AI adoption is moving faster than most organizations can redesign their assurance models. New tools are being embedded in customer interactions, operations, decision making, etc., across the enterprise. For chief audit executives (CAEs), the central question that emerges is whether AI risk is embedded in the way the entire audit function assesses risk and provides assurance.
A standalone AI governance audit may still be useful, but alone, it is not enough. Policies, committees, inventories, and approvals can show that governance exists, but they do not demonstrate that risk is understood where AI is actually used, that controls are operating effectively, or that the intended business and customer outcomes are being achieved. Internal audit needs an approach that follows AI into the business process.
The Imperative for CAEs
Internal audit should establish a common AI risk lens, apply it to a complete and risk-tiered inventory of AI use cases, and assess accountability through the organization’s existing ownership and control structure.
1. How AI Risk Is Embedded Into the Internal Audit Methodology
AI is a horizontal risk that can alter decisions, data flows, control execution, customer outcomes, operational resilience, third-party dependencies, and the way employees exercise judgment. It is also a risk consideration that should influence the audit universe, annual and dynamic risk assessment, engagement scoping, fieldwork, testing, reporting, and follow-up.
The shift is from asking, “Do we have an AI audit?” to asking, “How does AI change the risk profile and control expectations of every relevant audit?” That question should be built into the core methodology rather than left to individual teams or a small group of specialists.
What an Embedded Assessment Looks Like
- Risk assessment and planning: Identify where AI is used or materially influences the process, decision, control, output, or customer experience; reflect that exposure in the risk assessment and audit coverage decisions.
- Scoping and walkthroughs: Determine the purpose of the use case, where it sits in the process, what decisions or actions it affects, what data it uses, who relies on its output, and what happens when it is wrong or unavailable.
- Control evaluation: Assess whether existing controls remain fit for purpose and whether AI creates new expectations for authorization, data quality, access, validation, explainability, monitoring, human review, change management, incident response, and third-party oversight.
- Testing: Go beyond evidence that an approval occurred. Test whether the use case and its controls produce the intended outcomes within defined tolerances and risk appetite.
- Reporting and follow-up: Connect findings to the responsible process and risk owners, describe the business consequence, and monitor whether remediation remains effective as the use case, model, data, or workflow changes.
- Ongoing assurance: Use trigger-based or more continuous monitoring for higher-risk use cases rather than relying solely on an annual audit cycle.
Six-Pillar Lens: Support the Assessment, Maintain Ownership
A principles-based framework gives auditors a consistent way to assess AI across different businesses and technologies. The six pillars below should be used as an enterprise risk lens for each use case. They organize the questions and control expectations; they do not create a separate AI organization or transfer accountability away from existing owners.
| Risk Lens | What Internal Audit Should Consider |
|---|---|
| Governance & Policy | Purpose, approvals, accountability, acceptable use, escalation, oversight, and alignment with enterprise policy and risk appetite. |
| Risk Taxonomy & Assessment | Identification, classification, tiering, aggregation, monitoring, and reassessment of risks created or amplified by the use case. |
| Model Risk Management | Development and validation expectations, performance, explainability, limitations, change, drift, and ongoing monitoring where applicable. |
| Technology, Data Management, & Cybersecurity | Architecture, access, data provenance and quality, privacy, security, integration, logging, and technical change controls. |
| Operational Resilience & Third-Party Management | Service continuity, fallback processes, concentration, incident response, vendor dependencies, contractual expectations, and exit planning. |
| Talent & AI Literacy | Role-appropriate knowledge, informed use, effective human challenge, escalation, documentation, and understanding of control responsibilities. |
Used consistently, this lens creates a single enterprise view of AI risk while still allowing audit scope and testing depth to reflect the use case, its risk tier, and the organization’s maturity.
2. Make the Risk-Tiered Use-Case Inventory the Foundation
Internal audit cannot provide complete assurance over an AI footprint the organization cannot see. A reliable enterprise inventory is, therefore, foundational—but it must inventory AI in use, not merely the applications, platforms, or models the organization has acquired.
