Under the EU AI Act (Regulation 2024/1689), not all artificial intelligence is regulated the same way. Most AI — spam filters, recommendation engines, video game AI — faces little or no legal obligation. But a subset of AI applications, classified as “high-risk,” must meet strict requirements before they can enter the European market. Understanding this classification is the key to understanding what the Act actually demands from businesses.
The EU AI Act’s four risk tiers
The Act divides AI systems into four broad tiers based on the potential harm they can cause:
- Prohibited AI — applications banned outright, such as mass real-time biometric surveillance in public spaces or systems that manipulate people through subliminal techniques. This ban took effect on 2 February 2025.
- High-risk AI — systems with meaningful potential to harm health, safety, or fundamental rights. These face the Act’s strictest compliance requirements.
- General-purpose AI (GPAI) — large foundation models like those powering ChatGPT-style services, subject to transparency and documentation rules.
- Minimal or no risk — the vast majority of AI applications, covering video games, customer chatbots, and most productivity tools. No specific obligations apply.
Two paths to high-risk status
An AI system qualifies as high-risk through one of two routes:
Annex II — product-embedded AI. AI safety components built into physical products already governed by EU product-safety law: medical devices, industrial machinery, toys, lifts, and civil aviation equipment, among others. These follow the conformity procedures of the relevant product regulation.
Annex III — standalone use cases. A list of eight sectors where AI has direct power over important decisions in people’s lives:
- Biometrics — remote identification, categorisation by protected attributes, emotion recognition
- Critical infrastructure — AI managing electricity grids, water systems, or road traffic
- Education and training — systems deciding access to schools, grading students, or detecting cheating
- Employment — tools that screen job applicants, manage performance, allocate tasks, or trigger dismissal
- Essential public services — AI assessing eligibility for social benefits, healthcare, or loans; insurance pricing
- Law enforcement — risk-profiling for crime or reoffending, polygraph-equivalent tools, evidence analysis
- Migration and asylum — border security risk assessments, asylum application processing
- Justice and democratic processes — AI assisting courts, or systems that can influence electoral behaviour
If an AI system does not fall into any of these categories, it is almost certainly in the minimal-risk tier.
What high-risk AI providers must do
Providers of high-risk AI systems face a demanding compliance checklist before they can lawfully deploy in the EU:
- Technical documentation. A detailed technical file describing the system’s design, training data, and intended use must be maintained throughout the product’s life.
- Conformity assessment. Providers must demonstrate compliance — through internal checks or, for biometric systems, an accredited third-party notified body. Passing the assessment earns a CE mark and an EU Declaration of Conformity.
- EU database registration. Standalone high-risk systems must be listed in a public EU database before deployment.
- Human oversight. Systems must be designed so that a human can monitor, interpret, and override their outputs. Over-reliance on AI decisions is explicitly not permitted.
- Automatic logging. Events and outputs must be recorded automatically so that incidents can be investigated after the fact.
- Post-market monitoring. Providers must watch for problems after deployment and report serious incidents to national authorities.
When compliance is required
The EU AI Act was published on 1 August 2024. Implementation is phased:
| Date | Requirement |
|---|---|
| 2 February 2025 | Prohibited AI applications banned |
| 2 August 2025 | GPAI (general-purpose AI model) rules apply |
| 2 December 2027 | Standalone high-risk AI (Annex III) must comply |
| 2 August 2028 | Product-embedded high-risk AI (Annex II) must comply |
The December 2027 deadline for standalone high-risk systems was extended by 16 months from the original August 2026 date, following the EU’s Digital Omnibus legislative package. The extension was granted because harmonised technical standards and national supervisory bodies were not yet in place — businesses therefore have more time to prepare, but the requirements themselves have not changed.
In the news
The EU recently confirmed this timeline change. Read our brief: EU Finalizes AI Act Overhaul, Delaying High-Risk Compliance by 16 Months.
For a broader overview of the Act and what it means for your organisation, see: What Is the EU AI Act — and What Do Businesses Need to Know?.
Frequently asked questions
Does the EU AI Act apply to companies outside the EU?
Yes. If an AI system’s output is used inside the EU — or if the users are EU-based — the Act applies regardless of where the provider is headquartered.
Can a single AI product fall into multiple high-risk categories?
Yes. An AI system used for both recruiting and ongoing performance monitoring, for instance, falls under the employment category and all obligations that entails.
What are the penalties for non-compliance?
Fines can reach €15 million or 3% of global annual turnover (whichever is higher) for violations of high-risk obligations, and up to €35 million or 7% of turnover for deploying prohibited AI.
Is the list of high-risk use cases fixed?
Not entirely. The European Commission can amend Annex III by delegated act if new AI applications emerge that pose equivalent risks to those already listed.