The short version
Use this sheet for the last pass before the exam, not as your only study source. It follows the three domains of ISACA’s AAIA outline and adds term pairs, decision rules and traps in our own words. It does not cover everything on the exam; ISACA’s content outline is the full list.
Exam facts at a glance
| Item | Fact (checked October 22, 2026) |
|---|---|
| Exam | ISACA Advanced in AI Audit (AAIA) |
| Questions | 90 multiple choice, four options, one best answer |
| Time | 150 minutes, one break up to 10 minutes (clock keeps running) |
| Passing score | 450 on a scaled 200 to 800 range |
| Wrong answers | No penalty, so answer everything |
| Domain 1 | AI Governance and Risk, 33% |
| Domain 2 | AI Operations, 46% |
| Domain 3 | AI Auditing Tools and Techniques, 21% |
| Retakes | 4 attempts in 12 months; waits of 30, 90, 90 days |
| After passing | Apply for the credential within 5 years |
Details are on the exam format page and the renewal and retakes page.
Domain 1: AI governance and risk (33%)
ISACA’s five areas: models and requirements; governance and program management; AI risk management; privacy and data governance; leading practices, ethics, regulations and standards.
Key concepts
- Model families. Know the difference between rule-based systems, supervised and unsupervised learning, and generative models. The audit approach depends on how the system learns and what it outputs.
- Governance layers. Policy says what is allowed. Standards and procedures say how. Roles (owner, approver, reviewer) say who. An AI use case register says what exists.
- Risk vocabulary. Inherent risk is before controls, residual is after. Appetite is how much risk the organization wants to take; tolerance is the acceptable variation around it.
- Privacy. Purpose limitation, data minimization, retention, and extra care for special categories of data.
Pairs people mix up
| This | Versus this |
|---|---|
| Inherent risk | Residual risk |
| Risk appetite | Risk tolerance |
| Policy | Procedure |
| Pseudonymization (can be reversed with extra data) | Anonymization (cannot reasonably be reversed) |
| Accountability (who answers) | Responsibility (who does the work) |
Decision rules
- Scenario names no owner for an AI risk: think ownership and governance first, tools second.
- Regulation is mentioned: pick the answer about principles (transparency, accountability, human oversight), not memorized article numbers.
- A constraint is stated (budget, time, data location): the best answer fits the constraint, not the most thorough answer.
Domain 2: AI operations (46%)
The biggest domain. Seven areas: data management; development methods and lifecycle; change management; supervision of AI; testing; threats and vulnerabilities; incident response.
Key concepts
- Data. Source, quality, bias, lineage, versioning, and who may use it for training.
- Lifecycle. Design, build, validate, deploy, monitor, retire, with a control or gate at each stage.
- Change. A retrained model is a change. Expect a request, testing, approval and a rollback path.
- Supervision. Monitor accuracy, drift, fairness and decisions; define when a human steps in.
- Testing. Performance, bias, robustness and repeatability.
- Threats. Poisoned training data, prompt manipulation, model extraction, inference about training data, plus ordinary access and supply chain weaknesses.
- Incidents. Detect, contain, investigate, recover, learn.
Pairs people mix up
| This | Versus this |
|---|---|
| Data drift (inputs change) | Concept drift (the relationship between inputs and outcome changes) |
| Validation (built the right thing) | Monitoring (still works in production) |
| Preventive control | Detective control |
| Poisoning (attack on training data) | Prompt injection (attack on the running model through input) |
| Model bias | Data bias |
| Testing the design of a control | Testing its operating effectiveness |
Decision rules
- A model is retrained or swapped by a supplier without notice: think change management and vendor oversight.
- Outputs get worse over time with no code change: think drift and monitoring.
- A decision harms a group: think fairness testing, human review and escalation.
- An AI incident happens: contain first, then investigate, then report and fix the root cause.
- Two options are good: the one with a stated threshold, owner and evidence usually beats a vague one.
Domain 3: AI auditing tools and techniques (21%)
Five areas: audit planning and design; testing and sampling; evidence collection; data quality and analytics; outputs and reports. The outline’s supporting tasks also cover using AI inside the audit and weighing the risks of doing so.
Key concepts
- Planning. Scope from risk, objectives, criteria, and what evidence you need.
- Sampling. Statistical sampling lets you project results; judgmental does not. Sample size depends on risk, tolerable error and population.
- Evidence. Sufficient means enough; appropriate means relevant and reliable. Evidence from an independent source is stronger than from the auditee.
- Analytics. Reconcile extracts to source before relying on them. Keep scripts reproducible.
- Reports. Findings need condition, criteria, cause, effect and a management action with an owner.
Pairs people mix up
| This | Versus this |
|---|---|
| Sufficient evidence | Appropriate evidence |
| Observation | Inquiry (observation is stronger evidence) |
| Statistical sampling | Judgmental sampling |
| Finding | Recommendation |
Decision rules
- The data extract does not match the source: stop and resolve before drawing conclusions.
- The auditee cannot explain the model: ask for documentation and independent validation evidence before testing outputs.
- You use an AI tool in the audit: assess its risks (accuracy, confidentiality) and keep human judgment on conclusions.
Common traps
- Answering the topic instead of the qualifier (BEST, FIRST, MOST likely).
- Choosing a technical fix when the root cause is ownership or policy.
- Picking a detective control when the scenario needs a preventive one.
- Forgetting third parties: much AI comes from vendors.
- Spending too long on a long scenario. Mark a guess and move on.
Last-day checklist
- Read the outline once more; tick each area you can explain in a sentence.
- Skim your wrong answers from practice, not new material.
- Know the format: 90 questions, 150 minutes, 450 to pass, no wrong-answer penalty.
- Check ID, name match, workspace and, for remote exams, the room-scan rules.
- Sleep. A tired brain misses qualifiers.
More guides for AAIA
- AAIA Sample Questions With Answers
- Is the AAIA Exam Hard? An Honest Look
- AAIA Prerequisites: Who Can Take the Exam
- Is AAIA Worth It? An Honest Look
- AAIA vs AAIR: Which ISACA AI Exam First?
- AAIA vs AAISM: Which ISACA AI Exam First?
Next step
For a full review of each area, see domains explained and the study plan; the overview is the AAIA exam guide. When you want to test yourself under time, our AAIA practice tests have five original sets. They show readiness and do not replace ISACA’s own material.