What is AI risk classification?
AI risk classification is the process of assigning an AI system or use case to a defined risk level based on its purpose, context, capabilities and potential impact. The classification determines which reviews, approvals, controls and monitoring requirements should apply. It allows an organisation to focus the strongest governance on systems that could materially affect rights, safety, access to opportunities or essential services.
Classification is not simply a technical rating of model accuracy. It considers who may be affected, whether the system influences significant decisions, the sensitivity of the data, the degree of autonomy, the scale of use, the reversibility of outcomes and the possibility of discrimination, security failure or misuse.
How are AI risk levels determined?
Organisations normally begin with an intake questionnaire that captures intended purpose, users, affected groups, data, outputs and decision authority. The answers are evaluated against internal criteria and relevant legal categories. Some frameworks use labels such as low, medium, high and prohibited, while others use numerical scores or separate impact dimensions.
The classification should be supported by written reasoning. Two systems with similar technology may receive different ratings because one recommends entertainment content and another screens job candidates. The context of deployment, not only the model type, is central to the assessment.
What controls follow from the classification?
Lower-risk systems may require basic inventory, acceptable-use controls, security review and owner approval. Higher-risk systems may require a formal impact assessment, independent validation, bias testing, explainability, human oversight, enhanced vendor due diligence, legal review, executive approval and continuous monitoring. A prohibited category may mean that the use cannot proceed unless the design changes.
The control matrix should be clear enough that project teams know what evidence is required before launch. It should also distinguish between the initial risk of the use case and the residual risk remaining after safeguards have been implemented.
How does classification relate to the EU AI Act?
The EU AI Act uses a risk-based structure that includes prohibited practices, high-risk systems, transparency obligations and requirements for general-purpose AI. An internal classification framework can help organisations identify potential legal categories, but it should not replace a detailed role and scope analysis.
Organisations operating across jurisdictions often create one enterprise classification method and then map local requirements onto it. This reduces duplication while preserving the need for legal review where a system may fall into a regulated category.
Frequently asked questions
Is AI risk classification the same as a full risk assessment?
No. Classification is an early triage step that determines the depth of assessment and controls. A detailed risk or impact assessment may follow for medium- or high-risk systems.
Can an AI system change risk level?
Yes. A new purpose, wider deployment, different data, increased automation or use with a vulnerable population can materially change the classification.
Who should approve the classification?
The business owner should provide accurate context, while a governance or risk function should review the result. Higher-risk classifications may require legal, privacy, security or executive approval.
Should vendor AI be classified?
Yes. Buying a system does not remove the organisation’s responsibility for its use. The deployment context, configured features and affected people still need to be assessed.
How often should classifications be reviewed?
They should be reviewed at defined intervals and whenever there is a material change, incident, regulatory update, model drift or evidence that the original assumptions are no longer valid.