How to classify the risk of an AI system
A practical method to classify an AI system: prohibited practices, Annexes I and III, Article 6(3) exceptions, transparency and documentation.
AI ACT · CLASSIFICATION AND DECISION
Legal risk is not determined by the model name, parameter count or the impression that a technology is “advanced”. The AI Act starts with intended purpose, context of use, affected people and the output’s actual influence.
There is no universal risk score
A risk-based approach does not mean assigning every tool a fixed score from one to five. The same model may draft a product description, rank job applicants or become a safety component in a device. Classification differs because intended purpose and effect differ. [EU-REG] [EU-RISK]
Before classification, describe the concrete function: who provides the system, who uses it, which data it receives, what output it produces, who sees it and which decision it may influence. “We use AI for efficiency” is not enough information for a conclusion. [EU-NAV] [EU-RISK]
Gate 1: check prohibited practices
The first test is not whether a system is high-risk, but whether its purpose or method falls under Article 5. The Regulation prohibits, among other things, certain harmful manipulation, exploitation of vulnerabilities, social scoring, some predictive policing, untargeted facial-image scraping and certain biometric or emotion-recognition uses. [EU-REG] [EU-PROH]
A possible match with a prohibition stops launch and requires legal analysis, not a cosmetic mitigation. Commission guidance provides examples but is non-binding; authoritative interpretation belongs to the courts. The file should retain the purpose description and the reason Article 5 applies or does not apply. [EU-PROH]
Gate 2: products in Annex I
Article 6(1) classifies a system as high-risk when it is a safety component of a product covered by Annex I legislation, or is itself such a product, and the product must undergo third-party conformity assessment before market placement or commissioning. [EU-REG] [EU-RISK]
Both conditions are cumulative. An AI feature in a regulated product is not automatically high-risk if it lacks the relevant safety role or the required third-party assessment. Classification should be performed with the product-compliance specialist, not by software staff alone. [EU-REG]
Gate 3: intended purposes in Annex III
Annex III lists uses in biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration and asylum, justice and democratic processes. “Intended” is decisive: a general tool is not high-risk merely because it could technically be used in such a field. [EU-REG] [EU-RISK]
For example, AI used to filter CVs or evaluate applicants enters the Annex III employment area. A tool that spell-checks a job advert does not have the same influence. Document the output, user and stage in the process; the department label alone does not determine classification. [EU-REG]
The narrow Article 6(3) exceptions
An Annex III system may be considered non-high-risk where it poses no significant risk to health, safety or fundamental rights, including because it does not materially influence the decision outcome. The Regulation identifies four situations: a narrow procedural task, improvement of a completed human activity, pattern detection without replacing or influencing human assessment, and a preparatory task. [EU-REG] [EU-RISK]
The exception is not automatic. Profiling people remains high-risk in this context, and a provider relying on Article 6(3) must document the assessment before launch and meet the registration obligation. The Commission draft helps with examples, but the conclusion must be tied to facts and reviewed when purpose changes. [EU-REG] [EU-RISK]
Transparency does not mean high-risk
A chatbot, deepfake or content-generation system may fall under Article 50 transparency duties without being high-risk. Keep separate register fields for prohibited practice, Annex I path, Annex III path, transparency and other applicable law. Otherwise teams confuse a disclosure duty with the full high-risk regime. [EU-A50] [EU-REG]
Ordinary uses still require governance. An SEO draft at AYSA.RO, stock forecast in ProFlorist or recommendation in AdverLink or CanUHelp APP is not automatically high-risk. Yet new purpose, influence on people, personal data, transparency and human control need separate assessment; GDPR, consumer and contract law still apply. [EU-REG] [EU-NAV] [EU-A50]
The decision must be reproducible
The classification record should retain system version, purpose, company role, affected people, checked annexes and articles, any Article 6(3) condition relied upon, evidence, approver and next review date. Attach provider instructions and archive the version reviewed. [EU-REG] [EU-RISK]
Repeat the analysis when model, data, audience, autonomy, integration, offering brand or influenced decision changes. An old conclusion does not protect a new use. Mature compliance means not merely naming a category quickly, but explaining why it was selected and which event triggers reassessment. [EU-REG] [EU-RISK]