation. The chief audit executive's interaction with the board should support a shared understanding of organizational risks and assurance priorities. Internal audit strategy and planning should respond to the organization's objectives and risks, while resource constraints that could impair delivery should be communicated to the board. For directors, the practical question is whether those professional responsibilities are producing sufficient assurance over AI risk.

Nine questions that change the conversation

Board question	What it reveals	What a strong answer should contain
1. What material AI uses exist, and how confident are you that we know about all of them?	Real visibility over exposure.	A management-owned inventory with business owners, purpose, data, third parties, risk classification, and methods for detecting undeclared or embedded AI.
2. Which AI risks could materially affect our objectives?	Whether AI is treated as enterprise risk rather than only technology risk.	Prioritization linked to strategy, operations, fraud, compliance, privacy, cybersecurity, third parties, people decisions, reputation, and resilience.
3. Is AI governance operating in practice, or does it mainly exist on paper?	Control design and operating effectiveness.	Evidence on accountability, risk appetite, approvals, human oversight, monitoring, incidents, vendors, and escalation of exceptions.
4. What independent assurance have we performed over AI, and what remains uncovered?	The difference between audit activity and risk coverage.	A clear view of completed work, AI embedded in business-process audits, thematic reviews, unassessed risks, and planned coverage.
5. Is the internal audit plan changing as quickly as our AI exposure?	Adaptability.	Evidence that new deployments, incidents, regulation, vendors, or technologies can change audit priorities, scope, or timing.
6. Do we have the skills and resources to challenge AI credibly?	Third-line capability.	A skills-gap assessment, training, access to data, cyber, or model-risk specialists, and transparent decisions about hiring or co-sourcing.
7. Can we trust the AI information being reported to the board?	Reliability of management reporting.	Testing over data integrity, performance and risk metrics, incidents, exceptions, model degradation, and escalation criteria.
8. What are our most important readiness gaps for AI assurance, and what is needed to close them?	Maturity and concrete priorities.	A short list of material gaps, accountable owners, actions, timelines, required resources, and measures of progress.
9. How is Internal Audit governing its own use of AI?	Whether the function applies the discipline it expects elsewhere.	Approved tools, data rules, human review, output verification, documentation, access controls, vendor oversight, incident handling, and quality controls.

None of these questions requires directors to become model validators or AI engineers. They require something more useful: the ability to distinguish between known exposure, existing controls, and independent confidence obtained.

AI coverage is not the same as one “AI audit”

A common mistake is to ask whether Internal Audit has scheduled a dedicated artificial intelligence audit and then use the answer as a proxy for assurance coverage.

AI exposure may require several forms of work. A cross-functional review may assess governance, inventory, and policy. A human resources audit may cover AI-assisted recruitment. A procurement engagement may examine a generative AI contract assistant. A third-party audit may address dependency on external models or providers. A cybersecurity engagement may include AI-specific threats.

The better board conversation therefore starts with an AI exposure and assurance map, not a count of “AI audits.” That map should show where AI could affect material objectives, what first- and second-line mechanisms exist, what Internal Audit has tested, and where a confidence gap remains.

Conceptual assurance map showing AI use cases, audit coverage, and remaining gaps

Answers that should concern the audit committee

Some responses do not automatically prove a control failure, but they justify deeper challenge:

“IT knows which AI tools we use,” without a verifiable inventory and business ownership.
“We have an AI policy,” without evidence of approvals, exceptions, monitoring, or control testing.
“AI is covered by cybersecurity,” even though the risks also affect decisions, fraud, compliance, vendors, or customers.
“We do not need specialists,” despite planned assurance over complex models or high-impact systems.
“The annual audit plan has already been approved,” as a reason not to respond to material change during the year.
“Internal Audit already uses GenAI,” without clear rules for confidential information, output validation, documentation, and professional accountability.
“The models have good accuracy,” without tolerances, risk indicators, incidents, bias, degradation, or an explanation of what “good” means for the specific use.

The common pattern is confusing the existence of an activity with evidence that the risk is controlled.

Turn the questions into an assurance agenda

A useful practice is to ask the chief audit executive for a concise, recurring view of five things: material AI exposure, governance, assurance coverage, Internal Audit capability, and readiness gaps. It does not need to be a complex dashboard. It should allow the board to see what changed, what has been assessed, what remains outside coverage, and which limitations may reduce the confidence Internal Audit can provide.

The IIA-Australia's 2026 Internal Audit Excellence Framework offers a complementary maturity lens for this discussion by linking internal audit excellence to strategic alignment, capability, technology, agility, innovation, and assurance over emerging technologies. It is not a requirement of the Global Internal Audit Standards, but it is useful for challenging whether the function is prepared for risks that can evolve faster than traditional audit cycles.

A mature discussion does not end by asking whether Internal Audit “knows AI.” It ends by asking whether the function can demonstrate, with the right competence and evidence, where the organization needs more independent confidence and how that confidence will be obtained.

The final question

If the board could ask only one question, it should be this:

Where are we relying on AI without yet having sufficient independent assurance over the risks that matter most?

A useful answer forces strategy, inventory, controls, coverage, capability, and resources into the same conversation. More importantly, it makes visible the difference between using AI and being ready to govern it.