AIGP Study Guide

Certified AI Governance Professional (IAPP)

Based on Body of Knowledge v2.1 — effective 2 February 2026

---

How to use this guide

The exam tests four domains. Each domain breaks into competencies (broad areas) and performance indicators (the specific tasks you're scored on). This guide mirrors that structure, then adds the actual concepts behind each indicator — because the blueprint tells you what is tested but not what the answer is.

Exam style: Questions sit mostly at the remember/understand and apply/analyze levels of Bloom's Taxonomy. Translation: about half are "do you know this term/requirement," and half are "here's a scenario — what should the governance professional do?" Study definitions cold, but practice applying them to situations.

Watch the verbs. Identify / define / list = recall a fact. Understand / explain = describe a concept. Apply / evaluate = use judgment in a scenario. The blueprint's verbs predict the question type.

---

Exam blueprint at a glance

Takeaway: Domains III and IV (the operational "how to govern" domains) are the largest. Domain II (law) is close behind and is the most fact-dense. Domain I is foundational and overlaps with everything — learn it first, but it's the smallest slice.

---

DOMAIN I — Foundations of AI Governance (16–20 Q)

What AI governance is, plus the principles and pillars for building a program. Industry- and size-agnostic best practice.

I.A — Understand what AI is and why it needs governance

Definitions & types of AI to know:

• AI — systems that perform tasks normally requiring human intelligence (inference, prediction, generation) from data.
• Classic / narrow AI — built for one task (classification, recommendation, forecasting).
• Generative AI — produces new content (text, image, code) from learned patterns.
• Agentic AI — systems that plan and take multi-step actions toward a goal with limited human input.
• GPAI (general-purpose AI) — foundation models adaptable to many downstream tasks.

Types of risks and harms AI can pose to individuals, groups, organizations and society:

• Misalignment with intended objectives
• Ethics and bias risk (discriminatory or unfair outcomes)
• Complexity and scalability risk (small errors propagate fast and wide)

Unique characteristics that demand a special governance approach (high-value memory list):

Complexity · Opacity · Autonomy · Speed & scale · Potential for harm/misuse · Data dependency · Probabilistic (not deterministic) outputs

That last one matters most conceptually: traditional software is deterministic (same input → same output), AI is probabilistic — outputs vary and can't be fully predicted, which is why it needs monitoring rather than one-time testing.

The responsible AI principles (know all six — frequently tested):

Fairness · Safety & reliability · Privacy & security · Transparency & explainability · Accountability · Human-centricity

I.B — Establish & communicate organizational expectations

• Define roles and responsibilities for governance stakeholders (who owns what).
• Cross-functional collaboration — legal, security, data science, product, HR, ethics. The point is diversity of expertise and perspective, which reduces blind spots and bias.
• Training & awareness for all stakeholders on AI terminology, strategy and governance — not just technical staff.
• Differentiate governance approaches by company size, maturity, industry, products/services, objectives and risk tolerance. There's no one-size program.

The four actor roles (governance perspective — appears in Domains II and IV too):

One organization often plays multiple roles — the blueprint stresses this explicitly. (The EU AI Act also adds importer and distributor — see Domain II.)

I.C — Establish policies & procedures across the AI life cycle

Create oversight and accountability policies covering every life-cycle stage:

Use-case assessment → risk management → ethics by design → data acquisition & use → model/system development → training & testing → deployment & monitoring → documentation & reporting → incident management

Also:

• Update existing policies for AI — data privacy, security, data governance, intellectual property.
• Manage third-party risk via policies, assessments and contracts — procurement, supply chain, HR, and acceptable-use policies.

---

DOMAIN II — Laws, Standards & Frameworks (19–23 Q)

Existing laws applied to AI, plus AI-specific laws, standards and frameworks. Anchored on the EU AI Act, South Korea's AI Basic Law, and US federal/state AI laws.

II.A — How existing data privacy laws apply to AI

Privacy law (think GDPR-style concepts) maps directly onto AI:

• Transparency, choice, lawful basis, purpose limitation — you still need a legal reason to process data and can't repurpose it freely for training.
• Data minimization & privacy by design — collect only what's needed; bake privacy in from the start.
• Controller obligations applied to AI: privacy/data protection impact assessments, vetting third-party processors, cross-border transfer rules, data subject rights, automated decision-making rules, incident management, breach notification, record keeping.
• Sensitive / special-category data (e.g., biometrics) carries heightened requirements.

