Skip to main content
CertRampLearning

ISACA · Free practice questions

AAIR Practice Questions

15 original ISACA AAIR practice questions across the three domains of the official exam content outline, spread roughly by weight: six on AI Risk Governance and Framework Integration (37%), three on AI Life Cycle Risk Management (21%) and six on AI Risk Program Management (42%). Each one comes with the answer, a short explanation and the reason every other option is wrong.

Written for the current AAIR exam · Updated November 11, 2026

These are original practice questions written by CertRamp to the official ISACA exam outline. They are not real exam questions and do not come from any exam. Try each question first, then open the answer to see the explanation for every option.

Domain 1

AI Risk Governance and Framework Integration

1A AI Models, Frameworks, Strategies, and Use CasesFoundationDomain: AI Risk Governance and Framework Integration

Question 1. A risk manager is aligning the organisation's AI risk program with the NIST AI Risk Management Framework. Which core function is designed to be cross-cutting, informing and running through the other three functions?

  1. Option A: Map
  2. Option B: Govern
  3. Option C: Measure
  4. Option D: Manage
Show answer and explanation

Correct answer: B. Govern

The AI RMF core has four functions: Govern, Map, Measure and Manage. Govern sets the risk culture, policies and accountability that every other activity depends on, which is why NIST positions it across the whole life cycle rather than as a single step.

  • A (incorrect)Map establishes context and frames risks for a specific AI system. Its outputs feed Measure and Manage, but it is not the cross-cutting function.
  • B (correct)Correct. NIST describes Govern as a cross-cutting function that is infused throughout Map, Measure and Manage, covering culture, policies and accountability structures.
  • C (incorrect)Measure analyses, benchmarks and monitors AI risks using methods and metrics. It draws on Map and informs Manage, but does not span the others.
  • D (incorrect)Manage allocates resources to mapped and measured risks and plans treatment, response and recovery. It is one of the three functions that Govern informs.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 5: AI RMF Core (opens in a new tab)

1B AI Organizational Processes and AlignmentIntermediateDomain: AI Risk Governance and Framework Integration

Question 2. Several business units submit AI use cases for approval. The AI risk committee notices that each team rates its residual risk as acceptable, but the teams use very different reasoning. What should the committee establish FIRST?

  1. Option A: A single approved AI vendor list for all business units
  2. Option B: A requirement that every use case be red-teamed by an external firm before approval
  3. Option C: A central inventory of all AI models in production
  4. Option D: Documented AI risk tolerances that every assessment is measured against
Show answer and explanation

Correct answer: D. Documented AI risk tolerances that every assessment is measured against

Without a common, documented risk tolerance, every team decides for itself what is acceptable, and the committee cannot compare or aggregate AI risk. Setting tolerances aligns AI decisions with the enterprise risk appetite and makes approvals consistent.

  • A (incorrect)A vendor list may help procurement, but it does not tell teams what level of risk the organisation is willing to accept.
  • B (incorrect)Testing produces risk information, but without a defined tolerance the results still cannot be judged consistently.
  • C (incorrect)An inventory is an important governance control, but it records what exists rather than setting the threshold for acceptance.
  • D (correct)Correct. Acceptability is only meaningful against an agreed tolerance. The NIST AI RMF (MAP 1.5) calls for organisational risk tolerances to be determined and documented.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 5: AI RMF Core (opens in a new tab)

1C AI Ownership, Oversight, and AccountabilityIntermediateDomain: AI Risk Governance and Framework Integration

Question 3. The data science team built a model that pre-screens loan applications for the retail lending department, which now uses it daily. Who should be accountable for the risk of this AI use case?

  1. Option A: The data science team, because it designed and trained the model
  2. Option B: The chief information security officer, because the model runs on corporate infrastructure
  3. Option C: The internal audit function, because it provides independent assurance
  4. Option D: The head of retail lending, as owner of the business process that relies on the model
Show answer and explanation

Correct answer: D. The head of retail lending, as owner of the business process that relies on the model

AI risk ownership follows the business process and its outcomes, not the technology. Clear accountability means the process owner accepts or treats the risk, while developers, the second line and internal audit each support oversight from their own position.

