The IIA AI Auditing Framework Explained

An organization can have an AI policy, an AI steering committee, and a catalog of technical controls and still struggle with a basic question: who directs AI, who manages it, and who provides independent assurance over it?

That is the organizing problem behind The IIA’s Artificial Intelligence Auditing Framework, 2nd Edition, updated in September 2024. Its value is not that it offers one more universal control checklist. It separates three dimensions that are often blurred together — Governance, Management, and Internal Audit — and shows how they interact so auditors can understand the environment, ask better questions, and turn expectations into assurance work.

The practical thesis is straightforward: the framework is most useful as an architecture of responsibilities and audit criteria, not as a checklist for “certifying” AI.

What the framework is — and what it is not

The IIA describes its Artificial Intelligence Auditing Framework as guidance that helps internal auditors understand AI-related risks, identify best practices, and recognize relevant internal controls. The second edition expanded coverage of generative AI, governance, ethics, transparency, and accountability, while retaining a structure linked to The IIA’s Three Lines Model and International Professional Practices Framework.

That positioning matters. The framework does not have the same authority as the Global Internal Audit Standards. The Global Internal Audit Standards are a mandatory component of the IPPF and establish professional requirements for the internal audit function. The AI framework is topic-specific guidance. It helps practitioners apply audit thinking to a complex risk area, but it does not by itself create regulatory obligations for the organization or replace applicable law, contracts, internal policies, or standards such as ISO/IEC 42001 when those are relevant audit criteria.

This distinction prevents a common mistake: treating every consideration in the document as a universal requirement. In a real engagement, evaluation criteria still need to come from the obligations and expectations that actually apply to the organization.

Three domains, three different questions

Part 3 of the framework is organized into three domains. Each answers a different question.

Domain	Primary question	What the auditor looks for
Governance	Is the organization directing and overseeing AI consistently with its objectives and values?	Strategy, objectives, risk appetite, board oversight, ethics, responsibilities, and accountability.
Management	Is the organization translating that direction into effective processes, controls, and monitoring?	Ownership, policies, data, access, cybersecurity, lifecycle controls, third parties, testing, reporting, and risk management.
Internal Audit	How should Internal Audit advise or provide assurance without taking management responsibility?	Scope, competence, criteria, evidence, testing, specialist support, conclusions, and communication.

Governance: set direction and demand accountability

In the framework, AI governance refers to the structures and processes used to direct, manage, and monitor AI activities. The emphasis is not the algorithm itself. It is whether the organization knows why it is using AI, what it intends to achieve, and within what boundaries.

That is why the framework starts with strategy. AI objectives should be linked to organizational objectives, strategy should not be developed in isolation, and values, risk, and ethical, social, and legal consequences should be considered. The framework also points to governing-body involvement when AI could materially alter strategy or the organization’s risk profile.

For Internal Audit, this domain translates into questions about the quality of board oversight, alignment between strategy and risk, accountability, and acceptable-use boundaries.

Management: turn intent into controls that operate

The Management domain moves from direction to execution. It encompasses first-line responsibilities for designing and operating controls and second-line activities such as risk management and compliance that provide expertise, monitoring, and challenge.

The topics include leadership, acceptable-use policies, data, access, cybersecurity, change management, third parties, testing, monitoring, reporting, and risk treatment. For a new AI auditor, the key insight is that many AI failures materialize through familiar business and technology processes. Weak data governance or vendor oversight can be as important as a model weakness.

Internal Audit: advise and assure without becoming the owner

The third domain focuses on Internal Audit’s role. The framework contemplates both advisory and assurance services. In a lower-maturity organization, Internal Audit may add value by helping leaders frame questions, identify risks, and assess the governance architecture. In a more mature environment, the function can provide assurance over established processes and controls.

The boundary is essential: Internal Audit can advise on how AI should be governed, but management must retain responsibility for designing, implementing, and operating controls. Independence does not require isolation; it requires avoiding a situation where Internal Audit later evaluates decisions it effectively owned.

No audit team will master every technical dimension of AI. The goal is to develop enough knowledge to assess risk, request appropriate evidence, and recognize when specialist support is needed.

How to turn the framework into an audit engagement

The document includes a practitioner’s guide, but its strongest use is to translate its considerations into an audit chain. A simple five-step approach works well.

1. Understand how the organization uses AI. Identify the strategy, current and planned use cases, accountable owners, and the inventory of systems or applications.

2. Select the relevant domains. A governance review may concentrate on strategy, oversight, and accountability. An engagement focused on a specific AI use case will usually need deeper Management coverage: data, changes, testing, third parties, monitoring, and operational controls. The Internal Audit domain shapes how assurance or advisory work should be delivered.

3. Convert expectations into criteria and expected controls. Suppose the organization requires high-impact AI models to receive an independent review before production. The audit criterion may be the approved policy; the expected control may be documented independent approval; and the evidence may include review records, exceptions, and release decisions.

4. Test operation, not just existence. Select AI use cases, inspect approvals and test evidence, and reconcile inventories or vendor changes to determine whether controls actually operate.

5. Conclude within the team’s competence. A governance review can often be performed by internal auditors with appropriate AI literacy. Statistical model validation, adversarial robustness testing, or deep technical architecture assessment may require specialists. That boundary should be explicit in the scope rather than hidden in the conclusion.

This approach also connects the AI framework to the Global Internal Audit Standards. Competence and due professional care, risk-based planning, evaluation criteria, relevant and reliable information, documentation, conclusions, and communication remain professional requirements even when AI is the subject matter.

The most common mistake: treating it as a universal checklist

The framework contains many useful considerations and questions, but The IIA states that its practitioner’s guide should be customized to the organization’s level of AI use and maturity. That caveat should change how auditors use the document.

An organization that only permits approved generative AI tools does not need the same audit program as a company developing proprietary models for lending, hiring, pricing, or autonomous actions. Nor does the existence of an AI policy prove that governance is effective. Relevance depends on the use case, impact, data, third parties, level of autonomy, and applicable legal or contractual requirements.

The framework therefore works better as a question map than as a binary compliance list. It helps identify what might need to exist, but the auditor still has to determine materiality, applicable criteria, persuasive evidence, and the depth of testing justified by the risk.

How it fits with the Global Internal Audit Standards

The AI framework explains the subject matter; the Global Internal Audit Standards govern professional audit practice. They are complementary, not competing, sources.

The Global Internal Audit Standards require Internal Audit to understand governance, risk management, and control, plan based on risk, and work with appropriate criteria and evidence. The AI framework supplies topic-specific content for applying those requirements to artificial intelligence, but assurance quality still depends on methodology, evidence, and professional judgment.

The IIA’s Three Lines Model reinforces the same separation of responsibilities: the governing body oversees, management acts to achieve objectives and manage risk, and Internal Audit provides independent assurance and advice. That distinction is particularly important in AI because the pressure to “help” deploy new capabilities can pull Internal Audit too close to management decisions.

A framework for starting, not finishing

The IIA AI Auditing Framework is useful because it translates a technically intimidating subject into a structure familiar to internal auditors: direction, risk management, controls, evidence, and assurance.

Its most mature use is not to audit every line of the document. It is to use the framework to understand who should do what, identify where that architecture can fail, and convert applicable expectations into audit procedures proportionate to risk.

AI will continue to change faster than any single guidance document. The framework therefore treats audit capability as a progression. Internal Audit does not need to own AI or become the ultimate technical expert on every model. It needs enough competence, independence, and rigor to ask the right questions, evaluate reliable evidence, and know when a conclusion requires specialist support.