What an AI risk assessment produces, the six steps to run one, and how it differs from an impact assessment and a fundamental rights impact assessment.
An AI risk assessment is a structured evaluation of what could go wrong with a specific AI system or use case, how likely each harm is and how severe it would be, so the organization can decide whether and how to proceed.
It is one step inside AI risk management, repeated whenever the system changes.
The sequence below follows ISO/IEC 23894 and Article 9(2) of the EU AI Act, with NIST AI RMF references for each step.
Name the system, its intended purpose and the scale you will score on. ISO/IEC 23894 calls this defining risk criteria (clause 6.3.4).
List known and reasonably foreseeable risks to health, safety and fundamental rights, including from foreseeable misuse (Article 9(2)). Add security, legal and third-party risks.
Score each risk on one scale. NIST’s Playbook describes the common method as impact multiplied by likelihood, on a scale as simple as red, amber, green.
Compare each score with your documented risk tolerance (MAP 1.5) and rank treatment by impact, likelihood and available resources (MANAGE 1.2).
Mitigate, transfer, avoid or accept (MANAGE 1.3). Record the residual risk and the person who accepted it.
Repeat on a schedule and when the model, data, vendor or use changes. Article 9 describes a continuous, iterative process, not a one-time review.
Three assessments that are often confused. They ask different questions and have different legal weight.
| AI risk assessment | AI impact assessment | Fundamental rights impact assessment | |
|---|---|---|---|
| Core question | What could go wrong, how likely and how severe? | Who is affected, and how, for better or worse? | How could this deployment harm the rights of the people it affects? |
| Reference | ISO/IEC 42001 clause 6.1.2; ISO/IEC 23894 | ISO/IEC 42001 clause 6.1.4; ISO/IEC 42005:2025 | EU AI Act, Article 27 |
| Legal weight | Voluntary in the standards; part of the Article 9 duty for high-risk providers | Voluntary guidance | Mandatory for certain deployers of high-risk AI |
| Output | Scored risk register and treatment plan | Documented impacts on individuals, groups and society | Results notified to the market surveillance authority |
Straight answers, kept current. Last updated .
Ask us anything →Use NIST’s seven characteristics of trustworthy AI as the prompt list: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. Then add the risks that arrive with third parties, which GOVERN 6.1 and MAP 4.1 address: models, data and software you did not build.
Providers of high-risk AI systems, as part of the risk management system in Article 9, which must identify, estimate and evaluate risks and adopt targeted measures. Testing against predefined metrics has to happen before the system is placed on the market (Article 9(8)). Certain deployers have a separate duty, the fundamental rights impact assessment in Article 27.
No. A questionnaire collects a vendor’s statements. An assessment scores the risk of your use of that vendor’s AI, using those statements as one input alongside documentation, test results and contract terms. NIST AI RMF GOVERN 6.1 asks for policies and procedures that address AI risks associated with third-party entities.
Primary sources for the figures and frameworks cited on this page. Last verified .
See how Cranium AI discovers, observes, governs, secures and proves every AI model, agent and vendor in your enterprise.
Cookie Policy · Privacy Policy
Global Privacy Control detected. Opt-out request honored: marketing tags stay off in this browser.