AAIR lesson 1: what makes AI risk different
AI risk differs from classic IT risk because the behaviour of an AI system comes from data and training rather than from rules a person wrote, so the system can be wrong in ways nobody coded and can change without anyone touching it. The AAIR exam (ISACA Advanced in AI Risk) covers this field: it has 90 questions in three domains weighted 37%, 21% and 42%, and each domain follows the work of a risk professional from governance, through the life of a system, to running the risk programme.
Three differences from classic IT risk
A traditional application does what its code says. If a payroll calculation is wrong, a defect can be found and fixed, and the same input gives the same output tomorrow. AI systems break those assumptions in three ways.
- Behaviour is learned, not written. A model infers patterns from training data. Flaws such as bias come from that data and cannot be found in a line of logic.
- Outputs are probabilistic. The same question can produce different answers, and “correct” becomes a threshold (how often wrong, and how badly) rather than a yes or no.
- Risk moves over time. When real-world data drifts away from the training data, accuracy can fall with no change to the system. A control that was sufficient at launch may not be later.
None of this replaces classic risk thinking: confidentiality, availability, vendor risk and compliance still apply, and the vocabulary (likelihood, impact, appetite, tolerance, residual risk) is unchanged. AI adds harm to people outside the organisation, such as a customer wrongly declined. The risk management page covers the terms.
How the exam is organised
The figure above shows the structure. ISACA publishes the domain names, their weights and the subtopic headings, plus a list of tasks that cut across domains. It does not publish detailed content, so a study plan rests on the headings and on well-established risk practice.
| Domain | Weight | What it covers | Lessons |
|---|---|---|---|
| 1. AI Risk Governance and Framework Integration | 37% | Use cases, structures, ownership, policy, regulation, ethics | 1 to 4 |
| 2. AI Life Cycle Risk Management | 21% | Design to decommissioning, data and assets | 5 to 7 |
| 3. AI Risk Program Management | 42% | Scenarios, treatment, controls, metrics, suppliers, incidents | 8 to 12 |
Subtopic 1A in practice: what a use-case decision weighs
The first governance heading covers AI models, frameworks, strategies and use cases. The practical skill is deciding what risk level a proposed use carries, so that effort is proportionate.
Consider a 200-person insurance broker with two requests on the same week. The marketing team wants a generative assistant to draft newsletter copy that a person edits before sending. The claims team wants a model that scores incoming claims and automatically declines those it rates as likely fraud. A risk professional compares them on a few questions.
| Question | Newsletter drafting | Automatic claim declines |
|---|---|---|
| Who is affected if it is wrong? | Staff, who see and fix it | Customers, who lose a payout |
| Is a person in the loop? | Yes, an editor approves | No, decisions are automatic |
| Can the harm be reversed? | Easily | Slowly, via complaints |
| Is the data personal or sensitive? | Little | Yes |
| Does a regulation apply? | Unlikely | Quite possibly |
The first needs a short approval; the second needs an impact assessment, testing for unfair outcomes, human review and monitoring. Neither answer is “ban AI”: the decision compares each case with risk appetite and records who accepted what.
Easy to confuse
- AI risk versus AI engineering. Questions are about decisions, not model internals.
- A framework versus the programme. Choosing a framework is one input. The programme is the roles, register, controls and reporting that make the framework operate.
- Separate versus integrated. AI risk is integrated into the existing enterprise risk approach (register, appetite, reporting). A parallel structure is usually the weaker answer.
Key points
- AI behaviour is learned from data, outputs are probabilistic and risk can shift over time through drift.
- Classic risk vocabulary and the enterprise risk approach still apply; AI adds new harms, including to outside parties.
- AAIR has three domains: governance (37%), life cycle (21%) and programme management (42%).
- A use case is judged by who is affected, human involvement, reversibility, data and regulation.
Check yourself
Question 1. A 150-person retailer is asked to approve two AI projects: a tool that suggests product descriptions for staff to edit, and a tool that sets credit limits for customers without review. Which BEST explains why they need different governance?
- A. Only the credit limit tool uses a machine learning model
- B. Only the product description tool depends on a third-party vendor
- C. The credit limit tool can harm outside parties directly, with no human check
- D. The credit limit tool is technically more complex to build
Show answer
Answer: C
Governance is proportionate to the harm the use case can cause, and automatic credit limits affect customers directly with no human check. Both projects use AI and both could involve a vendor, so those facts do not separate them, and technical complexity is not the deciding factor.
Want more questions like this? Full AAIR practice tests →
Question 2. A model that approved loans accurately at launch now performs worse, although nobody has changed the model or its code. Which AI characteristic BEST explains this?
- A. Real-world data has drifted from the training data
- B. A recent software update introduced a code defect
- C. The model consumed its training data through use
- D. The licence for the model has expired
Show answer
Answer: A
Drift occurs when live data moves away from the data the model learned from, so quality can fall with no change to the system. A code defect would require a change or an existing bug in logic, and a model does not run out of licences or lose its training data by being used.
Want more questions like this? Full AAIR practice tests →
Question 3. An enterprise already has a risk register, appetite statement and reporting cycle. How should AI risk be handled?
- A. Create a separate AI register with its own appetite statement
- B. Wait until a regulation requires an AI-specific process
- C. Leave AI risk to the technology team to track informally
- D. Add AI risks to the existing register and reporting, using the same appetite
Show answer
Answer: D
Integrating AI risk into the existing enterprise risk framework keeps appetite, rating and reporting consistent and lets AI risk be compared with other risks. A parallel programme duplicates effort and produces inconsistent ratings, and ignoring AI risk until a regulation applies leaves exposure unmanaged.
Want more questions like this? Full AAIR practice tests →
This lesson opens Domain 1, AI Risk Governance and Framework Integration (37% of the exam), and sets up the rest of the course. The recap in AAIR governance and frameworks condenses the domain, and the topic page on AI governance and risk compares the main instruments.