The AI Act is a product-safety law. Here is the operator's map.

The AI Act reads differently once you notice which legal family it comes from. It is a product-safety law, a sibling of the rules that govern elevators, toys, and medical devices in Europe, and its unit of regulation is the use rather than the model. Its central instrument is the conformity logic behind the CE mark: classify the use, attach obligations to the class, make the obligations checkable.

The classification has four tiers, and the tier follows what you do with the system rather than which model sits inside it.

Unacceptable risk means prohibited. These practices have been banned since 2025-02-02: manipulation techniques that cause significant harm, exploitation of a group's vulnerabilities, social scoring, crime prediction based on profiling, untargeted facial-image scraping, emotion recognition in the workplace and in schools, biometric categorization that infers sensitive traits, and (with narrow exceptions) real-time biometric identification by law enforcement in public spaces.

High risk means permitted with heavy obligations. This is Annex III: credit scoring, education admissions, critical infrastructure, law enforcement, and employment. Recruitment and CV filtering sit here, in point 4. The high-risk wave was scheduled for August 2026 and was moved to 2027-12-02 by the AI Omnibus, in force 2026-07-27.

Transparency risk means disclosure duties. Article 50 has applied since 2026-08-02: a chatbot must present itself as a machine, synthetic content must carry machine-readable marking, deepfakes and AI-generated public-interest text must be visibly labeled.

Minimal risk is everything else, where the regulation asks nothing specific.

The same engine sits in different tiers depending on the use. A model that summarizes your internal meetings is unremarkable. The same model filtering job applications is a high-risk system with obligations for its provider and for you as its deployer. Europe regulated the driving rather than the motor, the way traffic law does, and the way pharmaceutical law attaches its rules to prescriptions rather than molecules.

The text splits responsibility between the provider, who builds the system and carries the heavy documentation duties, and the deployer, the organization that uses it and carries its own, lighter, very real duties: use the system as instructed, run human oversight with competent people, keep the logs, inform workers. Like the GDPR, the scope is extraterritorial. Serving EU users or EU subjects brings you inside the perimeter wherever you are incorporated.

The penalty structure has two floors. A prohibited practice costs up to 35 million euros or 7% of worldwide turnover, whichever is higher. An ordinary breach, transparency included, costs up to 15 million or 3%.

The unit of compliance is a register of uses, not a list of models. Every AI system you run needs a declared use, an owner, a tier, and evidence of the controls that tier demands. Organizations that maintain that register treat each new obligation as a row update; without one, the same inventory tends to get rebuilt under deadline each time a new wave applies.

Our own bias here comes from operating an agent fleet: the register only stays current when it falls out of daily operation rather than out of a compliance project. Keeping every agent behind a declared owner, a declared scope, and an approval path that leaves evidence turned out to be the same work as keeping the compliance picture, an observation our BAMS control plane grew out of. A register kept that way is an inventory with evidence and nothing more; it does not make a use legal, and the legal determination stays with counsel.

Sources: Regulation (EU) 2024/1689 (AI Act), Arts 4, 5, 50, 99, 101, Annex III; the AI Omnibus amending it; analysis video: "Le véritable problème de l'AI Act", IA et Stratégie, Le SamourAI, 2026-08-03. This article is legal information, not legal advice.