Overview
Understanding the foundations of AI governance is the first of the four domains in the AIGP Body of Knowledge v2.1, in force since 2 February 2026. It has three competencies and 12 performance indicators. Competency I.A asks what AI is, which types exist, what risks and harms it can cause, why its characteristics need a dedicated governance approach, and which responsible AI principles apply. Competency I.B covers organizational expectations: stakeholder roles, cross-functional collaboration, training and awareness, tailoring governance to the organization, and how developers, providers, deployers and users differ. Competency I.C turns this into policy, from life cycle oversight and accountability to updating privacy, security and IP policies and managing third-party AI risk.
IAPP allocates 16 to 20 of the 85 scored items here, about 21% of the exam. The AIGP certification is current, with no newer Body of Knowledge announced as of October 2026.
How it's tested
Domain I questions are usually framed around an organization starting or maturing its AI governance program. You are asked what should come first, who should own a decision, which principle is at stake or which policy needs updating. The expected answer is proportionate, documented and shared across functions rather than owned by one team.
-
Name the risk or principle
A harm such as biased screening or an unexplained denial is described, and you identify the risk category or responsible AI principle involved.
-
Assign the role
A scenario involves a model builder, a vendor and a business unit using the tool, and you decide which party holds which governance duty.
-
Choose the first policy step
A company is adopting AI quickly, and you pick the policy, inventory, committee or training action that should come before anything else.
A mid-sized insurer finds staff pasting claim files into a public chatbot. Banning all AI tools is tempting but drives use underground. Buying an enterprise tool fixes nothing on its own. The best answer is to update the acceptable use and data handling policies for AI, back them with training, and route new tools through an intake review, because that addresses behavior, data and future tools together.
Performance indicators 12
Official performance indicators from the IAPP outline, in outline order.
I.A.1 Know the generally accepted definitions and types of AIAI definitions, machine learning, generative AI and other types.
- What it means
- This indicator covers what counts as an AI system and the main families of AI. Widely used definitions come from the OECD, which describes an AI system as a machine-based system that infers from input how to generate outputs such as predictions, content, recommendations or decisions, and the EU AI Act, which builds on that wording. Types include machine learning (supervised, unsupervised, reinforcement), deep learning and neural networks, natural language processing, computer vision, expert systems and generative AI.
- What to understand
- Definitions matter because they decide whether a law or policy applies at all. A simple rules-based script may fall outside an AI definition, while a model that infers outputs from data falls inside. Know the difference between narrow AI, which performs specific tasks, and general-purpose models adapted to many tasks. Generative AI creates new content; discriminative or predictive models classify or score.
- How it shows up
- A system is described and you decide whether it meets a common AI definition or which type of AI it is.
- Common confusion
- Assuming any automated decision tool is AI; fixed if-then rules without inference may not qualify.
I.A.2 Identify the types of risks and harms posed by AI to individuals, groups, organizations and societyHarms to people, groups, organizations and society.
- What it means
- This indicator asks you to recognize harm at four levels. Individuals can suffer discrimination, privacy loss, physical injury or denial of opportunity. Groups can face systemic bias or exclusion. Organizations face reputational, legal, financial and operational harm. Society faces threats such as misinformation, erosion of democratic processes, labor disruption and environmental cost. IAPP's examples include misalignment with objectives, ethics and bias risk, and complexity and scalability.
- What to understand
- Map each harm to the party who bears it, because governance responses differ. A biased hiring model harms candidates directly and the employer through legal exposure. Deepfakes can harm one person and also public trust. Expect scenarios where one AI failure produces harms at several levels; the best answer usually names the most direct and severe harm first.
- How it shows up
- A failure is described and you identify which level of harm it represents or which harm is most significant.
- Common confusion
- Treating organizational harm such as fines as the main concern when the scenario centers on harm to individuals.
I.A.3 Identify the unique characteristics of AI that require a comprehensive approach to governanceComplexity, opacity, autonomy, speed, scale and data dependency.
- What it means
- This indicator explains why ordinary IT governance is not enough. IAPP lists traits such as complexity, opacity, autonomy, speed and scale, potential for harm or misuse, and data dependency. Models learn behavior from data rather than following written rules, can change over time, can act with limited human input and can repeat a single error across millions of decisions.