Automated decision-making is a hot spot: individuals often have rights to information about, and human review of, decisions made solely by automated means.

II.B — How other existing laws apply to AI

II.C — Main elements of AI-specific laws (e.g., EU AI Act)

The EU AI Act risk classification — memorize the four tiers and examples:

High-risk system requirements (know these as a cluster):

Risk management · data governance · technical documentation · conformity/impact assessments · record keeping · human oversight · transparency & notification · quality management

Plus:

• General-purpose AI (GPAI) models have their own distinct requirements, with extra obligations for models posing systemic risk.
• Enforcement & penalties — tiered fines (the EU AI Act tops out at the greater of €35M or 7% of global annual turnover for prohibited-practice violations).
• Obligations differ by role — providers, deployers, importers and distributors carry different duties.

II.D — Main industry standards & tools

Memory hook for NIST: "G-M-M-M" — Govern wraps the other three. Govern is continuous; Map/Measure/Manage cycle.

---

DOMAIN III — Governing AI Development (21–25 Q)

Responsibilities when designing, building, training, testing and maintaining AI systems. This is the "you're building it" domain.

III.A — Govern designing & building

• Define business context and use case before anything else.
• Perform/review an impact assessment on the system.
• Apply policies, procedures, best practices and ethics to design: purpose, requirements gathering, architecture/model selection, human oversight, data analysis, metrics and thresholds, stakeholder engagement, operational controls.
• Identify and manage risks using tools like a probability/severity harms matrix, a risk mitigation hierarchy, stakeholder mapping, use-case evaluation, benchmarking, and pre-deployment pilots/testing.
• Document the whole design/build process — for compliance and risk management.

III.B — Govern data in training & testing

• Data governance requirements — confirm and document the lawful right to collect/use data; assess quality, quantity, integrity, and fit-for-purpose.
• Data lineage and provenance — establish and document where data came from and how it moved. (Provenance = origin; lineage = the journey.)
• Plan and perform testing across types:

  unit · integration · validation · performance · security · bias · interpretability

• Manage issues and risks during training/testing, and document to validate results and prove compliance.

III.C — Govern release, monitoring & maintenance

• Assess readiness for production — create the model card, satisfy conformity requirements.
• Continuous monitoring + a regular schedule for maintenance, updates and retraining. (AI degrades — it's never "done.")
• Periodic safety/performance assessments: audits, red teaming, threat modeling, security testing.
• Manage and document incidents, issues and risks.
• Diagnose why incidents arise with stakeholders: brittleness, lack of robustness, poor-quality data, insufficient testing, and model or data drift.
• Public disclosures for transparency: technical documentation, instructions for use to deployers, post-market monitoring plans.

Model card = a standardized doc describing a model's intended use, performance, limitations and training-data characteristics. Drift = the world (data drift) or the relationship being modeled (concept/model drift) changes over time, degrading accuracy — the main reason continuous monitoring exists.

---

DOMAIN IV — Governing AI Deployment & Use (21–25 Q)

Selecting a model and deploying/using it responsibly — whether it's your own proprietary model or a third party's.

IV.A — Evaluate factors & risks in the deploy decision

• Understand the use-case context: business objectives, performance requirements, data availability, ethical considerations, workforce readiness.
• Know the model-type tradeoffs:

  classic vs. generative · proprietary vs. open source · small vs. large · language vs. multimodal

• Know the deployment-option tradeoffs:

  cloud vs. on-premise vs. edge · model as-is vs. fine-tuning vs. RAG vs. agentic architectures

RAG (retrieval-augmented generation) = feed the model relevant external/company data at query time to improve accuracy without retraining. Fine-tuning = further-train the model on your data. Edge = runs on local device, not the cloud.

IV.B — Assess the AI system before deploying

• Perform/review an impact assessment on the selected system.
• Evaluate vendor/licensing agreements — identify key terms and risks (liability, data use, IP, indemnification, termination).
• Understand the extra exposure of deploying your own proprietary model — increased obligations and higher potential liability vs. buying third-party.