An application inventory answers what technology exists. A use-case inventory answers how AI is being applied inside a business process: the objective, decision or activity supported, users, data, output, affected stakeholders, technology dependencies, owner, risks, controls, and potential impact. The same application may support several use cases with very different risk profiles. Conversely, a seemingly low-risk model can create significant exposure when it influences a consequential decision or high-volume process.
The use case is, therefore, the more useful unit of risk and assurance.
Minimum Information for an Assurance-Ready Inventory
- Business process, purpose, and decision point or activity affected
- Named business/process owner and relevant risk and control owners
- Users, customers, employees, or other stakeholders affected
- Technology, model, data, interfaces, and third-party dependencies
- Level of autonomy and nature of human involvement or override
- Potential financial, customer, regulatory, operational, legal, privacy, security, resilience, and reputational impacts
- Risk tier, rationale, applicable standards, required controls, monitoring, and review cadence
- Lifecycle status, material changes, incidents, exceptions, and retirement or exit requirements
Risk Tiering Must Trigger Action
A tier is useful only if it changes what the organization does. Tiering criteria should consider the consequence of error or misuse, scale and reach, customer or employee impact, regulatory relevance, sensitivity of data, degree of autonomy, explainability, complexity, reliance on third parties, reversibility, and availability of effective human or system safeguards.
| Illustrative Tier | What the Tier Should Drive |
|---|---|
| Higher-risk use cases | Enhanced approval and independent challenge; stronger validation and testing; explicit human oversight and escalation; more frequent monitoring and reporting; tighter change controls; and greater internal audit attention. |
| Moderate-risk use cases | Proportionate review, documented control requirements, periodic performance and control monitoring, reassessment after material change, and targeted audit coverage. |
| Lower-risk use cases | Baseline registration, acceptable-use and data controls, owner attestation where appropriate, event-driven reassessment, and streamlined oversight without eliminating accountability. |
Internal audit should assess both the design of the tiering methodology and whether the assigned tier actually drives differentiated governance, control, monitoring, testing, reporting, and assurance. An inventory that records use cases without connecting them to requirements creates visibility but not risk management.
3. Anchor AI Accountability in the Existing Risk Ownership Structure
AI generally should not create a parallel risk ownership model. The business or process owner remains accountable for the process, its outcomes, and its risks and controls—whether the activity is performed manually, through traditional technology, or with AI. Introducing AI may change the risk profile and the controls required, but it does not eliminate or transfer that accountability.
Specialists in model risk, cybersecurity, data, technology, privacy, legal and compliance, operational resilience, procurement, and third-party risk remain essential. Their role is to establish domain standards, provide expertise, perform required review or challenge, and support monitoring and escalation. It is not to become the default owner of every risk that appears in an AI-enabled process.
Avoid Fragmented Ownership
When responsibility is distributed across committees and specialist functions without one accountable process owner, approvals can multiply while gaps remain. Internal audit should be able to trace each material use case to a named owner, applicable risk owners, control owners, required specialist challenge, and a clear escalation path.
Education Is a Control Prerequisite
Ownership is meaningful only when the owner understands what is being owned. AI literacy should, therefore, be treated as a prerequisite for effective risk and control ownership, not as a general awareness exercise. Education should be tailored to the decisions each role must make and the evidence each role must produce.
| Role | Education & Understanding Required |
|---|---|
| Business and process owners | Understand the use case, its limitations and dependencies, the risks introduced into the process, the required controls, performance expectations, and escalation triggers. |
| Control owners and operators | Understand how AI changes the control objective or execution, what evidence demonstrates performance, how exceptions are resolved, and how to avoid automation bias or routine “rubber-stamp” review. |
| Risk and specialist functions | Understand where their standards apply, how they challenge the use case, what constitutes an acceptable exception, and how issues are aggregated and escalated. |
| Internal auditors | Understand enough about AI, the business process, data, and applicable risk domains to scope the engagement, involve specialists when needed, challenge management, test outcomes, and reach a defensible conclusion. |
Clear control expectations are equally important. Policies should translate into actionable requirements that owners can apply at the use-case level. Internal audit should be able to determine what controls were required because of the use case and its tier, who was responsible for them, what evidence demonstrates their operation, and how exceptions or changes are handled.