- What to understand
- Each trait links to a governance response. Opacity calls for explainability and documentation. Autonomy calls for human oversight. Speed and scale call for monitoring and kill switches. Data dependency calls for data governance and provenance checks. Probabilistic outputs mean the same input can produce different answers, so testing must be statistical rather than one-off.
- How it shows up
- A governance gap is described and you identify which AI characteristic created it or which control addresses it.
- Common confusion
- Thinking opacity only applies to deep learning; vendor secrecy can make any model opaque to its deployer.
I.A.4 Identify and apply the common principles of responsible AIFairness, safety, privacy, transparency, accountability, human-centricity.
- What it means
- This indicator covers the shared principles that appear across the OECD AI Principles, the NIST AI RMF trustworthiness characteristics and company frameworks. IAPP names fairness, safety and reliability, privacy and security, transparency and explainability, accountability, and human-centricity. You must recognize each in a scenario and apply it, not only list it.
- What to understand
- Fairness concerns unjustified differences in outcomes between groups. Safety and reliability mean the system works as intended without causing harm. Transparency is openness that AI is used and how; explainability is the ability to give reasons for a specific output. Accountability means a named person or body answers for outcomes. Human-centricity keeps people's rights and wellbeing at the center. Principles can conflict, such as accuracy against privacy, and governance documents the trade-off.
- How it shows up
- A design choice or failure is described and you identify which principle is affected or how to apply it.
- Common confusion
- Using transparency and explainability as synonyms; disclosure that AI is used does not explain an individual decision.
I.B.1 Define roles and responsibilities for AI governance stakeholdersBoard, executives, committee, owners, risk, legal, privacy.
- What it means
- This indicator covers who does what in an AI governance program. Typical stakeholders are the board and senior leadership, an AI governance committee or council, a chief AI or responsible AI officer, model and use case owners, data science and engineering, legal, privacy, security, compliance, risk management, procurement, HR and internal audit.
- What to understand
- Clear ownership prevents gaps where everyone assumes another team checked the risk. Leadership sets risk appetite and funds the program. A committee sets policy and decides escalated cases. Business owners are accountable for each use case. Second-line risk and compliance functions challenge, while internal audit gives independent assurance. A RACI matrix is the usual way to record this, and accountability should sit with a person, not a tool or vendor.
- How it shows up
- A decision needs an owner and you choose the role that should be accountable or consulted.
- Common confusion
- Making the data science team accountable for business outcomes it does not own.
I.B.2 Establish cross-functional collaboration in the AI governance programDiversity of expertise and perspective across functions.
- What it means
- This indicator recognizes that AI risk crosses legal, technical, ethical and business lines, so no single function can govern it. Collaboration means bringing privacy, security, legal, engineering, product, HR, ethics and affected business units into intake reviews, risk assessments and decisions, with diversity of expertise and perspective.
- What to understand
- Practical mechanisms include a standing committee with a charter, shared intake forms, joint impact assessments and agreed escalation paths. Diverse perspectives catch harms that a homogeneous team misses, such as accessibility problems or cultural bias. Collaboration also avoids duplicate reviews, for example by combining a privacy impact assessment with an AI impact assessment where they overlap.
- How it shows up
- A program is stalling or missing risks and you choose the collaboration step that fixes it.
- Common confusion
- Assuming collaboration means consensus on every decision; the charter still names who decides.
I.B.3 Create and deliver a training and awareness program to all stakeholders on AI terminology, strategy and governanceRole-based AI training and awareness.
- What it means
- This indicator covers building AI literacy across the organization. Training should explain AI terms, the company's AI strategy, approved and prohibited uses, how to request a new tool and how to report concerns. Content should differ by audience: executives, developers, procurement, HR and general staff need different depth.
- What to understand
- Role-based training is more effective than one generic course. Developers need secure and fair design practices; reviewers need assessment methods; general staff need acceptable use rules. The EU AI Act includes an AI literacy provision for providers and deployers, so training also supports compliance. Track completion and refresh content as tools and rules change.
- How it shows up
- A program has policies but staff misuse tools, and you pick the training response that fits the audience.
- Common confusion
- Treating a single annual awareness module as enough for engineers who build models.
I.B.4 Differentiate approaches to AI governance based upon company size, maturity, industry, products and services, objectives and risk toleranceProportionate, risk-based governance tailored to the organization.
