The Risks of Auditing AI Without a Rulebook

Artificial intelligence, particularly generative AI, is rapidly being embedded into core business processes, influencing how organizations analyze information, interact with customers, and make decisions. But while adoption accelerates, governance expectations are still forming, and internal audit is being asked to provide reasonable assurance without clear standards or consistent practices.

The challenge is not awareness of AI risk, but clarity around what to audit, how to assess it, and what good looks like.

What Falls Within AI Audit Scope?

The first challenge is defining scope. With traditional systems, audit boundaries are clear: applications, data flows, controls, and outputs are evaluated within a defined process. With AI, those boundaries blur. Risk does not sit neatly in one place, because AI operates across a chain of dependencies rather than a single system.

So, what is in scope exactly?

• The Model Itself: How it was selected, configured, maintained, and overseen. This includes understanding who made decisions about model selection, what assumptions were built into it, and whether there is appropriate oversight as it evolves.
• The Data Feeding the Model: Data carries risk around quality, completeness, bias, and lineage. Inaccurate or biased data can influence outputs in ways that are not immediately visible but still material to the business.
• The Outputs: Organizations increasingly rely on outputs, raising the question of how those are used and validated.
• The Business Use Case: The greatest risk is often the degree of reliance placed on technology. An AI-supported decision used for internal analysis carries a far different risk profile than one affecting financial reporting, customer outcomes, or regulatory obligations.

In practice, risk exists across all of these dimensions, as they are interconnected. Weaknesses in one area can ripple across the entire system, often without clear visibility. That creates real complexity for internal audit.

Most audit functions are structured to evaluate controls within defined processes, not across an interconnected ecosystem where ownership is distributed and boundaries are not clearly defined. As a result, audit coverage often becomes fragmented, with different pieces assessed in isolation rather than as part of a continuous risk flow.

Navigating AI Without a Single Framework

Even when scope is understood, internal audit faces a second challenge: there is no single framework that defines how AI governance should be evaluated.

Meaningful guidance exists and continues to evolve.

• COSO (Committee of Sponsoring Organizations of the Treadway Commission) extends its internal control framework to address emerging technology risks.
• NIST (National Institute of Standards and Technology) offers an AI Risk Management Framework focused on identifying and mitigating AI risks.
• COBIT (Control Objectives for Information and Related Technologies) provides a governance perspective linking technology to enterprise objectives.

Each framework approaches the problem from a different angle, and together they offer valuable building blocks. Individually, however, none provides a complete or prescriptive blueprint for how internal audit should evaluate AI.

At the same time, regulatory expectations continue to evolve, with growing emphasis on responsible AI, transparency, and data governance. Yet audit and assurance guidance has not kept pace. It remains largely high-level, leaving internal audit with limited direction on how to assess AI risk in practice, particularly in areas such as financial reporting and decision-making where the stakes are highest.

Building Visibility and Defining Risk

Despite uncertainty, meaningful progress is not only possible; it is already underway. Importantly, it does not depend on waiting for a perfect framework to emerge. Instead, it begins with building visibility.

For many organizations, a clear understanding of how AI is being used does not yet exist. Internal audit can close this gap by identifying shadow AI, assessing AI embedded in third-party tools, and mapping how these technologies connect to business processes. Building an AI inventory and classifying use cases by risk, reliance, and impact enables a shift from reactive reviews to a more focused, risk-based approach.

Attention then turns to governance. While no single framework provides a complete answer, organizations are adopting hybrid models that combine COSO, NIST, and COBIT, aligning them to how AI is actually used. The framework foundation remains critical; however, its effectiveness ultimately depends on how it is operationalized. Clearly defined roles, expectations for model and data use, and consistent control points create a practical foundation. Without this operational backbone, internal audit lacks a reliable baseline, and reasonable assurance remains subjective and difficult to scale.

Evolving Audit Approaches for AI

As governance matures, audit approaches themselves must evolve. Leading internal audit functions are moving beyond traditional control testing and incorporating more dynamic techniques, such as scenario-based testing to evaluate how AI behaves under real-world conditions. This evolution is supported by ongoing investment in upskilling internal teams, as well as the strategic use of co-sourcing to bring in specialized expertise where needed.

Together, these shifts reflect a broader transition from trying to fit AI into existing audit approaches to reshaping those approaches to match the complexity of AI itself.

Organizations that move early to bring structure and discipline to AI risk will be better positioned to build trust, meet evolving expectations, and scale their use of these technologies with confidence. For those looking to accelerate that journey, engaging with our Risk Solutions Practice can provide practical guidance, from establishing AI inventories and governance frameworks to enhancing audit approaches and building internal capabilities.