Audit Outcomes, Not Just Approvals
Traditional evidence—committee approval, a completed risk assessment, a validation memo, or a documented human review—remains relevant, but it does not by itself demonstrate that an AI-enabled process is controlled or effective. Internal audit should test the result as well as the governance surrounding it while considering the following:
- If AI is expected to reduce processing time, determine whether cycle time improved without unacceptable deterioration in accuracy, control, or customer outcomes.
- If AI is expected to reduce errors, examine error and exception trends, including whether errors shifted to a different point in the process.
- If a human remains in the loop, assess whether the person has the knowledge, authority, time, information, and incentives to provide a meaningful challenge.
- If the control relies on human review, monitor whether effectiveness degrades as users become more familiar with or overly reliant on AI outputs.
- If a use case changes through model updates, new data, prompt or workflow changes, vendor releases, or expanded deployment, determine whether reassessment and control change were triggered.
For the highest-risk use cases, assurance may need to become more dynamic. Periodic audits can be supplemented by indicators tied to performance, incidents, override behavior, drift, exceptions, user reliance, and control operation. The purpose is not to make internal audit the owner of monitoring; it is to use current risk signals to determine when assurance coverage should change.
A Practical Agenda for CAEs
1. Embed the Lens
Update the audit methodology and risk assessment so teams identify and assess relevant AI use in planning, scoping, walkthroughs, testing, reporting, and follow-up.
2. Validate the Inventory
Assess whether the enterprise inventory captures AI use cases comprehensively and maps each to a process, owner, technology, risks, controls, tier, and affected stakeholders.
3. Challenge Tiering
Determine whether tier criteria are consistently applied and whether tiers trigger proportionate governance, control, monitoring, reporting, and audit requirements.
4. Confirm Accountability
Trace material use cases to existing process, risk, and control owners; clarify specialist roles and escalation paths; and identify fragmented or orphaned ownership.
5. Assess Education
Evaluate whether owners, operators, specialists, and auditors have role-appropriate knowledge of the risks, limitations, controls, evidence, and escalation expectations relevant to them.
6. Test Outcomes & Adapt
Test performance and control outcomes, monitor degradation and material change, and adjust assurance as risk signals evolve.
The Bottom Line
Three Disciplines Make AI Assurance Durable
Build AI risk assessment into the audit methodology and every stage of the audit lifecycle.
Treat the use case—not the application—as the unit of risk, tiering, and assurance.
Align AI accountability to current process, risk, and control owners, supported by education and specialist challenge.
Internal audit does not need to predict every development in AI. It does need a repeatable way to understand where AI is used, how it changes risk, who owns the outcome, what control is expected, and whether that control remains effective. When those disciplines are connected, internal audit can provide assurance that keeps pace with innovation without creating a parallel risk universe.
How Forvis Mazars Can Help
As organizations move from experimentation to broader AI adoption, internal audit functions need practical, scalable ways to embed AI risk into assurance. Professionals at Forvis Mazars can help organizations:
- Embed AI risk considerations into internal audit risk assessment, planning, methodology, audit programs, and reporting
- Design or assess enterprise AI use-case inventories, risk-tiering methodologies, and tier-based control and monitoring requirements
- Align AI governance and assurance to existing process, risk, and control ownership structures
- Define role-specific control expectations, accountability, and evidence requirements
- Develop education and capability programs for process owners, control owners, risk functions, and internal auditors
- Perform AI governance, use-case, control framework, or internal audit readiness assessments
- Design horizontal and more continuous assurance approaches for higher-risk AI use cases
Whether your organization is establishing foundational oversight or expanding assurance over more advanced deployments, Forvis Mazars can help build a risk-focused approach that supports responsible innovation while strengthening accountability and control.
Contact a professional at Forvis Mazars to discuss your organization’s AI risk and internal audit priorities.