- What it means
- This indicator asks you to tailor the program. A ten-person startup using one vendor tool does not need the same structure as a global bank building credit models. Factors include company size, governance maturity, industry and its regulators, the products and services involved, business objectives and risk tolerance.
- What to understand
- Proportionality is the core idea: controls scale with risk and context. Regulated sectors such as health, finance and employment need heavier documentation and oversight. Mature organizations can extend existing privacy or model risk programs, while new programs may start with an inventory and acceptable use policy. Risk tolerance set by leadership decides where the line for approval and escalation sits.
- How it shows up
- Two organizations are compared and you choose the governance approach that fits each one.
- Common confusion
- Picking the most comprehensive framework for every organization regardless of size or risk.
I.B.5 Identify differences among AI developers, providers, deployers and users from a governance perspectiveResponsibilities, opportunities and needs by role.
- What it means
- This indicator separates the parties in the AI value chain. Developers build or train models. Providers place a model or system on the market or into service under their name. Deployers use an AI system under their authority in their own operations. Users interact with the system, sometimes as employees of the deployer and sometimes as members of the public.
- What to understand
- Each role has different responsibilities, opportunities and needs. Developers and providers control design, training data and documentation. Deployers control context of use, human oversight and monitoring in their setting. One company can hold several roles, for example building a model and also using it internally. Under the EU AI Act, a deployer that substantially modifies a high-risk system or puts its name on it can become a provider.
- How it shows up
- A value chain is described and you identify each party's role and resulting duties.
- Common confusion
- Assuming a company that only buys an AI tool has no governance duties because the vendor built it.
I.C.1 Create and implement policies to ensure oversight and accountability across all AI life cycle stagesLife cycle policies: use case review to decommissioning.
- What it means
- This indicator covers the policy set that governs AI from idea to retirement. IAPP's examples include use case assessment, risk management, ethics by design, data acquisition and use, and model and system development and deployment. Policies typically require an AI inventory, intake and risk tiering, approval gates, documentation, monitoring and decommissioning.
- What to understand
- Oversight works when policy creates checkpoints that cannot be skipped. A use case is registered, risk-tiered, assessed and approved before build or purchase, then reviewed again before release and on a schedule afterward. Each gate names an accountable owner. Policies should cover both built and bought AI, and both predictive and generative systems.
- How it shows up
- A company has AI in production without approvals, and you select the policy control that restores oversight.
- Common confusion
- Writing policies only for the development stage and forgetting monitoring and retirement.
I.C.2 Evaluate and update existing policies (e.g., data privacy, security, data governance and intellectual property) for AIAdapting privacy, security, data and IP policies.
- What it means
- This indicator recognizes that most organizations already have policies that AI touches. Privacy policies must address training data, inference about people and automated decisions. Security policies must address prompt injection, model theft and data poisoning. Data governance must cover data quality and lineage for training. IP policies must cover use of copyrighted material and ownership of AI outputs.
- What to understand
- Start with a gap analysis rather than writing everything from scratch. Extend retention, access and classification rules to prompts, outputs and model artifacts. Update acceptable use policies to cover generative AI tools. Align terms so that one policy does not permit what another forbids. Existing programs give AI governance an established home and owner.
- How it shows up
- A new AI use exposes a gap and you identify which existing policy needs updating.
- Common confusion
- Creating a standalone AI policy that ignores and contradicts existing privacy and security policies.
I.C.3 Create, update and implement policies, assessments and contracts to manage third-party riskProcurement, supply chain, HR and acceptable use controls.
- What it means
- This indicator covers AI that arrives from outside: purchased tools, embedded AI features in existing software, model APIs and open-source models. IAPP's examples include procurement, supply chain, human resources and acceptable use. Controls include vendor due diligence questionnaires, AI-specific contract clauses, ongoing vendor monitoring and rules on which tools staff may use.
- What to understand
- Due diligence asks how the vendor trained and tested the model, what data it uses and retains, whether customer data trains the model, and what documentation and audit rights exist. Contracts allocate liability, indemnity, incident notice and compliance duties. Software already in use can gain AI features through updates, so third-party reviews must be repeated, not done once.
- How it shows up
- A vendor tool is proposed and you choose the due diligence or contract step that manages its risk.
- Common confusion
- Assuming a well-known vendor's standard terms already cover the buyer's AI risks.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
AI system (OECD definition)
A machine-based system that infers from input how to generate predictions, content, recommendations or decisions.