IV.C — Govern deployment & use

• Apply policies/procedures/ethics to deployment: data governance, risk management, issue management, user training.
• Continuous monitoring + scheduled maintenance, updates, retraining.
• Periodic assessments: audits, red teaming, threat modeling, security testing.
• Document incidents, issues, risks, and post-market monitoring plans.
• Forecast and reduce secondary/unintended uses and downstream harms.
• Establish external communication plans.
• Build a kill switch — policy and controls to deactivate or localize a system when required (regulatory or performance reasons).

---

PART TWO — Framework Deep Dives

The four domains above tell you what the exam covers. This part drills into the specific laws, frameworks and standards named in the Body of Knowledge — the high-yield, fact-dense material that Domain II leans on but that also surfaces in III and IV. Each one uses the same template: what it is → status → structure → must-know → exam traps.

Quick mental model: Laws are binding (EU AI Act, GDPR, Korea's AI Basic Act). Frameworks and standards are voluntary (NIST AI RMF, OECD Principles, ISO). The exam loves to test whether you know which is which.

---

1. The EU AI Act — the single most-tested framework

What it is: Regulation (EU) 2024/1689 — the world's first comprehensive, horizontal AI law. Risk-based: the obligations on a system scale with the risk it poses. Applies extraterritorially (anyone placing AI on the EU market or whose output is used in the EU).

Status (enacted timeline — what the exam tests):

⚠️ Real-world status (good to know, but the exam follows the enacted dates): A "Digital Omnibus" simplification package (proposed Nov 2025; Council/Parliament provisional agreement May 2026) would push standalone high-risk obligations to 2 Dec 2027 and embedded-product high-risk to 2 Aug 2028, and add a prohibition on AI-generated CSAM and non-consensual intimate imagery. It's still in trilogue and not final. The BoK v2.1 (approved Sept 2025) reflects the original timeline.

The risk tiers (the spine of the Act):

• Prohibited (Art. 5): government social scoring; manipulative subliminal techniques; exploiting vulnerabilities; untargeted facial-image scraping; emotion recognition in workplaces/schools; most real-time remote biometric ID in public by law enforcement; certain biometric categorization.
• High-risk (Annex III categories): biometrics · critical infrastructure · education · employment/HR · access to essential services (e.g., credit scoring) · law enforcement · migration/border · administration of justice. Plus AI as a safety component of regulated products (Annex I).
• Limited/transparency (Art. 50): chatbots must reveal they're AI; deepfakes and synthetic media must be labeled.
• Minimal: everything else — no obligations.

High-risk obligations cluster (memorize as a group):

Risk management system · data governance · technical documentation · logging/record-keeping · transparency & instructions for use · human oversight · accuracy, robustness & cybersecurity → then conformity assessment + CE marking + EU database registration + post-market monitoring

GPAI is a separate layer:

• All GPAI models: technical documentation, copyright policy, summary of training data.
• Systemic-risk GPAI (training compute > 10²⁵ FLOPs): add model evaluations, adversarial testing, incident reporting, cybersecurity.

Penalties (tiered — note these exceed GDPR):

• Prohibited practices: up to €35M or 7% of global annual turnover
• Most other obligations (incl. GPAI): up to €15M or 3%
• Supplying incorrect info to authorities: up to €7.5M or 1.5%
• (SMEs/startups pay the lower of the fixed sum or the percentage.)

Exam traps: EU AI Act ≠ GDPR — different scope, but both can apply to the same system. Know provider vs. deployer vs. importer vs. distributor. A high-risk use often triggers a GDPR DPIA and possibly a fundamental rights impact assessment (FRIA) and an AI Act conformity assessment — three different assessments.

---

2. NIST AI Risk Management Framework (AI RMF 1.0)

What it is: A voluntary US framework (NIST, January 2023) for managing AI risk and building trustworthy AI. Not a law — adoptable by any organization, anywhere.

The four core functions (the thing you must know cold):

Each function breaks into categories → subcategories (concrete outcomes).

