Control · C-056 · Governance & Accountability
AI risk tiering and classification
Each system is assigned an internal risk tier and a regulatory class by documented criteria, the result is reviewed independently, and it is revisited whenever the system or its use changes.
- Control type
- Directive
- Risks addressed
- 6
- Audience
- Heads-of · Practitioner
- Frameworks
- Article 6 · MAP 2 · Principle 1
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-10
What it does
It decides how much governance each system receives. Documented criteria sort systems into tiers that set the depth of validation, testing, oversight and monitoring, and determine legal class, including whether a system is high-risk under the EU AI Act and what role the organisation holds for it. The system owner proposes the tier; a governance or second-line function confirms it. It operates at intake and at every material change.
How to implement it
- Write criteria based on use and consequence: who is affected, how seriously, how reversibly, with how much autonomy.
- Include materiality, complexity and the degree of reliance on the output.
- Add the legal tests, with legal review for borderline cases.
- Record the reasoning as well as the result.
- Have someone outside the project confirm each classification.
- Reclassify on change of purpose, population, autonomy or supplier terms, and test whether a modification makes the organisation a provider.
Evidence it operates
Design: tiering criteria and a classification procedure. Operation: a recorded tier and rationale for each system; evidence of independent confirmation; examples of a tier raised on challenge or after a change; a spread of tiers that is plausible for the portfolio.
Where it fits
Follows intake and drives everything after it. It is the primary control for misclassified AI risk tier and unrecognised provider obligations, sets the depth of independent model validation, and identifies the systems that need high-risk conformity assessment and registration.
Limits
Classification is a judgement made early, on a description supplied by people who want approval. Criteria cannot anticipate every use. A tier is correct only for the use that was described, which is why it depends on the intended-use definition and on change management to bring systems back.
Risks this mitigates
Primary mappings
- Misclassified AI risk tier High
- Applies documented criteria, independently reviewed, to assign tier and regulatory class. A system is assigned a lower risk tier or regulatory class than it warrants, so it receives lighter controls than the harm it can cause.
- Unrecognised provider obligations Medium
- Classification on change tests whether a modification or rebranding makes the organisation a provider. An organisation rebrands, substantially modifies or repurposes an AI system and takes on a provider's legal duties without realising it.
Supporting mappings
- Conformity assessment and registration failure High
- Correct classification identifies which systems need it. A high-risk system is placed on the market or put into use without the required conformity assessment, declaration, marking or registration.
- Inadequate independent model validation High
- Determines how deep validation must go. Validation is absent, is carried out by the model's own developers, or is not proportionate to the model's tier, so weaknesses are discovered in production.
- Inadequate pre-deployment testing High
- Sets how much testing is proportionate. A system goes live without testing proportionate to its risk, including functional, safety, security and user testing in realistic conditions.
Compensating mappings
- Opaque model logic High
- Where the logic cannot be explained, restricts the model to uses where opacity is tolerable. Developers, validators and owners cannot explain why a model produces its outputs well enough to challenge, correct or defend it.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| EU AI Act | Article 6 | Classification rules for high-risk AI systems Regulatory classification. |
| NIST AI Risk Management Framework (AI RMF 1.0) | MAP 2 | The AI system is categorised Categorisation. |
| SS1/23: Model Risk Management Principles for Banks | Principle 1 | Model identification and model risk classification Risk-based tiering. |
Related terms
How to cite
Axiom Verity (2026). AI risk tiering and classification. Reference Library, C-056. https://axiom-verity.com/controls/ai-risk-tiering. Reviewed 2026-11-10.
Found an error or a better source? Report a correction.