What the Act Is, Precisely
Regulation (EU) 2024/1689, published in the Official Journal of the European Union on July 12, 2024, is a product safety and fundamental rights regulation applied to AI systems. It does not regulate AI as a technology category. It regulates specific AI applications based on the risk those applications pose to health, safety, and fundamental rights. The regulation establishes four risk tiers, assigns mandatory obligations to each tier, and phases those obligations in across a four-year implementation window. Noncompliance carries fines up to €35 million or 7% of global annual turnover for the most serious violations, whichever is higher, under Article 99.
Non-EU companies should pause at that last clause: global annual turnover.
How the Tier Structure Works
The Act classifies AI systems into four tiers. The classification is not discretionary. The law assigns it based on what the system does, not on how the deployer characterizes it.
Unacceptable risk (prohibited). Article 5 enumerates practices that are banned outright: real-time remote biometric identification in public spaces by law enforcement (with narrow exceptions), social scoring by public authorities, subliminal manipulation that causes harm, and exploitation of vulnerabilities based on age or disability. No compliance pathway exists for these. The prohibition is the compliance requirement.
High risk. High risk carries the heaviest operational burden. Annex I covers AI embedded in regulated products — medical devices, machinery, aviation systems, vehicles — where existing EU product safety law already applies. Annex III covers standalone high-risk applications across eight domains: biometric identification, critical infrastructure management, education and vocational training, employment and worker management, access to essential services, law enforcement, migration and border control, and administration of justice. Systems in these categories require mandatory conformity assessments, technical documentation, human oversight mechanisms, and registration in the EU database before deployment.
Limited risk. Transparency obligations only. Chatbots must disclose they are AI systems. Deepfakes must be labeled. The obligations are real but narrow.
Minimal risk. No mandatory requirements. Most current commercial AI deployments fall here: recommendation engines, spam filters, inventory optimization tools.
For any given deployment, two questions determine everything else: which Annex applies, and does the use case match one of the eight Annex III domains?
The Four Implementation Waves
The regulation entered into force August 1, 2024. From that date, a staged clock started running.
February 2, 2025. Prohibited practices under Article 5 became enforceable. AI literacy requirements under Article 4 — which require providers and deployers to ensure staff have sufficient AI literacy for their roles — also took effect. These are already law. Organizations that haven't addressed them are already out of compliance.
August 2, 2025. Obligations for general-purpose AI (GPAI) models under Chapter V became applicable. GPAI covers large foundation models — the systems underlying most enterprise AI products today. Models exceeding a training compute threshold of 10^25 floating-point operations are classified as posing "systemic risk" and face additional requirements: adversarial testing, incident reporting to the European AI Office, cybersecurity measures, and energy consumption reporting. The compute threshold is specified in Article 51 and Annex XIII. Models below that threshold still face baseline GPAI obligations: technical documentation, copyright compliance policies, and publication of training data summaries. That deadline has also passed.
August 2, 2026. The full high-risk regime under Annexes I and III becomes enforceable. For enterprise AI deployments in regulated domains, this is the operative date. Conformity assessments, technical documentation, human oversight mechanisms, and EU database registration are all required before a high-risk system can be placed on the market or put into service. Ninety days from today.
August 2, 2027. A one-year extension applies specifically to high-risk AI systems embedded in products already covered by existing EU product safety legislation — the Annex I category. Systems already on the market before August 2026 under existing sector-specific rules get this additional runway. New Annex I deployments after August 2026 do not.
Extraterritorial Reach: The Buyer Question
Article 2 defines scope. The regulation applies to providers that place AI systems on the EU market or put them into service in the EU, regardless of where those providers are established. It also applies to deployers of AI systems located in the EU.
Location of the company is not the determining factor. Location of use is.
Procurement conversations often frame this as a binary — does this apply to us? — but the analysis runs at the deployment level. A US-headquartered company running an AI-driven hiring tool for EU-based employees is operating a high-risk system under Annex III, domain 4 (employment and worker management), subject to the full August 2026 regime. The same company's internal IT ticketing AI is almost certainly minimal risk and outside mandatory scope entirely. Scope attaches to specific deployments, not to organizations.
For public sector buyers, the analysis has additional layers. Law enforcement and border control applications sit in Annex III domains 6 and 7. Agencies deploying AI in those contexts face the strictest high-risk requirements, including fundamental rights impact assessments under Article 27 before deployment.
Okta Concept Mapping
The EU AI Act's tier structure most resembles NIST SP 800-63's authenticator assurance levels — AAL1, AAL2, AAL3 — where higher risk demands stronger controls and specific technical requirements. Both frameworks classify by risk, apply proportional obligations, and require documented evidence of compliance at higher levels. The analogy has a load limit, though. In NIST 800-63, the relying party chooses the assurance level based on transaction sensitivity. In the EU AI Act, the tier is assigned by law based on use case category. Deployers do not select their risk tier based on risk appetite; the regulation assigns it, and the prohibited tier has no compliance pathway at all — a concept with no equivalent in assurance level frameworks.
Production note: This article is a preview/demonstration. Every compliance threshold, article citation, and date requires verification against the official text of Regulation (EU) 2024/1689 and a corroborating specialist press source before publication. The compute threshold figure (10^25 FLOPs) and fine structure cited here reflect the author's understanding of the regulation as of May 2026 and must be confirmed against current official guidance, including any implementing acts issued by the European Commission. Triggered accuracy review required upon any major EU AI Act implementation development.