- Remember
- Inference from input is the defining trait.
- Tested as
- Scope questions.
- Trap
- Labeling fixed rule scripts as AI.
-
Generative vs. predictive AI
Generative AI creates content; predictive AI classifies, scores or forecasts.
- Remember
- Different risks: hallucination versus misclassification.
- Tested as
- Type questions.
- Trap
- Treating a credit score model as generative.
-
Opacity
Difficulty understanding how a model reached an output.
- Remember
- Drives explainability and documentation controls.
- Tested as
- Characteristic questions.
- Trap
- Thinking only neural networks are opaque.
-
Autonomy
Ability to act with limited human input.
- Remember
- Drives human oversight requirements.
- Tested as
- Characteristic questions.
- Trap
- Ignoring autonomy in agentic tools.
-
Fairness
Avoiding unjustified differences in outcomes between groups.
- Remember
- Needs measurement, not good intentions.
- Tested as
- Principle questions.
- Trap
- Assuming removing protected attributes removes bias.
-
Transparency vs. explainability
Disclosing AI use versus giving reasons for a specific output.
- Remember
- Both are needed for affected people.
- Tested as
- Principle questions.
- Trap
- Using the terms interchangeably.
-
Accountability
A named person or body answers for AI outcomes.
- Remember
- Cannot be delegated to a tool or vendor.
- Tested as
- Role questions.
- Trap
- Blaming the model.
-
AI governance committee
Cross-functional body that sets AI policy and decides escalations.
- Remember
- Needs a charter and decision rights.
- Tested as
- Structure questions.
- Trap
- Making it advisory only with no authority.
-
Developer, provider, deployer, user
Builder, market supplier, operator in context, and person interacting.
- Remember
- One company can hold several roles.
- Tested as
- Role questions.
- Trap
- Thinking buyers carry no duties.
-
Risk tolerance
The level of AI risk leadership will accept.
- Remember
- Sets approval and escalation thresholds.
- Tested as
- Tailoring questions.
- Trap
- Setting it inside the data science team.
-
AI inventory
Register of AI systems in use and in development.
- Remember
- First step to oversight.
- Tested as
- Policy questions.
- Trap
- Listing only internally built models.
-
Third-party AI risk
Risk from vendor tools, APIs, open models and embedded features.
- Remember
- Reassess when vendors add AI features.
- Tested as
- Procurement questions.
- Trap
- Reviewing a vendor once at onboarding only.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Value chain roles
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Developer | Builds or trains the model | Controls design and training data | Assuming it controls how buyers use the tool |
| Provider | Places the system on the market under its name | Carries most product obligations | Forgetting a rebranding deployer can become one |
| Deployer | Uses the system under its authority | Controls context, oversight and monitoring | Thinking buying a tool shifts all duties to the vendor |
| User | Person interacting with the system | Needs notice, training or recourse | Confusing users with deployers |
Responsible AI principles in practice
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Fairness | Equitable outcomes across groups | Measured with bias metrics | Equating it with equal treatment of inputs |
| Transparency | Openness that AI is used and how | System-level disclosure | Treating a notice as an explanation |
| Explainability | Reasons for an individual output | Decision-level | Assuming every model can be fully explained |
| Accountability | Ownership of outcomes | Assigned to a person or body | Assigning it to the AI vendor |
Governance approach by organization
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Small, low-risk adopter | Uses a few vendor tools | Inventory, acceptable use policy, vendor review | Building a full model risk function |
| Regulated enterprise | Builds high-impact models | Committee, tiered assessments, audit | Relying on informal sign-off |
| Mature privacy program | Existing DPIA and policy machinery | Extend it to AI rather than duplicate | Starting a parallel program from scratch |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Ban instead of govern
- What candidates pick
- Prohibit all generative AI use across the company.
- Why it looks right
- It removes the risk immediately.
- Why it's wrong
- Bans push use to unmanaged personal accounts; governance with approved tools, policy and training controls the risk.
Clue to watch for Staff are already using AI and the business wants the benefits.
-
Vendor owns the risk
- What candidates pick
- Rely on the vendor's assurances for a purchased hiring tool.
- Why it looks right
- The vendor built and tested the model.
- Why it's wrong
- The deployer still decides how the tool is used on its candidates and remains exposed under employment and AI laws.
