AAIA
Can you pass this 10-question AAIA quiz?
Ten questions taken from our AAIA Practice Questions, at least two from each of the three domains, with four options each, as on the real exam. Tap an answer to see why, with the source. The questions are scenarios, so allow a quarter of an hour.
Can you get 10 out of 10?
1 of 10A payments company's fraud owner reports that a new model is 99.9% accurate on held-back test data. In that data, about 1 transaction in 1,000 is confirmed fraud. Which measure would BEST show you whether the model catches fraud?
Show answer
Answer: A. When fraud is 1 case in 1,000, a model that calls every transaction genuine is right 999 times in 1,000 and catches no fraud at all. A 99.9% accuracy figure can therefore hide a model that misses everything. Recall is the share of real positives the model caught, which is the question asked. Source: ISACA, AAIA Examination Content Outline (June 2025), Domain 1, subdomain 1A: AI Models, Considerations, and Requirements
2 of 10HR describes its AI hiring screen as decision support that only helps recruiters, and rates it low risk. In your walkthrough, you find that recruiters open only the applications the screen places in its top band. The rest are never read. Which conclusion BEST describes the screen's role?
Show answer
Answer: C. A system that filters options so a person never sees some of them has made those decisions. The applicants outside the top band are affected parties: people subject to the system's output who may never know it exists. The label "decision support" gives no assurance by itself, which rules out option A. Source: ISACA, AAIA Examination Content Outline (June 2025): Domain 1, AI Governance and Risk, and supporting task 5 (AI decision-making systems)
3 of 10A supplier's ISO/IEC 42001 certificate covers the claims-handling product you are buying. Its statement of applicability excludes the reference controls on third-party data, and the product is trained on data bought from brokers. What should you do?
Show answer
Answer: A. A statement of applicability lists each reference control, whether it is applied and the reason for any exclusion, and the exclusions tell you what the certificate does not cover. Here the excluded controls are the ones that would give assurance over bought data, so your reliance stops there and you ask for the reason and test that part yourself. Source: ISACA, AAIA Examination Content Outline (June 2025), Domain 1, subdomain 1E: Leading Practices, Ethics, Regulations, and Standards for AI
4 of 10A vendor's monthly update to your HR platform switched on a feature that ranks job applicants. The feature never went through your organization's own design, development or deployment, and it is ranking people today. Which control would BEST have caught it?
Show answer
Answer: B. The control that would have caught this sits at intake: a rule that new AI features in existing products pass the same intake as new AI systems, backed by a contract term requiring notice before AI features are switched on. A feature a vendor switches on without that intake is shadow AI, and an inventory built only from project records cannot see it. Source: ISACA, AAIA Examination Content Outline (June 2025): Domain 2, subdomain 2B, AI Solution Development Methodologies and Lifecycle, and supporting task 16 (vendors and supply chain)
5 of 10The owner of a benefits-information assistant, which answers from retrieved policy documents, cites a 92% thumbs-up rate from users as evidence that its answers are accurate. Which evidence would BEST test answer accuracy?
Show answer
Answer: C. Satisfaction is not correctness, because a user told confidently that they qualify for a benefit will be satisfied whether or not they do. A generative system has no column of right answers, so supervision rests on a review sample of live outputs, drawn on a stated basis and scored against a written rubric by qualified reviewers, with groundedness checked against the retrieved documents. Source: ISACA, AAIA Examination Content Outline (June 2025), Domain 2, subdomain 2D: Supervision of AI Solutions
6 of 10You open the risk register for a new customer assistant built on a language model. Its one security entry reads "cyber attack," rated medium, and the controls column says "covered by the information security program." Which is your BEST conclusion?
Show answer
Answer: B. The entry could have been written for a payroll server. It names no prompt injection, poisoning or other threat that works only on AI, so the finding is about the risk assessment before it is about any control, and B states it. Source: ISACA, AAIA Examination Content Outline (June 2025), Domain 2, subdomain 2F: Threats and Vulnerabilities specific to AI
7 of 10A purchasing agent built on a language model can raise orders of any value, and it raises large orders with no person approving them. Using OWASP's three root causes of excessive agency, which statement is correct?
