How to use this book
Caution
Not legal advice or an audit opinion. This book is a course book for preparing for the AIDAN Specialty track. It merely summarizes and explains auditing standards and statutes to facilitate understanding, and does not constitute a legal opinion or audit conclusion regarding any specific organization or incident. In an actual audit, verify the original text of the standards followed by your organization and the statutes in effect on that day, and if necessary, obtain a review by legal counsel and applicable subject-matter experts.
Verification date: October 22, 2026. All standards and statutes cited in this book were compiled after reading the original texts on the official pages of each institution as of this date.
- Framework Act on the Development of Artificial Intelligence and Establishment of Trust-Based Foundations (abbreviated as the AI Basic Act) [Enforced July 21, 2026] [Act No. 21311, Jan. 20, 2026, Partial Amendment]
- Enforcement Decree of the same Act [Enforced Aug. 20, 2026] [Presidential Decree No. 36580, Aug. 18, 2026, Amendment by Other Act]
- Personal Information Protection Act (PIPA) [Enforced Sept. 11, 2026] [Act No. 21445, Mar. 10, 2026, Partial Amendment]
- The IIA Global Internal Audit Standards 2024 Edition, U.S. GAO Government Auditing Standards 2024 Revision (Yellow Book), PCAOB Auditing Standards AS 1105, 2201, 2315, 2601
- NIST AI RMF 1.0 and Generative AI Profile (NIST AI 600-1), GAO Artificial Intelligence Accountability Framework (GAO-21-519SP), ISO/IEC 42001:2023 guidance page
PCAOB standards are for public company financial statement audits, and the Yellow Book is for U.S. government audits. This book borrows the principles of evidence, sampling, and control evaluation formulated by both standards and applies them to AI audits. If your organization has separate standards that it must follow, those standards take precedence.
Who this book is for
This book was written for three groups of readers. First are internal and external auditors who must include an organization's use of AI in an audit plan for the first time. Second are risk and compliance managers who design and inspect AI-related controls. Third are audit committees and their support organizations that receive AI governance reporting. While the questions from these three readers differ, they ultimately converge on the same five tasks: identifying what exists, examining whether controls are properly designed and actually operating, gathering evidence to substantiate that judgment, verifying whether models and data are as reliable as claimed, and reporting findings so they can be remediated and verified through to completion.
An understanding of AI at the White belt (AI fundamentals) level and even a modest amount of experience in audit and internal control work will be sufficient. To ensure that those new to auditing can follow along, terms are explained in plain language whenever they first appear, and statistics are covered at the level of basic arithmetic. Conversely, data scientists who know AI well but are new to auditing can also use this book to understand what auditors ask and why.
Upon completing this book, you should be able to produce the following deliverables yourself: an AI inventory whose completeness is verified through independent evidence; an audit plan whose depth is determined by risk; a test plan for human review controls; a sampling design tailored to audit objectives; a model performance claim verification table; and audit findings equipped with criteria, condition, cause, and effect, along with a follow-up verification plan. The two hands-on tasks in each chapter guide you to produce these deliverables step by step.
This course does not cover financial statement audits in general, the formulation of audit opinions, or directly developing or tuning models. Statutes are covered only to the extent they are used as audit criteria, leaving conclusions on whether a particular system is subject to legal obligations to legal counsel's judgment.
How this book is organized
The main text consists of five chapters arranged in the same order as the exam domains. Chapter 1 covers planning to determine where to direct the audit; Chapter 2 covers evaluation of whether controls are properly designed and functioning; Chapter 3 covers evidence and sampling to support that judgment; Chapter 4 covers verification of models and data, which are risks unique to AI; and Chapter 5 covers reporting and follow-up actions. It is structured so that deliverables produced in earlier chapters serve as materials for later chapters. The AI inventory from Chapter 1 becomes the subject of control evaluation in Chapter 2, the test plan from Chapter 2 leads to the sampling design in Chapter 3, and the verification results from Chapter 4 become the findings in Chapter 5.
Each chapter consists of ten sections: Learning objectives, Introduction, Key concepts, Going deeper (trade-offs and edge cases), At work (hypothetical case studies showing even discarded alternatives), Common mistakes, Summary at a glance, Check the official docs, Self-check (explanations in Appendix B), and Hands-on tasks.
The appendixes contain a glossary (A), Self-check explanations (B), an exam-day guide (C), and a list of official documents cited in this book (D). When reading for the first time, read sequentially by chapter; when reviewing, we recommend starting with "Summary at a glance" and "Common mistakes" in each chapter.
The exam at a glance
AI Audit (AUD) is a Specialty track. The Specialty exam consists of 35 questions to be answered in 60 minutes, and passing requires earning 75% or more of the total points. The free practice set consists of 10 questions and 25 minutes.
| Chapter | Domain | Weight |
|---|---|---|
| Chapter 1 | Audit Planning, Scope, and AI Inventory | 20% |
| Chapter 2 | Control Design and Operating Effectiveness | 20% |
| Chapter 3 | Evidence Collection, Testing, and Sampling | 25% |
| Chapter 4 | Model and Data Risk Verification | 20% |
| Chapter 5 | Reporting and Follow-up Actions | 15% |
Chapter 3 carries the largest weight. This is because audit conclusions ultimately depend on the quality and quantity of evidence, and AI systems generate vast amounts of a new kind of evidence in the form of logs and automatically generated materials. Questions on statutes are based on statutes in effect at the time of the exam, while questions on standards are based on original texts as of the verification date of this book.