Clue to watch for The company uses the tool on its own customers or staff.
-
Technical team as owner
- What candidates pick
- Make data scientists accountable for a lending model's outcomes.
- Why it looks right
- They understand the model best.
- Why it's wrong
- Accountability belongs with the business owner and governance body; technical teams build and advise.
Clue to watch for The question asks who is accountable.
-
Disclosure equals explanation
- What candidates pick
- Answer a request for reasons by stating that AI was used.
- Why it looks right
- It satisfies a transparency principle.
- Why it's wrong
- Explainability requires the main factors behind the specific outcome, not only notice that AI was involved.
Clue to watch for An individual asks why they were rejected.
-
Maximum framework everywhere
- What candidates pick
- Adopt full ISO/IEC 42001 certification for a five-person firm using one chatbot.
- Why it looks right
- More rigor sounds safer.
- Why it's wrong
- Proportionate governance matches controls to size, maturity and risk; a lighter approach fits here.
Clue to watch for Small organization, low-risk use, limited resources.
-
One-time vendor review
- What candidates pick
- Approve a SaaS vendor at onboarding and stop reviewing.
- Why it looks right
- Due diligence was completed.
- Why it's wrong
- Vendors add AI features through updates, so reviews must repeat when functionality or data use changes.
Clue to watch for An existing tool has added a new AI feature.
Answer framework
Use these steps for Domain I scenarios about building or adjusting a governance program.
- Confirm the system is AI under a common definition and identify its type, since this decides which policies and laws apply.
- Identify who could be harmed, at the level of individuals, groups, the organization and society, and which AI traits drive that harm.
- Name the responsible AI principle at stake, such as fairness, transparency or accountability, and the control that puts it into practice.
- Determine each party's role (developer, provider, deployer, user) and assign an accountable owner, using a committee for escalations.
- Choose a proportionate policy response: update existing policies first, add life cycle gates, cover third parties and train each audience.
Practice question
Want more like this? The AIGP practice test has questions for every domain.
Checklist
0 of 13
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Begin with vocabulary and definitions. Read the OECD AI system definition and the AI Act's version, and write a one-line meaning for each AI type the BoK expects, including generative AI. Then build a harms grid with four rows (individuals, groups, organizations, society) and fill it with two examples each.
Next, learn the principles by applying them. Take a familiar AI failure, such as a biased resume screener, and write which principles it breaks and which control would have prevented it. Do the same for opacity, autonomy and scale.
For competency I.B, sketch a RACI for a fictional company with a board, AI committee, legal, privacy, security, data science and HR. Then rewrite it for a 20-person startup to practice proportionality. Practice placing companies in the developer, provider, deployer and user roles.
Finish with I.C by drafting a one-page AI policy outline: inventory, intake, risk tiers, approval gates, monitoring and retirement, plus a vendor checklist. IAPP's official AIGP training and the BoK itself are the primary references. Then attempt practice questions, especially case-study sets, to check that you choose the proportionate answer.
FAQ
How much of the AIGP exam is Understanding the foundations of AI governance?
IAPP allocates 16 to 20 of the 85 scored items to this domain, roughly 21%. It is the smallest of the four domains in BoK v2.1.
Which Body of Knowledge version is the exam based on?
Version 2.1, effective 2 February 2026. It replaced v2.0.1 and keeps the four-domain structure introduced in 2025.
Do I need technical AI knowledge for this domain?
You need conceptual knowledge: AI types, how models learn from data and why that creates risk. You do not need to code or train models.
What is the difference between a provider and a deployer?
A provider places an AI system on the market or into service under its own name. A deployer uses a system under its authority in its own operations and controls the context of use.
Which responsible AI principles does IAPP list?
Fairness, safety and reliability, privacy and security, transparency and explainability, accountability, and human-centricity.
Should an organization write a brand-new AI policy?
Usually it should first evaluate and update existing privacy, security, data governance and IP policies, then add AI-specific life cycle and third-party controls where gaps remain.
Is the AIGP exam scenario based?
Yes. All questions are multiple choice with one correct answer, and about 30% are linked to case studies describing an organization's AI situation.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 1 knowledge
You have covered AI types, harms, responsible AI principles, governance roles and life cycle policies. Check whether you can choose the proportionate governance answer in scenario questions.