  • A (incorrect)Developers are responsible for building and documenting the model, but they do not own the business decision or the risk it creates.
  • B (incorrect)The CISO owns security controls, not the business risk of how a lending decision tool is used.
  • C (incorrect)Internal audit must stay independent; owning the risk would impair its ability to assure it.
  • D (correct)Correct. The risk owner is the person accountable for the business process and its outcomes. Technical teams, security and audit support that owner in their own roles.

Reference: ISACA: AAIR Exam Content Outline (opens in a new tab)

1D AI Policies, Procedures, and Organizational TrainingFoundationDomain: AI Risk Governance and Framework Integration

Question 4. A review finds that employees in several departments paste customer data into free public generative AI chatbots to draft emails. There is no AI policy. What is the BEST first step for the AI risk practitioner to recommend?

  1. Option A: Block all AI websites at the web proxy for every user and consider the issue closed
  2. Option B: Issue an AI acceptable use policy, backed by staff training
  3. Option C: Deploy data loss prevention rules for AI sites as the only control
  4. Option D: Start disciplinary proceedings against the employees involved
Show answer and explanation

Correct answer: B. Issue an AI acceptable use policy, backed by staff training

Unmanaged use of public AI tools is a governance gap first. A policy that defines sanctioned tools and data classes, combined with awareness training, gives a basis for technical controls and for accountability. The NIST AI RMF Govern function covers both policies and personnel training.

  • A (incorrect)Blocking without a policy or approved alternative tends to push use to personal devices and does not set expectations for future tools.
  • B (correct)Correct. A policy sets the rules for sanctioned use and data handling, and training makes staff aware of them. Technical controls can then enforce the policy.
  • C (incorrect)DLP is a useful enforcing control, but on its own it lacks the policy basis that defines what is allowed and why.
  • D (incorrect)Without a policy in place, staff had no defined rule to break; this treats a governance gap as individual misconduct.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 5: AI RMF Core (opens in a new tab)

1E AI Regulatory Compliance and Legal ConsiderationsIntermediateDomain: AI Risk Governance and Framework Integration

Question 5. An EU-based company plans to use an AI tool that sorts and ranks CVs for job openings. Under the risk-based approach of the EU AI Act, how is this use case MOST likely classified?

  1. Option A: High risk, with duties such as risk management, logging and human oversight
  2. Option B: Unacceptable risk, so the tool is prohibited
  3. Option C: Transparency risk only, so the company just needs to disclose to applicants that AI is used
  4. Option D: Minimal risk, with no specific obligations
Show answer and explanation

Correct answer: A. High risk, with duties such as risk management, logging and human oversight

The EU AI Act sorts AI systems into unacceptable, high, transparency and minimal risk. Employment uses such as CV sorting affect people's fundamental rights and fall into the high-risk tier, so the risk practitioner must plan for the related obligations before deployment.

  • A (correct)Correct. The European Commission lists CV-sorting software for recruitment as an example of a high-risk use case, which carries requirements such as risk mitigation, high-quality datasets, logging, documentation and human oversight.
  • B (incorrect)Prohibited practices include social scoring and emotion recognition in the workplace. CV sorting is regulated as high risk, not banned.
  • C (incorrect)Transparency obligations apply to systems such as chatbots and generated content. Employment use cases carry the heavier high-risk obligations.
  • D (incorrect)Minimal-risk examples are spam filters and AI in video games. A tool that affects access to employment is treated as high risk.

Reference: European Commission: AI Act (regulatory framework for AI) (opens in a new tab)

1F AI Trustworthiness, Ethical and Societal Implications (e.g., ESG)AdvancedDomain: AI Risk Governance and Framework Integration

Question 6. In the NIST AI Risk Management Framework, which characteristic of trustworthy AI is described as a necessary condition and the base for the other characteristics?

