The EU AI Act, Regulation (EU) 2024/1689, sorts AI systems into four tiers by the risk of their use rather than the sophistication of the model: prohibited (Article 5), high-risk (Annex III areas and safety components of regulated products, with narrow Article 6 derogations), limited-risk (transparency duties, such as disclosing that content is AI-generated), and minimal-risk (no specific obligations). What the system does and where it is used decides the tier, so classification follows from declared facts about purpose and deployment context, and can be computed rather than debated.
A regulation organised by risk
The EU AI Act, Regulation (EU) 2024/1689, is the first horizontal law for artificial intelligence, applying across every sector. Its central idea is a risk-based pyramid: an AI system's obligations depend on the risk its use poses, sorted into four tiers. What the system does and where it's used determines the tier, however advanced the underlying model is.
| Tier | What it means |
|---|---|
| Prohibited | Uses banned outright, for example social scoring or certain manipulative or exploitative systems (Article 5). |
| High-risk | Permitted but heavily regulated: the systems in the Annex III areas (and safety components), carrying the bulk of the Act's obligations. |
| Limited-risk | Permitted with transparency duties, for example telling people they're interacting with an AI or that content is AI-generated. |
| Minimal-risk | Everything else: the majority of AI uses, with no specific obligations under the Act. |
What makes a system "high-risk"
High-risk is the tier that matters most, because it's where the obligations concentrate. A system is high-risk broadly if it's used in one of the areas listed in Annex III (think biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration, and the administration of justice), or if it's a safety component of a regulated product.
Article 6 sets out a narrow set of derogations: a system in an Annex III area may not be high-risk if it performs a narrow procedural task, doesn't materially influence a decision, and so on. But those derogations fall away if the system does profiling of individuals: profiling keeps a system high-risk regardless. Classification is therefore a chain of declared facts.
The risk category should follow from facts about the system (its use area, whether it profiles, whether a derogation genuinely applies) rather than from a box someone ticked. A hand-set category that disagrees with those facts is exactly the kind of thing that unravels under scrutiny.
What a high-risk system owes
For a high-risk system, a provider must maintain technical documentation (the Annex IV set: intended purpose, design, data governance, risk management, human oversight, accuracy and robustness), register the system in the EU database (the Annex VIII information), and operate a quality and risk-management system around it. There are also serious-incident obligations: reporting to the authorities on a clock measured in days from the moment you become aware.
Building the register with Qgentic
Qgentic AI Act keeps your AI estate on the record: it takes the declared facts about each system, computes the risk category from the Article 5 / Article 6 / Annex III logic, and refuses a hand-set category that disagrees. It runs the Annex IV documentation completeness gates for high-risk systems, shapes the Annex VIII registration record, and tracks the incident clocks. It also computes the general-purpose AI model duties (Articles 51, 53 and 55) and whether an Article 27 fundamental-rights impact assessment is required. It is prepare-only, on the same approval gate and audit trail as every Qgentic platform. For UK organisations there's a separate rulebook, covered in UK AI governance.
This is engineering-and-compliance orientation, not legal advice, and the AI Act's obligations phase in over time. Article references are for orientation; confirm classification and duties for your specific systems with counsel.
Qgentic AI Act computes each system's risk category from declared facts, and refuses a hand-set category that disagrees.
See the AI Act platform Talk to us