本文目前仅提供英文版本。
AI Internal Audit: What Audit Committees Should Demand and How to Test AI Controls
AI is now a material risk for most organisations, but few internal audit functions have developed the methodology to audit it effectively. Here is the framework for auditing AI — what to test, how to test it, and what good AI audit evidence looks like.
Key Takeaways
AI internal audit is not the same as technology audit. Testing whether a server is patched is different from testing whether an AI model is producing biased outcomes. AI audit requires a specific methodology that most internal audit functions are still developing.
The audit committee's role in AI: request a complete inventory of material AI systems from management; confirm that AI risk is explicitly within the internal audit universe; ensure internal audit has access to the technical expertise (whether in-house or co-sourced) needed to audit AI effectively; and receive annual reporting on AI audit findings.
The five AI audit workstreams: governance (AI risk framework, board oversight, policies); model development and validation (pre-deployment review, testing, documentation); data governance (training data quality, consent, representativeness); ongoing monitoring (performance monitoring, bias testing, incident response); and third-party AI (vendor due diligence, contractual provisions, ongoing oversight).
AI bias testing methodology: internal audit should request demographic disparity analysis for any AI system making consequential decisions about individuals. Where material disparities exist, audit should assess whether they are justified by actuarially sound risk factors or indicate potential proxy discrimination.
Sampling AI decisions: a core audit technique is sampling a statistically meaningful number of AI decisions and reviewing them for accuracy, consistency, and compliance with governance requirements. Auditors should be able to trace a decision back through the AI system to the inputs and model parameters that produced it.
AI audit findings most commonly identify: no formal AI governance framework; AI systems not captured in the risk register; model changes deployed without revalidation; no demographic disparity monitoring; vendor AI not subject to the same oversight as internally developed AI.
"仅供参考。本文不构成法律、监管、财务或专业建议。如需具体指导，请咨询合格专家。"
Why AI audit is different
Internal audit of AI systems requires capabilities that most internal audit functions are still building. Traditional technology audit focuses on system security, availability, change management, and access controls — all of which remain relevant for AI, but are insufficient. AI audit additionally requires capability to assess: whether a model is producing statistically unbiased outcomes across demographic groups; whether a model is performing as validated; whether the training data was appropriate; and whether human oversight of AI decisions is genuine rather than nominal.
The IIA (Institute of Internal Auditors) Global Technology Audit Guide on AI provides a framework, and ISACA has published guidance on auditing AI and machine learning. These should be the starting reference point for internal audit functions developing AI audit capability.
The audit committee's AI responsibilities
Audit committees have direct responsibility for oversight of internal audit and for reviewing the organisation's risk management effectiveness. For AI, this translates to specific responsibilities that audit committees should be explicitly discharging.
Request an AI system inventory: audit committees should ask management to provide a complete inventory of material AI systems — what AI the organisation uses, for what decisions, at what scale, and what the key risks are. This inventory should form the basis for audit coverage planning.
Confirm AI is in the audit universe: internal audit's annual audit plan should explicitly include AI risk. Audit committees should satisfy themselves that AI audit coverage is proportionate to AI risk materiality — organisations where AI is central to business processes should have more extensive AI audit coverage than those where AI use is peripheral.
Ensure technical capability: audit committees should understand how internal audit is addressing the technical complexity of AI audit. Options include building in-house AI audit capability, co-sourcing with specialist firms, or using external technical experts to support audit fieldwork. Audit committees should confirm that the approach is adequate for the organisation's AI risk profile.
Five AI audit workstreams
Governance and framework: does the organisation have a documented AI governance framework? Is AI risk explicitly in the risk register? Is there board and senior management oversight? Are AI governance policies approved and current? Is there an AI model inventory? Audit objective: confirm that the AI governance framework is designed appropriately and operating as intended.
Model development and validation: is there a pre-deployment review process for AI systems? Does that process include technical validation, bias testing, documentation requirements, and approval authority? Are model changes subject to the same pre-deployment requirements as initial deployments? Audit objective: confirm that AI models are not deployed without appropriate validation and governance.
Data governance: how is training data selected and approved? Are data quality standards documented and tested? Is consent or lawful basis documented for personal data used in training? Is training data assessed for representativeness across demographic groups? Audit objective: confirm that training data governance is adequate to prevent biased or non-compliant AI outputs.
Ongoing monitoring: is there a performance monitoring process for each material AI system? Are demographic disparity tests conducted regularly? Is there an AI incident response process? Are incidents investigated and root causes addressed? Audit objective: confirm that AI performance is monitored and deteriorating performance is identified and addressed.
Third-party AI: is vendor AI subject to the same governance requirements as internally developed AI? Is there an AI-specific vendor due diligence process? Are vendor bias audit results obtained and reviewed? Are contractual provisions for AI vendors adequate? Audit objective: confirm that vendor AI risk is managed equivalently to internal AI risk.
Testing AI bias: a core audit technique
Testing for demographic bias in AI outputs is a distinctive AI audit technique. The methodology: obtain a representative sample of AI decisions; segment the sample by demographic characteristics relevant to the decision (for lending AI: race, gender, age, national origin; for employment AI: the same plus disability status); compare decision outcomes across demographic segments; and assess whether material disparities are present and, if so, whether they are explicable by legitimate risk factors or indicate potential discrimination.
For example, auditing a credit underwriting AI: obtain a sample of 1,000 recent underwriting decisions; segment by race using ZIP code proxies or applicant-disclosed data if available; compare approval rates and average credit limits across segments controlling for credit score and debt-to-income ratio; if Black applicants are approved at materially lower rates than white applicants with equivalent financial profiles, this warrants further investigation.
The audit finding is not that bias exists — it is that bias testing was not conducted, or that bias was detected and management did not investigate it adequately, or that bias was detected and the AI system was not appropriately modified. Internal audit should not reach conclusions about whether an AI model is discriminatory — that requires legal and technical expertise beyond audit scope. But audit can and should identify whether bias testing is being conducted, whether results are reviewed, and whether appropriate action follows concerning findings.