  1. Option A: Accountable and transparent
  2. Option B: Secure and resilient
  3. Option C: Valid and reliable
  4. Option D: Fair with harmful bias managed
Show answer and explanation

Correct answer: C. Valid and reliable

NIST lists seven characteristics: valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair with harmful bias managed. A system that does not perform as intended cannot be trustworthy, so validity and reliability come first, and trade-offs between the others must be balanced in context.

  • A (incorrect)NIST shows accountable and transparent as relating to all other characteristics, but it is not the base condition.
  • B (incorrect)Security and resilience are one of the characteristics built on the base; a secure system can still be untrustworthy if it is not valid.
  • C (correct)Correct. NIST names valid and reliable as a necessary condition of trustworthiness and shows it as the base for the other characteristics.
  • D (incorrect)Fairness is a key characteristic, but NIST does not position it as the foundation for the others.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 3: AI Risks and Trustworthiness (opens in a new tab)

Domain 2

AI Life Cycle Risk Management

2B AI Model Training, Testing, and ValidationIntermediateDomain: AI Life Cycle Risk Management

Question 7. Before go-live, a model team reports 98% accuracy for a new claims classification model. During review, the risk practitioner learns the accuracy was measured on the same records used for training. What is the PRIMARY concern?

  1. Option A: Possible overfitting; it must be evaluated on separate, representative data
  2. Option B: Accuracy is an unacceptable metric for any classification model
  3. Option C: The model should be retrained with more epochs to reach 100%
  4. Option D: The training data must be deleted before deployment to protect the privacy of claimants
Show answer and explanation

Correct answer: A. Possible overfitting; it must be evaluated on separate, representative data

A model evaluated on its own training data looks better than it will perform on new cases. The NIST AI RMF Measure function expects systems to be tested before deployment under conditions similar to their intended use, and the evidence should come from data the model has not seen.

  • A (correct)Correct. Scoring a model on its own training data hides overfitting. Validation needs held-out data that reflects the intended deployment setting.
  • B (incorrect)Accuracy can be misleading on imbalanced data, but it is not banned; the core problem here is the test data, not the metric.
  • C (incorrect)More training on the same data would increase the risk of overfitting, not prove that the model generalises.
  • D (incorrect)Data retention is a separate question; deleting training data would also remove the ability to retest and audit the model.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 5: AI RMF Core (opens in a new tab)

2C AI Implementation, Maintenance, and DecommissioningIntermediateDomain: AI Life Cycle Risk Management

Question 8. A customer churn model is being retired and replaced by a vendor service. Which action is MOST important to include in the decommissioning plan?

  1. Option A: Archiving the model's source code in the repository for future reference
  2. Option B: Revoking access and integrations, handling model artefacts and data per retention rules, and updating the AI inventory
  3. Option C: Announcing the retirement date to the model's users
  4. Option D: Running a final accuracy test on the retired model
Show answer and explanation

Correct answer: B. Revoking access and integrations, handling model artefacts and data per retention rules, and updating the AI inventory

Retired AI systems can leave behind open endpoints, service accounts, stored training data and stale inventory records. The NIST AI RMF (GOVERN 1.7) calls for processes to decommission and phase out AI systems safely, in a way that does not increase risk or reduce trustworthiness.

  • A (incorrect)Keeping code may be useful, but it does nothing about the live risks of the retired system.
  • B (correct)Correct. Safe decommissioning removes residual access and data exposure and keeps the inventory accurate, so retiring the system does not increase risk.
  • C (incorrect)Communication is helpful, but it does not remove access, data or downstream dependencies.
  • D (incorrect)Testing a model that is being switched off does not address the risks that decommissioning creates.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 5: AI RMF Core (opens in a new tab)

2D AI Data and Asset ManagementAdvancedDomain: AI Life Cycle Risk Management

Question 9. A team fine-tunes a language model on data aggregated from internal documents, a purchased dataset and public web sources. Which control BEST reduces the risk that poisoned data goes undetected?

