How Internal Audit Should Govern Its Use of AI

An internal auditor asks a generative AI tool to cluster preliminary findings, identify recurring themes, and help draft an audit committee summary. The output is polished and useful. A month later, however, the team cannot reconstruct which model version was used, what information was submitted, which claims were checked against source evidence, or how the auditor changed the analysis before the conclusion was issued.

The failure is not that AI was used. The failure is that Internal Audit introduced a new decision-support layer without defining its own control environment.

AI should therefore be governed as part of the audit function’s operating model, not treated as an informal productivity add-on. The control burden does not need to be heavy for every prompt. It does need to be proportionate to the sensitivity of the data, the degree of reliance on the output, the significance of the audit decision, and the level of automation involved.

The Standards already create a governance baseline

The Global Internal Audit Standards do not prescribe a dedicated AI policy. They do, however, establish mandatory requirements that make ungoverned AI use difficult to defend.

Under Standard 10.3, Technological Resources, the chief audit executive must strive to ensure the function has technology that supports the internal audit process, regularly evaluate the technology in use, provide appropriate training when new technology is implemented, and collaborate with information technology and information security functions on implementation. The implementation considerations further point to technology-related risks such as information security and privacy.

Standard 9.3, Methodologies requires the CAE to establish systematic and disciplined methodologies and update them when significant changes affect the function. If AI changes how auditors plan procedures, analyze populations, write code, create workpapers, or draft reports, those practices belong in Internal Audit methodology rather than in scattered team habits.

Standard 5.2, Protection of Information is equally important. Internal auditors must respect confidentiality, privacy, and ownership of information and manage the risk of inadvertent exposure. An enterprise-approved AI platform can reduce risk, but platform approval does not replace the auditor’s responsibility to decide whether particular engagement information may be processed there.

The operating principle follows naturally: Internal Audit’s AI policy should translate existing professional obligations into rules for AI-enabled work.

A controlled AI capability needs more than an acceptable-use statement

A useful governance model links each risk to a control expectation that can be evidenced and reviewed.

ISO/IEC 42001:2023 provides a useful management-system lens for many of these elements: AI policy, responsibilities, risk assessment and treatment, documented information, operational controls, monitoring, and improvement. It is an organizational AI management system standard for entities that develop, provide, or use AI. It is not an Internal Audit standard and does not mean the audit function must create a separately certifiable management system. Its value here is architectural: it encourages Internal Audit to manage AI as a lifecycle capability rather than as a collection of disconnected tools.

Reliance on AI changes the evidence question

AI can summarize interviews, classify documents, suggest procedures, generate code, detect anomalies, or draft report language. Those are valuable capabilities, but they are not interchangeable with audit evidence.

Standard 14.1, Gathering Information for Analyses and Evaluation, requires information supporting the engagement to be relevant, reliable, and sufficient. Standard 14.6, Engagement Documentation, requires documentation detailed enough for an informed and competent person to repeat the work and derive the same engagement results. When AI materially affects an audit conclusion, that creates a practical traceability obligation.

The answer is not to retain every interaction. A low-risk prompt used to improve grammar in non-sensitive text does not deserve the same record as an AI-assisted analysis that identifies the transactions behind a significant finding. A better test is: if the AI output disappeared tomorrow, could the team still demonstrate what source information was used, how the output was validated, and who made the professional judgment?

COSO’s 2026 publication Achieving Effective Internal Control Over Generative AI is particularly helpful on this point. It treats GenAI outputs as claims requiring validation, recommends controlled configurations and access, emphasizes human review gates, and highlights the need for evidence when management relies on AI-supported outputs. The guidance also recognizes that vendor updates, prompts, retrieval sources, and model behavior can change quickly, making monitoring and change control central rather than optional.

A risk-tiered model is therefore more durable than a tool-specific rulebook. Low-risk uses such as ideation on non-sensitive content can operate with light controls. Medium-risk uses that transform internal information or support audit analysis should require approved tools, source-based validation, and proportionate traceability. High-impact uses that shape evidence, engagement conclusions, critical communications, or autonomous actions should trigger pre-approval, stronger access and change controls, explicit human sign-off, and enhanced monitoring.

The vendor sits inside the audit control chain

Most audit functions will not build their own foundation models. AI arrives through cloud services, embedded features in existing audit platforms, coding assistants, retrieval tools, or agents connected to enterprise systems. That introduces a form of dependency that traditional software reviews can underestimate.

Vendor governance should establish at least four facts: what information the provider receives and how it may use it; what can change without the organization’s approval; which logs, configurations, and versions can be reconstructed; and how the organization exits the service without losing records needed to support audit work. Where AI is connected to document repositories or a retrieval-augmented generation environment, inherited permissions matter as much as the model itself. An AI assistant that can retrieve documents outside the auditor’s need-to-know boundary is a data-governance problem even if the chat interface appears secure.

The IIA’s Artificial Intelligence Auditing Framework is professional guidance rather than a mandatory Standard. It complements the GIAS by organizing AI considerations across governance, management, and internal audit and by emphasizing accountability, transparency, traceability, and control. Those same principles should be applied inward when Internal Audit adopts AI for its own work.

AI governance should become part of the QAIP

The easiest governance model to implement is also the easiest to forget: approve a tool, publish a policy, deliver training, and assume the risk is now stable. Generative AI does not behave like static office software. Models change, vendors modify services, retrieval sources age, prompts evolve, and auditors discover new use cases.

The control environment therefore needs a quality loop. Standard 12.1, Internal Quality Assessment, requires ongoing monitoring and periodic self-assessment of the internal audit function. Standard 12.3, Oversee and Improve Engagement Performance, reinforces engagement supervision and improvement. AI-related controls can be built into those existing mechanisms instead of creating a separate compliance bureaucracy.

A CAE can make meaningful progress in 90 days by inventorying current tools and use cases, defining risk tiers and data rules, embedding AI requirements into methodology and workpaper templates, refreshing vendor due diligence, and adding AI-assisted engagements to quality sampling. Metrics can remain practical: policy exceptions, outputs corrected during review, rework attributable to AI, unapproved tool use, failed validations, or material vendor changes.

The goal is not to log every prompt or prevent auditors from experimenting responsibly. It is to preserve what makes Internal Audit credible: protected information, reliable evidence, professional skepticism, reproducible documentation, competent supervision, and continuous improvement.

The most mature audit function will not be the one with the most autonomous AI. It will be the one that can explain where AI is used, why the use is appropriate, which controls apply, and how the function knows that professional judgment and audit quality remain under human accountability.

Sources

• The Institute of Internal Auditors, Global Internal Audit Standards
• The Institute of Internal Auditors, Artificial Intelligence Auditing Framework
• ISO, ISO/IEC 42001:2023 — Artificial intelligence management systems
• COSO, Achieving Effective Internal Control Over Generative AI