How to use this book

Who this book is for

This book is written for people who plan AI features. Service planners, product managers, Business Leads who select and prioritize AI adoption initiatives, and product leaders who must establish launch criteria for AI features are the primary audience. It does not cover how to train models or write code. Leaving those tasks to developers, this book covers the work of the person who judges what to ask developers to build and whether what has been built may be released.

Planning AI features differs from planning traditional features in two ways. First, the outputs are probabilistic. Even with the same input, slightly different answers are returned, and no matter how well built, a certain percentage is wrong. Therefore, one cannot promise in words that it "works well"; it must be defined numerically and verified through evaluation. Second, failure is part of planning. A button is either clicked or not clicked, but AI fails plausibly. From the outset, you must design what the user sees, who intervenes, and how to stop it when it is wrong.

An understanding at the yellow belt (AI questioning) level and service planning experience are sufficient. Even without planning experience, terminology is explained upon its first appearance so that you can follow along. Upon completing this book, you should be able to produce the following five deliverables yourself: an AI applicability memo, a requirements document containing measurable success criteria, an evaluation plan, a response matrix by failure type, and a phased rollout plan. The hands-on tasks in each chapter guide you to create these deliverables one by one.

How this book is organized

The main text consists of five chapters in the same order as the exam domains. It traces the exact path of a single AI feature from idea to operations.

Each chapter consists of ten sections: Learning objectives, Introduction, Key concepts, Going deeper, At work, Common mistakes, Summary at a glance, Check the official docs, Self-check, and Hands-on tasks. Key concepts comprehensively covers the knowledge and skills within the exam scope; Going deeper explores edge cases and trade-offs where judgment calls diverge; and At work illustrates a fictional organization's journey from problem to decision, including the discarded options.

The appendix contains a glossary (A), Self-check explanations (B), an exam-day guide (C), and a list of official documents cited in this book (D). We recommend reading in chapter order the first time, and starting from "Summary at a glance" and "Common mistakes" in each chapter when reviewing.

The exam at a glance

AI Product Management (APM) is a specialty course. The Specialty exam consists of 35 questions to be completed in 60 minutes. Questions that verify or apply a single principle are worth 2 points, while analytical questions requiring multiple conditions to be considered together are worth 3 points. Candidates pass if their raw score, converted to a 100-point scale, is 75 points or higher, and grades 1 through 9 are also displayed according to the score bracket. The free practice set consists of 10 questions and 25 minutes.

The domain with the largest weight is Evaluation Design. However, in actual questions, domains appear mixed together. It is common for the correct answer to a launch question to stem from success criteria (Chapter 2), or for the correct answer to a failure handling question to stem from assessing AI applicability (Chapter 1). Rather than memorizing chapters in isolation, understand them as a single continuous flow: "Is it necessary? What constitutes success? How do we verify it? What do we do if it is wrong? How do we roll it out?"

What the questions look like

Questions appear in five formats. Most first present a scenario from a fictional organization and ask for the most appropriate judgment or action in that situation.

The exam paper displays all questions converted into a format where a single number (①–⑤) is selected. For multiple select and statement evaluation, you choose one among the combinations in the options box; for sequencing and matching, you choose one among candidate sequences or pairings.

There is something to be particularly careful of in this exam: The same list of options may be used in different scenarios, and the correct answer varies depending on the scenario. For example, "A human approves every output" is the correct answer for a small volume of official correspondence that cannot be revoked once sent, but it is incorrect for high-volume tagging tasks where errors can be corrected immediately. The questions are designed so that if you choose what seems most responsible solely by looking at the choices, you will get it wrong. Read the numbers and conditions in the scenario first, and determine which choice those conditions lead you to select.

Distractors are generally constructed in three ways. First, bringing a practice that is correct in another situation into this situation (requiring 100% human approval for easily reversible tasks, or automated scoring when no criteria exist). Second, choosing an unnecessarily heavy solution (using an LLM for calculation, or an agent for classification). Third, swapping metrics and objectives (arguing that format validation is unnecessary because accuracy is high, or focusing on click counts instead of conversion rates). The "Common mistakes" section in each chapter organizes these three types by chapter.

Six-week study plan

This plan is designed based on 40–60 minutes per day, 5 days per week. If you have extensive planning experience, you may combine Weeks 1 and 2. At the end of each week, there must be a tangible deliverable left in your hands.

Notation and sources

Sources and verification dates. All pages relied upon by this book are listed in each chapter's "Check the official docs" section and Appendix D. The verification date for all citations is October 22, 2026. If the documentation has changed, the documentation takes precedence.

Vendor names and feature names. Product and feature names retain the exact original notation from official documentation (e.g., Message Batches API, structured outputs). Because model names and prices change frequently, they are listed along with their verification dates only when strictly necessary to explain underlying principles.

Statutory notation. The Framework Act on the Development of Artificial Intelligence and Establishment of Foundation for Trustworthy Artificial Intelligence is abbreviated as the AI Basic Act, and the Personal Information Protection Act is abbreviated as PIPA. Articles are cited in the format of "Article 31(1)".

Fictional organizations. All companies and institutions in the text are fictitious names. Any resemblance to real organizations is purely coincidental.

Alert boxes. NOTE provides supplementary context, TIP highlights key takeaways, WARNING flags common failure points, and EXAMPLE denotes hands-on tasks to try yourself.