Glossary
Updated
What is AI risk management?
AI risk management is the cycle an organization runs to decide which AI risks it will accept, reduce, share or avoid. It applies to every AI system in use, whether it was built in-house, bought from a vendor or arrived as a new feature in an existing tool. It runs for the life of the system, not once before launch, because AI changes in use: models are updated, data drifts and people find new ways to use a tool.
What are the steps?
Most programs follow the same sequence. First, know what you have: an AI inventory with an owner for each system. Second, classify each system by its use and impact, so effort goes where harm is most likely. Third, identify the risks for that use, such as wrong outputs, unfair treatment, data leakage, security attacks and vendor failure.
Fourth, score each risk for likelihood and impact before any controls, which gives its inherent rating. Fifth, choose a treatment: controls, human review, limits on use, or not going ahead. Sixth, score it again for the residual risk, and have someone with authority accept it. Last, monitor the system and repeat the assessment when something material changes.
Which frameworks cover AI risk management?
The NIST AI Risk Management Framework, published in January 2023, organizes the work into four functions: GOVERN, MAP, MEASURE and MANAGE. Its Generative AI Profile (NIST AI 600-1, July 2024) sets out 12 risks that generative AI creates or makes worse. ISO/IEC 23894 gives guidance on AI risk management built on ISO 31000, and ISO/IEC 42001 requires AI risk assessment, risk treatment and AI system impact assessment as part of a certifiable management system.
Under the EU AI Act, Article 9 requires providers of high-risk systems to run a risk management system across the system's whole life. For uses listed in Annex III, that duty applies from December 2, 2027.
How is it different from model risk management?
Model risk management grew up in banking and asks whether a model is sound and fit for its purpose. AI risk management covers that and more: vendor and embedded AI, tools staff use without approval, agents that take actions, and duties under several laws at once. In the US, SR 26-2, which replaced SR 11-7 on April 17, 2026, puts generative and agentic AI outside its scope, so banks need another way to govern those systems.
What makes it work?
Three things help. Each risk has an owner with the authority to accept it. Risks sit in one register, so security, privacy and legal don't score the same issue three different ways. And each treatment has evidence attached, so the rating holds up at audit.
In Enzai, each mitigation is tracked from not started to completed, with proof attached, next to the risk's rating before and after treatment. See AI risk and compliance.
Example
A regional bank plans to use a large language model to draft replies to customer complaints. It records the system in its inventory and classifies it as customer-facing with personal data. The assessment finds three main risks: wrong statements about a customer's account, personal data reaching the model provider, and replies that breach conduct rules. All three rate high before controls. The bank adds human review before any reply is sent, a contract term barring the provider from training on its data, and a monthly check of a sample of sent replies. The residual ratings fall to medium and low, the head of complaints accepts them, and the system is reassessed whenever the model version changes.