  1. Option A: Increasing the size of the dataset so that any poisoned records are diluted by clean ones
  2. Option B: Encrypting the training dataset at rest
  3. Option C: Tracking data provenance and versions (for example ML-BOM, DVC) and vetting suppliers
  4. Option D: Restricting the model to answer only in a fixed output format
Show answer and explanation

Correct answer: C. Tracking data provenance and versions (for example ML-BOM, DVC) and vetting suppliers

Data and model poisoning is an integrity attack that is hard to detect once a model is trained, especially from external or unverified sources. Knowing where every dataset came from and how it changed, plus vetting suppliers, makes manipulation detectable and traceable.

  • A (incorrect)Targeted poisoning and backdoors can work with a small number of records, so more data does not reliably neutralise them.
  • B (incorrect)Encryption protects confidentiality in storage, but it does not reveal whether the data was manipulated before it was stored.
  • C (correct)Correct. OWASP recommends tracking data origins and transformations (for example with CycloneDX or an ML-BOM), using version control such as DVC to detect manipulation, and vetting data vendors.
  • D (incorrect)Output format checks address other risks, such as improper output handling, but not the integrity of the training data.

Reference: OWASP Top 10 for LLM Applications 2025: LLM04:2025 Data and Model Poisoning (opens in a new tab)

Domain 3

AI Risk Program Management

3A AI Risk Scenario Identification and Assessment (e.g., threats, vulnerabilities, and attacks)IntermediateDomain: AI Risk Program Management

Question 10. An LLM assistant reads incoming supplier emails and drafts payment instructions for the finance team. While building risk scenarios, the practitioner notes that an attacker could send an email containing hidden instructions that the assistant follows. Which threat does this scenario describe?

  1. Option A: Direct prompt injection
  2. Option B: Model inversion
  3. Option C: Data poisoning
  4. Option D: Indirect prompt injection
Show answer and explanation

Correct answer: D. Indirect prompt injection

Indirect prompt injection is a key scenario wherever an AI system reads untrusted content and can trigger actions. OWASP notes it cannot be fully prevented, so the assessment should drive impact-limiting controls such as least privilege, segregation of external content and human approval for high-risk actions.

  • A (incorrect)Direct injection comes from the user's own input to the model. Here the instructions arrive in external content the model processes.
  • B (incorrect)Model inversion aims to reconstruct training data from a model; it is not about instructions hidden in input content.
  • C (incorrect)Poisoning manipulates training, fine-tuning or embedding data. The email is processed at run time, not used to train the model.
  • D (correct)Correct. OWASP defines indirect prompt injection as instructions hidden in external content, such as files or emails, that the model processes.

Reference: OWASP Top 10 for LLM Applications 2025: LLM01:2025 Prompt Injection (opens in a new tab)

3B AI Risk Treatment StrategiesFoundationDomain: AI Risk Program Management

Question 11. After assessment, a retailer concludes that a planned in-store facial recognition feature exceeds its risk tolerance and that no available control brings the risk down far enough. Management cancels the feature. Which risk response is this?

  1. Option A: Avoidance
  2. Option B: Mitigation
  3. Option C: Transfer
  4. Option D: Acceptance
Show answer and explanation

Correct answer: A. Avoidance

The four classic responses are mitigate, transfer, avoid and accept. When a risk exceeds tolerance and cannot be reduced, avoiding the activity is a legitimate and often the most responsible choice.