Show answer
Answer: D. OWASP traces excessive agency to functionality, permissions and autonomy, each with its own fix. A tool the agent does need, running with broader rights than the task requires, is excessive permissions, so the unlimited value belongs there. High-impact actions taken without a person approving them are excessive autonomy. Source: ISACA, AAIA Examination Content Outline (June 2025), Domain 2, subdomain 2F: Threats and Vulnerabilities specific to AI; OWASP Top 10 for LLM Applications
8 of 10During fieldwork on a benefits eligibility model, you find that it has been wrongly refusing a group of applicants since a release last week, and the refusals are continuing. What should you do FIRST?
Show answer
Answer: A. The harm is live, so the first job is to get the incident to the people with authority to contain it, which A does. B makes a management decision, and an auditor who takes it has joined the control they are there to judge. Source: ISACA, AAIA Examination Content Outline (June 2025), Domain 2, subdomain 2G: Incident Response Management specific to AI
9 of 10Your team ran an anomaly-detection tool over 120,000 auto-approved refunds and tested the 50 cases it flagged. Of those, 12 lacked the required second approval. A colleague's draft says: "About 24% of auto-approved refunds lack second approval." Which response BEST corrects the draft?
Show answer
Answer: A. The tool picked the 50 cases for their characteristics, so this is a targeted selection. The other refunds had no chance of selection, so the result speaks only for the items tested and gives no basis to project a rate. Report the 12 exceptions for what they are, and if the objective needs a conclusion about all refunds, draw a random sample as well. Source: ISACA, AAIA Examination Content Outline (June 2025), Domain 3, subdomain 3B: Audit Testing and Sampling Methodologies
10 of 10Your draft finding says: "The claims model treats rural customers unfairly." Your manager returns it, asking what it is measured against. Which revision BEST states the condition?
Show answer
Answer: C. The condition is what you found, stated as fact with its evidence: what happened, how many times, out of what population, over what period and for which model version. The revision that names the version, the period, the gap, the approved tolerance and the population is the one a reader can check against the workpapers. Source: ISACA, AAIA Examination Content Outline (June 2025), Domain 3, subdomain 3E: AI Audit Outputs and Reports
AAIA flashcards for Anki
front,back "A payments company's fraud owner reports that a new model is 99.9% accurate on held-back test data. In that data, about 1 transaction in 1,000 is confirmed fraud. Which measure would BEST show you whether the model catches fraud? | A) Recall: the share of confirmed fraud cases that the model flagged. / B) Precision: the share of flagged cases that turned out to be confirmed fraud. / C) Overall accuracy, recomputed by you on the same held-back test data. / D) The share of transactions the model scored inside the agreed response time.","A) Recall: the share of confirmed fraud cases that the model flagged.. When fraud is 1 case in 1,000, a model that calls every transaction genuine is right 999 times in 1,000 and catches no fraud at all. A 99.9% accuracy figure can therefore hide a model that misses everything. Recall is the share of real positives the model caught, which is the question asked. Source: Credential Press, https://credentialpress.com/guides/aaia-quiz" "HR describes its AI hiring screen as decision support that only helps recruiters, and rates it low risk. In your walkthrough, you find that recruiters open only the applications the screen places in its top band. The rest are never read. Which conclusion BEST describes the screen's role? | A) It is decision support, because a recruiter still makes every hiring decision. / B) It is low risk, because the organization faces little exposure from it. / C) It decides about the applicants outside the top band, who are affected parties no person ever sees. / D) It is a human-in-the-loop design, because recruiters approve each shortlist.","C) It decides about the applicants outside the top band, who are affected parties no person ever sees.. A system that filters options so a person never sees some of them has made those decisions. The applicants outside the top band are affected parties: people subject to the system's output who may never know it exists. The label ""decision support"" gives no assurance by itself, which rules out option A. Source: Credential Press, https://credentialpress.com/guides/aaia-quiz" "A supplier's ISO/IEC 42001 certificate covers the claims-handling product you are buying. Its statement of applicability excludes the reference controls on third-party data, and the product is trained on data bought from brokers. What should you do? | A) Treat the exclusion as a limit on what the certificate shows, ask for the reason, and test the data supply yourself. / B) Rely on the certificate in full, because its scope covers the product and the standard lets a supplier exclude controls. / C) Reject the certificate, because a management system