The 7 trustworthy-AI characteristics: valid & reliable · safe · secure & resilient · accountable & transparent · explainable & interpretable · privacy-enhanced · fair (with harmful bias managed).

Companions: the Playbook (suggested actions), the Generative AI Profile (NIST-AI-600-1, July 2024), plus the Roadmap and crosswalks to other standards.

Exam traps: GOVERN is continuous and cross-cutting — not "step 1 of 4." That's the classic trick. Remember NIST = voluntary framework; EU AI Act = binding law. NIST AI RMF maps neatly onto ISO 42001.

---

3. OECD AI Principles

What it is: The first intergovernmental AI standard (adopted May 2019, updated May 2024 to address generative AI and mis/disinformation). The basis for the G20 AI Principles and the conceptual ancestor of many later frameworks.

The five values-based principles:

1. Inclusive growth, sustainable development & well-being
2. Respect for human rights & democratic values, including fairness & privacy
3. Transparency & explainability
4. Robustness, security & safety
5. Accountability

Plus five recommendations to governments: invest in AI R&D · foster an inclusive digital ecosystem · shape an enabling policy environment · build human capacity / prepare for labor-market transition · international cooperation.

Exam traps: OECD = principles, not enforceable. Often the "where did this originate" answer. Don't confuse the 5 OECD principles with the 6 responsible-AI principles (Domain I) or the 7 NIST trustworthiness characteristics — three different lists.

---

4. ISO/IEC AI standards

Voluntary, consensus-based international standards. The BoK names three explicitly — know what each is for:

Also worth knowing: ISO/IEC 23894 (2023) — AI risk management guidance, aligned with ISO 31000.

Exam traps: The likely question is the mapping — 42001 = management system (certifiable) · 42005 = impact assessment · 22989 = terminology. Only 42001 is something an organization gets certified against.

---

5. GDPR — the privacy backbone (Domain II.A)

Why it's here: Domain II.A ("how existing data privacy laws apply to AI") is essentially GDPR concepts applied to AI. Not AI-specific, but heavily tested.

Core concepts applied to AI:

• A lawful basis is required to process personal data — including for training
• Purpose limitation & data minimization
• Transparency / notice to data subjects
• Data subject rights: access, rectification, erasure, portability, objection
• Article 22 — rights regarding decisions based solely on automated processing with legal/significant effects (information, human review, ability to contest)
• DPIA required for high-risk processing
• Controller vs. processor duties; vetting third-party processors; cross-border transfer rules
• Special-category data (biometrics, health) — stricter conditions
• Breach notification (72 hours) and record-keeping
• Penalties up to €20M or 4% of global turnover

Exam traps: Controller (decides the purpose & means) vs. processor (acts on the controller's instructions) — a favorite. Automated decision-making rights come up often. The same high-risk AI system can require a GDPR DPIA and an EU AI Act conformity assessment.

---

6. South Korea AI Basic Act

What it is: The Framework Act on the Development of Artificial Intelligence and Establishment of a Foundation for Trustworthiness ("AI Basic Act"). The second comprehensive national AI law after the EU — named in the BoK's Domain II intro.

Status: Passed December 2024, effective 22 January 2026, with roughly a one-year grace period before penalties are fully enforced.

Key features:

• Risk-based but lighter and more innovation-friendly than the EU model — pairs obligations with heavy industrial support
• Targets "high-impact AI" in critical sectors (e.g., healthcare, energy, public services): risk assessment, human oversight, documentation, user notification
• Also captures large-scale/"high-performance" AI by a compute threshold
• Transparency & labeling for generative AI — notify users that AI is in use; label AI-generated content
• Extraterritorial — foreign firms over set thresholds must designate a domestic representative in Korea
• Enforcement decrees issued by MSIT; comparatively modest fines

Exam traps: Position it against the EU — Korea is promotion + trust, lighter penalties, with a standout generative-AI labeling requirement. Know it as the "second comprehensive AI law."

---

7. The US AI legal landscape

What it is: No single comprehensive federal AI law. Instead, a patchwork: existing federal/sectoral laws applied to AI + a growing set of state laws + NIST AI RMF as the de facto voluntary federal framework. Executive-branch AI policy shifts between administrations and remains in flux — verify current status before relying on specifics.