  • A (correct)Correct. Ending or not starting the activity that creates the risk is avoidance, one of the response options the NIST AI RMF lists alongside mitigation, transfer and acceptance.
  • B (incorrect)Mitigation reduces likelihood or impact with controls while the activity continues. Here the activity was stopped.
  • C (incorrect)Transfer shifts the financial impact to another party, for example through insurance or contract. Nothing was shifted here.
  • D (incorrect)Acceptance means proceeding with the risk as it is. The retailer chose not to proceed.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 5: AI RMF Core (opens in a new tab)

3C AI Controls Management (e.g., Evaluation, Selection, Validation)AdvancedDomain: AI Risk Program Management

Question 12. A business unit bought a guardrail product that the vendor says blocks prompt injection and the release of personal data. Before the risk register shows the risk as reduced, what should the risk practitioner require?

  1. Option A: The vendor's marketing white paper describing the detection approach
  2. Option B: Confirmation that the product is widely used by other companies in the sector
  3. Option C: Independent testing of the control against defined attack cases in the target environment, repeated after changes
  4. Option D: A signed statement from the business unit that the control is in place
Show answer and explanation

Correct answer: C. Independent testing of the control against defined attack cases in the target environment, repeated after changes

Selecting a control is only half of controls management; its effectiveness must be validated before risk is recorded as reduced. The NIST AI RMF (MEASURE 2.7) calls for AI system security and resilience to be evaluated and documented, and OWASP recommends regular adversarial testing for prompt injection defences.

  • A (incorrect)A description of design intent is not evidence that the control works in your environment.
  • B (incorrect)Popularity says nothing about whether the control is effective for this system and its threats.
  • C (correct)Correct. Control validation needs evidence of operating effectiveness. Adversarial testing in the real configuration shows what the control actually blocks.
  • D (incorrect)Attestation that a control exists does not show that it operates effectively against the risk.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 5: AI RMF Core (opens in a new tab)

3D AI Risk Metrics, Monitoring, and ReportingIntermediateDomain: AI Risk Program Management

Question 13. Which metric is the BEST key risk indicator for the risk that a deployed fraud detection model becomes less effective over time?

  1. Option A: The number of models in the organisation's AI inventory
  2. Option B: The trend of detection rate and false positives against confirmed outcomes, with thresholds that trigger review
  3. Option C: The number of hours the data science team spent training the model
  4. Option D: The percentage of the AI budget spent this year
Show answer and explanation

Correct answer: B. The trend of detection rate and false positives against confirmed outcomes, with thresholds that trigger review

Model effectiveness degrades as data and fraud patterns drift. The NIST AI RMF (MEASURE 2.4) calls for the functionality and behaviour of the AI system to be monitored in production, and thresholds turn that monitoring into a reportable risk indicator.

  • A (incorrect)Inventory size describes exposure at portfolio level, not the effectiveness of this model.
  • B (correct)Correct. A KRI should be forward-looking and tied to the risk. Tracking production performance against ground truth, with agreed thresholds, signals degradation early.
  • C (incorrect)Training effort is an input measure and says nothing about current performance in production.
  • D (incorrect)Budget consumption is a financial metric unrelated to the model's effectiveness.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 5: AI RMF Core (opens in a new tab)

3E AI Supply Chain Risk Management (e.g., third party resources)AdvancedDomain: AI Risk Program Management

Question 14. A company uses a third-party generative AI service in a customer-facing process. The contract allows the provider to change the underlying model at any time without notice. Which contract change BEST addresses this risk?

  1. Option A: Advance notice of material model changes, a test window and an exit right
  2. Option B: A discount on the subscription fee if the service is unavailable
  3. Option C: A clause stating that the provider's model will never produce inaccurate or harmful output
  4. Option D: Moving all prompts to a private network connection
Show answer and explanation

Correct answer: A. Advance notice of material model changes, a test window and an exit right

A third-party model is part of the AI supply chain, and an unannounced change can invalidate earlier testing. The NIST AI RMF (GOVERN 6) calls for policies and procedures for third-party AI risks and for contingency processes for failures in high-risk third-party systems.