What the questions look like
Questions appear in five formats. Most present the situation of a hypothetical organization first, and then ask for the most appropriate judgment or procedure for the auditor to adopt in that situation.
| Format | Appearance | Scoring | Solution tips |
|---|---|---|---|
| Single choice | One correct answer out of 5 options | Exact match | First mark the conditions of the scenario (what remains, who is responsible, what changed) before reading the options |
| Multiple choice | 2–3 correct answers out of 5–6 options | Scored as correct only when all correct answers are selected and no incorrect answers are selected | Evaluate true/false independently for each option |
| Statement evaluation | Yes/No for each statement | Both judgment and rationale must be correct for Yes | Even if the conclusion is correct, if the rationale is incorrect, it is No |
| Sequence ordering | 4–5 steps in sequence | Entire sequence must match | Arrange them by asking, "Is the result of the preceding step the input for the next step?" |
| Matching | Connect left items to right options | All pairs must match | Eliminate the most certain pairs first, and remember that there is one unused option |
A significant number of single-choice questions in this exam are designed by reusing the same list of options across multiple scenarios. Because the correct answer among the same five options changes depending on the situation, reading only the options and choosing the "most model answer" will easily lead to errors. For example, a "random sample" is the correct answer if the population is uniform, but it yields to a stratified sample if there are subpopulations with different risks, and yields to full population analysis if the attribute to be verified is determined automatically. Read with the awareness that every single fact in the scenario changes the correct answer.
Incorrect options are generally constructed in four ways. First, they conflate design with operation ("Written procedures exist, so it is operating"). Second, they mistake the quantity of evidence for quality ("There are tens of thousands of logs, so it is sufficient"). Third, they treat statements from the audited department as evidence ("The person in charge replied that it was implemented"). Fourth, the auditor assumes management's responsibilities ("The audit team approves the tool"). The "Common mistakes" section in each chapter organizes these four types chapter by chapter.
Six-week study plan
This schedule is designed on the basis of 40–60 minutes per day, 5 days per week. If you have substantial audit practice experience, you may combine Week 2 and Week 3. At the end of each week, there must be a tangible deliverable in hand.
| Week | Scope | Tasks | Deliverables |
|---|---|---|---|
| Week 1 | Chapter 1 | Read Govern and Map in the NIST AI RMF Core in the original text | Draft AI inventory supplemented by independent sources |
| Week 2 | Latter half of Chapter 1 and hands-on tasks | Read Articles 2, 31, and 33 through 35 of the AI Basic Act and Article 27 of its Enforcement Decree | Plan table establishing risk priorities and audit depth |
| Week 3 | Chapter 2 | Read the design and operating effectiveness evaluation paragraphs of PCAOB AS 2201 and AS 2601 | Human review control test plan, vendor control verification sheet |
| Week 4 | Chapter 3 | Read AS 1105, AS 2315, and the evidence paragraphs of Chapter 8 of the Yellow Book | Sampling design document and log requirement list |
| Week 5 | Chapter 4 | Read the trustworthiness characteristics of the NIST AI RMF and Data and Performance in the GAO Framework | Performance claim verification table |
| Week 6 | Chapter 5, comprehensive review | Read Chapter 9 of the Yellow Book and IIA Principles 14 and 15 (Standards 14.1–15.2), and resolve the Self-checks | One finding along with a follow-up verification plan, Review notes |
Key point
Read the standards in their original text at least once, not just summaries. The difference between "should" and "may", or "generally" and "always", alters an auditor's judgment. This exam frequently tests those distinctions.
Notation and sources
Notation for standards. PCAOB standards are cited by standard number and paragraph number, such as "AS 2315 .41"; the Yellow Book is cited by paragraph number, such as "GAO 8.124"; and IIA standards are cited by standard number, such as "IIA 14.3". Subcategories of the NIST AI RMF are cited by function and number, such as "GOVERN 1.6" and "MEASURE 2.11".
Notation for provisions. Statutory provisions are cited as "Article 34 (1) 4", and enforcement decrees as "Article 27 (2) of the Enforcement Decree". Provisions that do not specify the name of an act are provisions of the AI Basic Act. When citing the Personal Information Protection Act, the name of the act is prefixed.
Strength of verbs. The verbs in standard statements convey the degree of requirement. In the Yellow Book, "must" indicates an unconditional requirement that must be complied with in all applicable cases, and "should" indicates a presumptively mandatory requirement that must be complied with in principle, so that in rare cases of departure, alternative procedures must be performed, and the documented justification for the departure and how the alternative procedures fulfilled the objective of the requirement must be documented (GAO 2.02–2.04). IIA standards divide each standard into "Requirements" (mandatory obligations) and "Considerations for Implementation" (considerations for implementation). In statutes, "shall" and "shall endeavor to" are distinguished in the same manner.
Hypothetical organizations. All companies and institutions in the text are fictitious names. Numbers were also created for explanatory purposes.
Status indication. Content that could not be verified against official documents was not included. Where standards do not speak directly and this book borrows and applies principles, it is explicitly indicated using phrasing such as "borrowing from" or "applying".
Callout boxes. Note (NOTE) provides supplementary context, Tip (TIP) highlights key points to remember, Warning (WARNING) marks frequent mistake points, and Example (EXAMPLE) represents hands-on tasks to perform yourself.