that excludes any reference control cannot be certified. / D) Ask the certification body to confirm that the supplier's models are accurate and fair.","A) Treat the exclusion as a limit on what the certificate shows, ask for the reason, and test the data supply yourself.. A statement of applicability lists each reference control, whether it is applied and the reason for any exclusion, and the exclusions tell you what the certificate does not cover. Here the excluded controls are the ones that would give assurance over bought data, so your reliance stops there and you ask for the reason and test that part yourself. Source: Credential Press, https://credentialpress.com/guides/aaia-quiz" "A vendor's monthly update to your HR platform switched on a feature that ranks job applicants. The feature never went through your organization's own design, development or deployment, and it is ranking people today. Which control would BEST have caught it? | A) An annual review of the vendor's independent assurance report on its development practices. / B) An intake rule that new AI features in existing products pass the same intake as new systems. / C) A post-implementation review of the HR platform at a set milestone after go-live. / D) Monitoring of the ranking feature's outputs by group, from the day it is found.","B) An intake rule that new AI features in existing products pass the same intake as new systems.. The control that would have caught this sits at intake: a rule that new AI features in existing products pass the same intake as new AI systems, backed by a contract term requiring notice before AI features are switched on. A feature a vendor switches on without that intake is shadow AI, and an inventory built only from project records cannot see it. Source: Credential Press, https://credentialpress.com/guides/aaia-quiz" "The owner of a benefits-information assistant, which answers from retrieved policy documents, cites a 92% thumbs-up rate from users as evidence that its answers are accurate. Which evidence would BEST test answer accuracy? | A) The thumbs-up rate broken down by topic and by month. / B) The provider's published accuracy benchmark for the underlying model. / C) Live answers sampled at random and on entitlement topics, scored against a rubric by reviewers who know the rules, with groundedness checked. / D) The count of conversations that the assistant escalated to a human agent.","C) Live answers sampled at random and on entitlement topics, scored against a rubric by reviewers who know the rules, with groundedness checked.. Satisfaction is not correctness, because a user told confidently that they qualify for a benefit will be satisfied whether or not they do. A generative system has no column of right answers, so supervision rests on a review sample of live outputs, drawn on a stated basis and scored against a written rubric by qualified reviewers, with groundedness checked against the retrieved documents. Source: Credential Press, https://credentialpress.com/guides/aaia-quiz" "You open the risk register for a new customer assistant built on a language model. Its one security entry reads ""cyber attack,"" rated medium, and the controls column says ""covered by the information security program."" Which is your BEST conclusion? | A) The medium rating understates the exposure of a public-facing system and should be raised before go-live. / B) The entry names no AI-specific threat, so nothing shows that the controls answer attacks on a language model. / C) The information security program should be tested next, to confirm that it covers the assistant's servers. / D) The entry is adequate, because the assistant sits inside a security program whose controls are working.","B) The entry names no AI-specific threat, so nothing shows that the controls answer attacks on a language model.. The entry could have been written for a payroll server. It names no prompt injection, poisoning or other threat that works only on AI, so the finding is about the risk assessment before it is about any control, and B states it. Source: Credential Press, https://credentialpress.com/guides/aaia-quiz" "A purchasing agent built on a language model can raise orders of any value, and it raises large orders with no person approving them. Using OWASP's three root causes of excessive agency, which statement is correct? | A) The unlimited value is excessive functionality, and the fix is to remove the ordering tool altogether. / B) The missing approval is excessive permissions, and the fix is to narrow the service account's rights. / C) Both conditions are excessive autonomy, and a single approval step on every order fixes both of them. / D) The unlimited value is excessive permissions, and the missing approval is excessive autonomy.","D) The unlimited value is excessive permissions, and the missing approval is excessive autonomy.. OWASP traces excessive agency to functionality, permissions and autonomy, each with its own fix. A tool the agent does need, running with broader rights than the task requires, is excessive permissions, so the unlimited value belongs there. High-impact actions taken without a person approving them are excessive autonomy. Source: Credential Press, https://credentialpress.com/guides/aaia-quiz" "During fieldwork on a benefits eligibility model, you find that it has been wrongly refusing a group of applicants since a release last week, and the refusals are continuing. What should you do FIRST? | A) Inform management and the incident owner at once. / B) Suspend the model yourself, using the access granted for the audit. / C) Carry out a root cause analysis to identify the faulty change. / D) Record the matter as a finding for the audit report.","A) Inform management and the incident owner at once.. The harm is live, so the first job is to get the incident to the people with authority to contain it, which A does. B makes a management decision, and an auditor who takes it has joined the control they are there to judge. Source: Credential Press, https://credentialpress.com/guides/aaia-quiz" "Your team ran an anomaly-detection tool over 120,000 auto-approved refunds and tested the 50 cases it flagged. Of those, 12 lacked the required second approval. A colleague's draft says: ""About 24% of auto-approved refunds lack second approval."" Which response BEST corrects the draft? | A) Limit the statement to the 50 cases the tool selected. / B) Keep the sentence, because every flagged case was tested in full. / C) Lower the figure to allow for the refunds the tool did not flag. / D) Add the confidence level and sampling risk that apply to the 24% figure.","A) Limit the statement to the 50 cases the tool selected.. The tool picked the 50 cases for their characteristics, so this is a targeted selection. The other refunds had no chance of selection, so the result speaks only for the items tested and gives no basis to project a rate. Report the 12 exceptions for what they are, and if the objective needs a conclusion about all refunds, draw a random sample as well. Source: Credential Press, https://credentialpress.com/guides/aaia-quiz" "Your draft finding says: ""The claims model treats rural customers unfairly."" Your manager returns it, asking what it is measured against. Which revision BEST states the condition? | A) ""Rural customers appear to receive worse outcomes, which management should investigate."" / B) ""The claims model breaches the organization's fairness policy for rural customers."" / C) ""For version 2.3, January to June, rural claims were approved 9 points below urban claims, against a 4-point tolerance, across all 22,400 claims."" / D) ""Rural customers raised 3 times as many complaints about claims decisions as urban customers.""","C) ""For version 2.3, January to June, rural claims were approved 9 points below urban claims, against a 4-point tolerance, across all 22,400 claims."". The condition is what you found, stated as fact with its evidence: what happened, how many times, out of what population, over what period and for which model version. The revision that names the version, the period, the gap, the approved tolerance and the population is the one a reader can check against the workpapers. Source: Credential Press, https://credentialpress.com/guides/aaia-quiz"
Where next?
Also free: the AAIA interview questions and the AAIA mind map, plus every other credential on our study tools page. For the full syllabus, the AAIA Exam Guide and the AAIA Practice Questions go domain by domain.
Frequently asked questions
Is this AAIA quiz like the real exam?
No. It is a 10-question recall check written by Credential Press, not exam questions. Each answer cites its source.
Can I study the AAIA quiz offline?
Yes. Download the free flashcard CSV and import it into Anki or any flashcard app that reads CSV.
A payments company's fraud owner reports that a new model is 99.9% accurate on held-back test data. In that data, about 1 transaction in 1,000 is confirmed fraud. Which measure would BEST show you whether the model catches fraud?
Recall: the share of confirmed fraud cases that the model flagged. When fraud is 1 case in 1,000, a model that calls every transaction genuine is right 999 times in 1,000 and catches no fraud at all. A 99.9% accuracy figure can therefore hide a model that misses everything. Recall is the share of real positives the model caught, which is the question asked.
Which books go deeper on AAIA?

A study companion for the ISACA Advanced in AI Audit (AAIA) examination. 13 chapters, 371 pages.

Practice questions with full rationales, sized to the published domain weights. 216 pages.
How to credit this page
Using a figure or a passage from this page in an article, a course, a forum answer or AI-generated text? Please credit Credential Press and link to this page, so your readers can check the source and the date it was reviewed. Copy a ready-made citation:
Plain text
Markdown
HTML
Sources
- ISACA, AAIA Exam Candidate Guide (2026), with the Examination Content Outline of June 2025, checked against ISACA's published copy on 14 November 2026
- ISACA, AAIA Exam Content Outline, read 14 November 2026
- ISACA, Advanced in AI Audit (AAIA): prerequisites and certification steps, read 16 November 2026
- ISACA, Maintain AAIA Certification, read 16 November 2026
- OWASP, Top 10 for Large Language Model Applications
Credential Press is independent of ISACA.