Existing laws applied to AI: FTC Act (unfair/deceptive practices) · anti-discrimination law (Title VII, ECOA, Fair Housing Act) · EEOC guidance on AI hiring tools · sector rules (FCRA for credit, HIPAA for health).

State level: Colorado's AI Act — the first comprehensive US state AI law (developer/deployer duties for high-risk AI, focused on algorithmic discrimination). Plus state privacy laws with profiling/automated-decision provisions, Illinois BIPA (biometrics), and NYC Local Law 144 (bias audits for automated employment decision tools).

Exam traps: Know it's a patchwork — no omnibus federal AI law. NIST AI RMF is voluntary, not binding. Federal direction is administration-dependent.

---

Framework comparison — one-glance table

---

Cross-domain quick-reference cheat sheet

The 6 responsible AI principles: Fairness · Safety & reliability · Privacy & security · Transparency & explainability · Accountability · Human-centricity

The 7 unique AI characteristics: Complexity · Opacity · Autonomy · Speed & scale · Potential for harm · Data dependency · Probabilistic outputs

The 4 actor roles: Developer · Provider · Deployer · User (+ Importer, Distributor in EU AI Act)

EU AI Act risk tiers: Prohibited → High → Limited → Minimal

NIST AI RMF functions: Govern · Map · Measure · Manage

OECD = principles · NIST = risk framework · ISO 42001 = management system · ISO 42005 = impact assessment · ISO 22989 = terminology

Testing types: unit · integration · validation · performance · security · bias · interpretability

Incident root causes: brittleness · lack of robustness · poor data quality · insufficient testing · drift

Deployment techniques: as-is · fine-tuning · RAG · agentic

---

Self-test prompts

Cover the answers and check yourself.

Domain I

1. Name all six responsible AI principles.
2. Why does "probabilistic vs. deterministic" output change how you govern AI?
3. List the nine AI life-cycle stages a policy must cover.

Domain II 4. Give one example for each EU AI Act risk tier. 5. What are the four NIST AI RMF core functions, and which one is cross-cutting? 6. Match: 22989 / 42001 / 42005 → terminology / impact assessment / management system. 7. Name three existing (non-AI-specific) law types and how each applies to AI.

Domain III 8. What goes in a model card, and when is it created? 9. Define data lineage vs. provenance. 10. Name four periodic activities used to assess a deployed system's safety.

Domain IV 11. RAG vs. fine-tuning — what's the difference? 12. Why does deploying your own model raise your liability vs. using a vendor's? 13. What is a "deactivate or localize" control and when would you use it?

Framework deep dives 14. Name the four EU AI Act risk tiers and one example each. 15. What compute threshold flags a GPAI model as posing "systemic risk"? 16. Which NIST AI RMF function is cross-cutting, and what do the other three do? 17. List the five OECD AI Principles (values-based). 18. Map ISO 22989 / 42001 / 42005 to terminology / management system / impact assessment. 19. GDPR: controller vs. processor — who decides the purpose and means? 20. Why is the South Korea AI Basic Act significant, and what's its standout GenAI requirement? 21. True or false: the US has a single comprehensive federal AI law. (And NIST AI RMF — binding or voluntary?)

---

Suggested study sequence

1. Lock Domain I cold — the principles, characteristics and roles recur in every other domain. ~2 sessions.
2. Grind Domain II — it's the most memorization-heavy (laws, the EU AI Act tiers, the framework/standard names). Use flashcards. ~3–4 sessions.
3. Domains III & IV together — they share a lot (impact assessments, monitoring, red teaming, documentation). Learn the development lens, then the deployment lens, and note where they overlap vs. differ. ~3–4 sessions.
4. Drill scenarios — for the apply/analyze questions, practice "given this situation, what should the governance professional do?" rather than just reciting facts.
5. Final pass — the cheat sheet above plus the self-test prompts.

---

Note: This guide expands the IAPP AIGP Body of Knowledge v2.1 with explanatory content for study purposes. For anything legal/regulatory (especially the EU AI Act and newer laws like South Korea's AI Basic Law), confirm current specifics against primary sources before relying on them — fines, thresholds and effective dates do shift.