  • A (correct)Correct. Notice and a test window let the company re-validate the service, and an exit right gives a remedy if the new model is not acceptable.
  • B (incorrect)Service credits address availability, not unannounced changes in model behaviour.
  • C (incorrect)No provider can guarantee this, and such a clause would not give the company practical control over changes.
  • D (incorrect)Private connectivity protects data in transit, but the provider could still change the model without notice.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 5: AI RMF Core (opens in a new tab)

3F AI Incident Response, BIA, Business Continuity, and Disaster RecoveryIntermediateDomain: AI Risk Program Management

Question 15. A customer service chatbot suddenly starts giving customers incorrect refund commitments. The incident team cannot quickly find who may switch the chatbot off or what the fallback is. Which control was MISSING?

  1. Option A: A larger model with better instruction following
  2. Option B: Monthly accuracy reporting to the steering committee
  3. Option C: A defined way to deactivate the system, with an assigned owner and a manual fallback
  4. Option D: A disclaimer on the website stating that chatbot answers are not legally binding commitments
Show answer and explanation

Correct answer: C. A defined way to deactivate the system, with an assigned owner and a manual fallback

AI incident response needs a pre-agreed way to take a system out of service and keep the business process running. Assigning that authority and planning the fallback in advance, as part of business continuity planning, keeps the impact of an AI failure small.

  • A (incorrect)A different model might reduce errors, but it does not give the team a way to stop the harm now.
  • B (incorrect)Periodic reporting is useful oversight, but it does not provide an immediate response mechanism.
  • C (correct)Correct. The NIST AI RMF (MANAGE 2.4) calls for mechanisms, with assigned responsibilities, to supersede, disengage or deactivate AI systems whose performance is inconsistent with intended use.
  • D (incorrect)A disclaimer may limit legal exposure, but it does not stop the system or restore the service safely.

Reference: NIST AI RMF 1.0 (NIST AI 100-1), Section 5: AI RMF Core (opens in a new tab)

Exam information last verified: November 4, 2026

Exam details can change. Always confirm them with the official provider before you book.

Want more practice?

15 questions are a sample. The exam is 90 questions items.

The full CertRamp AAIR practice exams have 6 tests that grow with you, from a diagnostic start to an exam-level Test 4 and a challenge set beyond the exam, with an explanation for every option.

Who writes the exams

Written by a practitioner, not a question mill

Joshua Ravnjak, founder of CertRamp Learning

Joshua Ravnjak

Senior Consultant, digital transformation and AI

Founder, CertRamp Learning

921 learners and 61 reviews on Udemy (as of November 3, 2026)

I have more than 12 years of professional experience and work as a senior consultant in an IT management consultancy, focusing on digital transformation, AI and requirements engineering in the insurance industry. I have earned many certifications myself, so I know what exam preparation feels like. That is why every CertRamp practice exam follows the same progression: an easy start to see where you stand, then real exam level, then beyond it. You walk into the exam with a buffer and a good feeling.

Selected certifications

  • IREB® CPRE Foundation Level
  • ITIL® 4 Foundation
  • Professional Scrum Master I (PSM I)
  • Professional Scrum Product Owner I (PSPO I)
  • Certified SAFe® 6 Scrum Master
  • Certified AI-Empowered SAFe® 6 Product Owner/Product Manager

Plus further certifications from Microsoft and Anthropic (Claude).

ITIL® is a registered trademark of PeopleCert. IREB® is a registered trademark of IREB e.V. SAFe® is a registered trademark of Scaled Agile, Inc. PSM and PSPO are trademarks of Scrum.org.

Also in print

Cover: CPRE Foundation Level Prüfungsvorbereitung 2026 by Joshua Ravnjak

CPRE Foundation Level Prüfungsvorbereitung 2026

6 Übungsprüfungen mit 270 Fragen und ausführlichen Erklärungen

Paperback, German, October 2026

My CPRE Foundation Level exam-prep book (in German) follows the same progressive approach.

View on Amazon (opens in a new tab)