AI Risk Classification Matrix

How to use this document

• What it is for. This is a reference matrix to classify each AI system or use case with traceable criteria, in parallel under each regime that may apply (for example, the EU AI Act, Peru's AI regulation and the Colorado and California rules on automated decisions) and under the organization's internal criteria, and to assign minimum controls to each level.
• Who should own it. The AI lead maintains the matrix; the AI committee (or management, in small organizations) validates High classifications. Each system owner proposes the classification of their system. Enter the function in [RESPONSIBLE FUNCTION].
• Decisions to take first. Decide in which jurisdictions the organization operates and therefore which columns of the parallel classification sheet (Section 5) it uses; whether it adjusts the internal-criteria thresholds (Section 6) within the risk appetite approved in the "AI Risk Appetite Statement"; and which additional controls it requires at each level (Section 7). Who approves each level, and by when, is set in the "AI Governance Operating Model".
• Fields to complete. Complete one classification sheet (Section 5) per system and carry the result over to the AI system inventory. Replace the bracketed fields.
• Country variants. The variants contain each regime's criteria: European Union, Peru and United States (Colorado and California). Keep those for the places where the organization develops, sells or uses AI, or where affected people are, and delete the rest.
• Review. Regulatory classification is a legal question. Have internal legal or outside counsel validate high-risk classifications, exclusions and borderline cases. The [COMPLETE] fields in the variants mark points to confirm against the current official text.
• Approval and use. Have the AI committee approve the matrix, record the approval in "Version history," and use it before buying, developing or deploying any AI system and whenever its purpose or use changes.

AI Risk Classification Matrix

Field	Value
Organization	[LEGAL ENTITY NAME]
Owning function	[RESPONSIBLE FUNCTION]
Matrix owner	[OWNER]
Classification validation	[AI COMMITTEE]
Approved by	[APPROVED BY]
Effective date	[DATE]
Review frequency	[REVIEW FREQUENCY]
Jurisdictions considered	[COMPLETE]

1. Purpose and scope

1.1 This matrix sets out how [ORGANIZATION NAME] classifies its AI systems and use cases by risk level, with traceable criteria, and which minimum controls apply to each level.

1.2 It applies to all AI systems developed, acquired, integrated or used by the organization, including AI features in products already under contract and AI agents, before they are put into use and whenever they change.

2. Classification principles

2.1 Purpose and actual use. Classification is based on the system's intended purpose and actual use in the organization, not on the technology or the vendor's marketing description.

2.2 Unit of analysis. When several components, models or agents together produce an output that significantly influences a decision, they are classified as a single system.

2.3 Parallel classification. The same system may be subject to several regimes at once. It is classified separately under each applicable regime (Section 5) and under the internal criteria (Section 6). Categories do not carry over from one regime to another: a system may be high risk under one regime and not covered by another.

2.4 Single internal level. The system's internal level is the highest one resulting from the parallel classification and the internal criteria. That level determines the minimum controls in Section 7, without prejudice to each regime's specific obligations.

2.5 Prohibitions first. Before any other assessment, check that the use is not prohibited in any applicable jurisdiction.

2.6 Written rationale. Every classification, and especially every conclusion that a system is not high risk, is justified in writing with reference to the criterion applied.

2.7 When in doubt, go higher. Until a borderline classification is resolved, the system is treated at the highest level under discussion.