Key Definitions
| Term | Definition |
|---|---|
| AI System | A machine-based system designed to operate with varying levels of autonomy, that may exhibit adaptiveness and that infers from inputs to generate outputs such as predictions, content, recommendations, or decisions (EU AI Act Art.3(1)) |
| AI Audit | A systematic, independent, and documented process for obtaining audit evidence and evaluating it objectively to determine the extent to which audit criteria are fulfilled, applied specifically to AI systems and their governance |
| Conformity Assessment | The process demonstrating whether specified requirements relating to an AI system have been fulfilled (EU AI Act Art.3(20)) |
| Audit Criteria | The set of requirements used as a reference against which audit evidence is compared (ISO 19011:2018 §3.7) |
| Audit Evidence | Records, statements of fact, or other verifiable information relevant to the audit criteria (ISO 19011:2018 §3.9) |
| Audit Finding | Results of the evaluation of collected audit evidence against audit criteria (ISO 19011:2018 §3.10) |
| High-Risk AI System | An AI system falling within categories listed in Annex III of the EU AI Act, subject to enhanced requirements |
| Risk Management System | A continuous iterative process planned and run throughout the entire lifecycle of a high-risk AI system (EU AI Act Art.9) |
| Technical Documentation | The documentation required under Art.11 and Annex IV of the EU AI Act for high-risk AI systems |
| Audit Programme | Arrangements for a set of one or more audits planned for a specific time frame and directed toward a specific purpose (ISO 19011:2018 §3.13) |
| Nonconformity | Non-fulfilment of a requirement identified during the audit process |
| Corrective Action | Action taken to eliminate the cause of a detected nonconformity and to prevent recurrence |
| Post-Market Monitoring | Activities performed by providers to proactively collect and review experience gained from AI systems placed on the market (EU AI Act Art.72) |
Chapter 1: Foundations of AI Auditing
AI auditing is the structured examination of an artificial intelligence system's design, development, deployment, and ongoing operations against established regulatory, ethical, and technical standards. Unlike traditional IT audits that focus primarily on data integrity and access controls, AI audits must address unique challenges including algorithmic transparency, bias detection, explainability requirements, and the dynamic nature of machine learning models that evolve over time.
1.1 Why AI Auditing Matters Now
The regulatory landscape for artificial intelligence has shifted from voluntary guidelines to binding obligations. The EU AI Act (Regulation 2024/1689), which entered into force on 1 August 2024, establishes a comprehensive legal framework with phased enforcement:
- 2 February 2025: Prohibited AI practices provisions apply (Art.5)
- 2 August 2025: General-purpose AI model obligations apply (Chapter V)
- 2 August 2026: Full application of high-risk AI system requirements (Annex III)
Organizations deploying AI systems in the European Union — or whose AI systems affect EU residents — must demonstrate compliance through systematic auditing. The penalties for non-compliance are significant: up to EUR 35 million or 7% of total worldwide annual turnover for prohibited practices violations (Art.99).
Beyond regulatory compliance, AI auditing serves critical business functions:
- Risk mitigation — Identifying potential failures before they cause harm
- Stakeholder trust — Demonstrating responsible AI governance to customers, regulators, and the public
- Operational excellence — Ensuring AI systems perform as intended and deliver expected value
- Legal protection — Creating documented evidence of due diligence
- Continuous improvement — Building feedback loops that strengthen AI governance over time
1.2 Audit Scope: What Gets Audited
AI audits can examine any combination of the following dimensions:
| Dimension | Description | Key Questions |
|---|---|---|
| Technical | Algorithm design, model performance, data quality | Does the model perform accurately? Is it robust? |
| Ethical | Fairness, bias, human rights impact | Does the system discriminate? Are affected parties considered? |
| Legal/Regulatory | Compliance with applicable laws and standards | Does the system meet EU AI Act requirements? |
| Governance | Organizational policies, roles, oversight structures | Is there adequate human oversight? |
| Operational | Deployment practices, monitoring, incident response | How are failures detected and addressed? |
| Data | Data collection, storage, processing, quality management | Is training data representative and properly governed? |
| Security | Cybersecurity measures, adversarial robustness | Is the system protected against attacks and manipulation? |
1.3 Types of AI Audits
Different audit types serve different purposes within an organization's AI governance framework:
Internal AI Audit
Conducted by the organization's own audit team or internal compliance function. These provide ongoing assurance and early detection of issues. Internal audits should follow the independence principles of ISO 19011, ensuring auditors are independent from the activities they audit.
External AI Audit
Performed by independent third parties to provide objective assurance to stakeholders. For high-risk AI systems under the EU AI Act, external conformity assessments may be required through notified bodies (Art.43).
Regulatory AI Audit
Conducted by or on behalf of regulatory authorities (market surveillance authorities under the EU AI Act). Organizations must cooperate and provide access to documentation, data, and systems (Art.74).
Supplier/Vendor AI Audit
Assessment of third-party AI systems or components before or during procurement. Critical for organizations deploying AI systems developed by external vendors.
Pre-deployment AI Audit
Conducted before an AI system goes live to verify it meets all requirements. For high-risk systems, this includes conformity assessment (Art.43).
Operational AI Audit
Ongoing or periodic examination of AI systems already in production. Focuses on continued compliance, performance monitoring, and drift detection.
1.4 Regulatory Framework Mapping
The following table maps key regulatory requirements to audit activities:
| Regulation/Standard | Requirement | Audit Activity |
|---|---|---|
| EU AI Act Art.9 | Risk management system | Verify risk management lifecycle process |
| EU AI Act Art.10 | Data governance | Examine data quality management practices |
| EU AI Act Art.11 | Technical documentation | Review completeness of Annex IV documentation |
| EU AI Act Art.12 | Record-keeping | Test logging and traceability mechanisms |
| EU AI Act Art.13 | Transparency | Assess user information and disclosure practices |
| EU AI Act Art.14 | Human oversight | Evaluate human-in-the-loop/on-the-loop controls |
| EU AI Act Art.15 | Accuracy, robustness, cybersecurity | Test technical performance requirements |
| ISO/IEC 42001 §6.1 | Risk assessment | Review AI risk assessment methodology |
| ISO/IEC 42001 §8.1 | Operational planning | Examine AI lifecycle management |
| NIST AI RMF Govern | Governance structures | Assess organizational AI governance |
| NIST AI RMF Map | Context and risk identification | Review risk identification processes |
| NIST AI RMF Measure | Risk analysis and tracking | Evaluate measurement methodologies |
| NIST AI RMF Manage | Risk response | Assess risk mitigation implementation |
1.5 The AI Audit Lifecycle
The AI audit lifecycle consists of five major phases, each detailed in subsequent chapters:
- Phase 1 — Audit Planning (Chapter 2): Defining scope, objectives, criteria, and resources
- Phase 2 — Preparation (Chapter 3): Risk assessment, audit programme design, team assembly
- Phase 3 — Fieldwork (Chapter 4-5): Evidence collection, testing, analysis
- Phase 4 — Reporting (Chapter 6): Findings documentation, recommendations, communication
- Phase 5 — Follow-Up (Chapter 7): Corrective actions, verification, continuous monitoring