AI risk management

In plain English

Think of a ship's captain. Nobody can remove the weather, so the captain checks forecasts, plots a route around storms, carries lifeboats and keeps watch for the whole voyage. AI risk management does the same. Find what could go wrong, judge how likely and how bad it would be, decide what to do, and keep watching.

The watching matters more for AI than for most technology. Models change as data, users and the world change, so a risk that was small at launch can grow. The NIST AI RMF defines risk as a combination of an event's probability and the size of its consequences.

Example

Calloway Health, a fictional hospital network, plans an AI tool that predicts which patients will miss appointments, so staff can offer reminders or transport.

1. Identify. The team maps stakeholders, including patients who never see the tool. One harm stands out: if flagged patients' slots are double-booked, those patients get shorter visits, and low-income patients may be flagged more often.
2. Assess. Each harm is scored on probability and severity. Unfair double-booking is likely and severe, so it ranks first. A badly timed reminder is likely but minor.
3. Evaluate. The scores are compared with the risk tolerance the board approved.
4. Treat. The team climbs the mitigation hierarchy from the top. It avoids the worst harm by design: a flag can trigger extra support but never double-booking. It reduces the rest with fairness testing by group.
5. Monitor. Flag rates by group are reviewed monthly and reassessed whenever the model is retrained.

The residual risk, its owner and the reasoning go into the risk register.

Why it matters

• NIST AI RMF 1.0 (voluntary) organizes the work into GOVERN, MAP, MEASURE and MANAGE. Each organization sets its own risk tolerance, and AI risk should be integrated into enterprise risk management. Where risk is unacceptable, NIST says development or deployment should stop in a safe manner until the risk can be managed.
• ISO. ISO/IEC 23894 adapts ISO 31000 risk guidance to AI. ISO/IEC 42001 requires AI risk assessment, risk treatment and AI system impact assessment inside a certifiable management system.
• EU AI Act. Art. 9 requires providers of high-risk systems to run a continuous risk management system. It must cover reasonably foreseeable misuse and consider effects on children and other vulnerable groups. Art. 9(5) sets an order: eliminate or reduce risk by design, then add controls, then inform and train deployers. After the Digital Omnibus, this applies from 2 Dec 2027 for Annex III systems and 2 Aug 2028 for Annex I systems.

How the AIGP exam tests it

Domain III tests the toolkit (lesson III.A.5); Domain II tests the NIST structure (lesson II.D.2).

Common traps:

• Rare but catastrophic gets ignored. Low probability is not zero. Rare and severe harms need a contingency plan.
• Disclaimers, insurance or compensation chosen first. These sit at the bottom of the mitigation hierarchy. Insurance transfers the organization's financial exposure but protects nobody. Avoiding the risk at source ranks highest.
• A one-time check. Both NIST and Art. 9 describe a continuous, life-cycle process.
• A separate AI risk silo. NIST wants AI risk integrated with cybersecurity, privacy and enterprise risk.

The tell: the best answer ranks risks before treating them, prefers prevention to cure, and keeps monitoring after launch.