What is AI risk management?
AI risk management is the ongoing process of identifying what could go wrong with an AI system, judging how likely and how serious each harm is, and deciding what to do about it. It applies the discipline that companies already use for financial, safety, and cybersecurity risk to a new kind of system, one that learns from data, gives different answers to similar inputs, and can act on other systems through tools.
The process repeats through the life of a system. Risks are identified at design, re-assessed before launch, and reviewed whenever the model, the data, or the use changes.
How AI risk management works
Four steps form the loop. Identification lists the harms: wrong outputs, biased outcomes, leaked data, unsafe actions, legal exposure, and dependence on a vendor that may change its model without notice. Analysis rates each harm for likelihood and impact, often on a simple scale, and separates harms to the company from harms to customers, employees, and the public. Treatment chooses a response for each harm: reduce it with a control, transfer it by contract or insurance, accept it with a documented sign-off, or avoid it by not building the feature. Monitoring checks that the controls still work, using model monitoring data, incident reports, and periodic review.
The NIST AI Risk Management Framework organizes this loop into its Map, Measure, and Manage functions. The EU AI Act makes a documented risk management system mandatory for high-risk AI systems. Banks apply a closely related discipline under model risk management.
AI risk management in practice
A hospital network plans to use a language model to draft discharge summaries. The risk workshop identifies a summary that omits a medication change as the most serious harm, with a leaked patient record close behind. Treatment includes a mandatory physician sign-off before any summary enters the chart, an automated check that every medication in the visit record appears in the draft, and running the model inside the hospital’s own environment so that no record leaves it. The chief medical officer accepts the residual risk in writing. Monitoring samples fifty summaries a month and tracks every correction a physician makes.
What to watch for
- Rate harms to people separately from harms to the business, since regulators care most about the first.
- Include reasonably foreseeable misuse, such as staff pasting data the system was never approved to handle.
- Treat a vendor’s model update as a change that reopens the assessment.
- Record who accepted each residual risk and when, because that record is what an auditor asks for first.
- Keep the register short enough that owners read it, because a hundred-line spreadsheet nobody maintains manages nothing.