AAIA Sample Questions With Answers

About these questions

These five questions are original. We wrote them in the best-answer style that ISACA describes for AAIA, with the domains in rough proportion to their weight. They are not real exam questions, and the real exam will word things differently. All five are single answer; ISACA’s candidate guide describes the real exam as four options with one best answer.

How to use these

Try each question without notes, and pick an answer before you open the explanation. Then read the explanation even when you were right, because the reason the closest wrong option fails is where the learning is. If you slip on a domain, revisit it in domains explained.

Question 1 (AI Governance and Risk)

A national regulator requires a mortgage lender to give each declined applicant the specific reasons for the decision. A neural network model is four points more accurate than a simple points-based scorecard that the credit team already understands. Which choice best fits this requirement?

  • A. Deploy the neural network and publish its overall accuracy rate
  • B. Deploy the neural network and send applicants a general policy statement
  • C. Deploy the scorecard and accept the small loss of accuracy
  • D. Deploy the neural network and log every input for later analysis
Show the answer and explanation

Correct answer: C. The regulator wants reasons for each applicant, and only a model that can be explained directly guarantees them for every case. Four accuracy points do not outweigh an obligation the lender cannot otherwise meet. Publishing overall accuracy, the closest wrong option, describes the population and says nothing about one applicant’s decision. A general policy statement lists rules rather than reasons, and logging inputs preserves evidence but does not explain an outcome.

Question 2 (AI Operations)

An intensive care unit uses a model that flags possible sepsis. Lowering the alert threshold would catch more genuine cases and add about thirty false alerts per day, which staff confirm they can handle. A missed case can be fatal and the unit has no second safety net. Which threshold decision is best?

  • A. Keep the higher threshold so staff trust each alert that appears
  • B. Lower the threshold and monitor how staff cope with the extra alerts
  • C. Set the threshold where false alerts and missed cases are equal in number
  • D. Keep the higher threshold and review missed cases at the end of each month
Show the answer and explanation

Correct answer: B. A threshold is a judgment about which error is worse. A missed case can be fatal, nothing else will catch it, and the unit has confirmed capacity for more alerts, so sensitivity should win, with alert fatigue monitored as the accepted cost. Setting equal error counts, the closest wrong option, treats a missed case as no worse than a needless check. Keeping a high threshold to protect trust puts convenience ahead of patient safety. Monthly review of misses is a learning control and comes too late to help the patient.

Question 3 (AI Operations)

A software company bans all external AI assistants and offers no internal alternative. A survey shows that many developers paste code into personal accounts on their own devices, where the company has no visibility. Which security decision is best?

  • A. Add network filtering and device restrictions to tighten the ban
  • B. Provide a sanctioned internal assistant with logging and data controls
  • C. Allow external assistants for code that is labelled as low sensitivity
  • D. Publish a disciplinary policy against using external assistants
Show the answer and explanation

Correct answer: B. A control that is too restrictive to work does not remove the risk. It moves the risk to places the company cannot see. A sanctioned assistant with logging, retention limits and data loss controls restores visibility and keeps the benefit. Tighter blocking is the closest wrong option, but it treats the symptom: developers use personal devices precisely because the block already fails. Discipline has the same weakness, and allowing low-sensitivity code relies on labels that people rarely apply reliably.

Question 4 (AI Operations)

A bank wants to release a new fraud-scoring model before a seasonal peak in three weeks. Validating every product line needs about eight weeks. Errors on card disputes delay refunds to customers, while errors on a low-value internal reporting use cause little harm. Which release approach is best?

  • A. Release for all uses and finish validation while the model is live
  • B. Release for all uses with a manual check of a sample of decisions
  • C. Release only the uses already validated and defer the others
  • D. Postpone the whole release until every use has been validated
Show the answer and explanation

Correct answer: C. Schedule pressure may change the scope of a release, but it must not lower the evidence required for what goes live. Releasing the validated uses delivers value on time, while customers in the sensitive areas stay on the existing process. The deferral should be documented as a condition of approval. Validating in production, the closest wrong option, moves the risk to customers. Sampling afterwards is a detective control that cannot prevent the harm, and postponing everything throws away benefit that is already proven.

Question 5 (AI Auditing Tools and Techniques)

An auditor must give the audit committee a defensible deviation rate for the 1,200 model change approvals recorded this year. Available hours allow about sixty items to be examined, however they are chosen. Which sampling approach best supports the conclusion?

  • A. A block of sixty consecutive approvals from the busiest month
  • B. A judgmental sample of approvals the model owners consider sensitive
  • C. A haphazard selection spread across the twelve months
  • D. A statistical attribute sample, because the rate must be projected
Show the answer and explanation

Correct answer: D. To project a deviation rate onto the whole population, each item needs a known chance of being selected, which statistical sampling provides along with a stated confidence level. A judgmental sample, the closest wrong option, is useful for targeting risk but supports no projection. Haphazard selection only imitates randomness, and a single busy month assumes the other eleven behave the same. The committee asked for a rate, so only the statistical approach answers it.

More guides for AAIA

Next step

For more questions in this style, with an explanation for each option, see our AAIA practice tests. To plan your preparation, read the study plan.

Try these practice exams