The European Commission Recently Published Its Guidance on High-Risk AI Systems Classification.

28 mei 2026, On May 19, 2026, the European Commission (the “Commission”) released its long-anticipated Draft Guidelines on the Classification of High-Risk AI Systems (accessible here) under Article 6 of the Regulation (EU) 2024/1689 (the “EU AI Act”).  

For any organization developing, deploying, or using AI systems in Europe, this is the guidance you have probably been waiting for; and it may have significant implications for your compliance roadmap.  

Why These Guidelines Matter 

The high-risk classification is one of the most consequential determinations under the EU AI Act. If your AI system is classified as high-risk, you may face the regulation’s most demanding compliance obligations; such as risk and quality management systems, strict data governance, technical documentation, conformity assessments, registration in the EU database, post-market monitoring, and more.  

Until now, organizations have been grappling with uncertainty about how to interpret the classification criteria in the EU AI Act. These draft guidelines, though not legally binding (yet), apparently reflect the Commission’s official interpretation. They provide around 150 pages of interpretive guidance and practical examples across the three draft documents, giving real-world clarity for the first time at this level of detail. 

The Two Routes to High-Risk Classification 

The guidelines are structured around the two independent pathways through which an AI system can be classified as high-risk: 

• Route 1: Product Safety (Article 6.1 and Annex I of the EU AI Act) 

An AI system is high-risk if it is a safety component of a product or is itself a product covered by EU harmonization legislation listed in Annex I, and that product must undergo a third-party conformity assessment. This includes, for example, products covered by regimes for machinery, toys, radio equipment, medical devices, automotive systems, and aviation. 

A key clarification is that how a feature is commercially positioned does not control its classification. The more important question is the system’s intended purpose and realistic failure mode. The guidelines give a vivid example: an AI system marketed as a “combustion-efficiency optimiser” in a household gas appliance may still be classified as a safety component if its failure could lead to carbon monoxide formation, explosion, or fire. By contrast, a heating-schedule optimiser whose failure merely causes discomfort, or higher bills would not typically be classified as a safety component. 

• Route 2: Specific Use Cases (Article 6.2 and Annex III of the EU AI Act) 

An AI system is also high-risk if its intended purpose falls into one of the use cases listed across the eight areas in Annex III: 

• Biometrics: remote biometric identification, biometric categorization, and emotion recognition  

• Critical infrastructure: AI systems used as safety components in the management and operation of critical digital infrastructure, road traffic, or the supply of water, gas, heating, or electricity  

• Education and vocational training: admissions decisions, grading, assessment, and proctoring  

• Employment and worker management: recruitment, screening, performance evaluation, promotion, termination, and task allocation  

• Access to essential private and public services: social welfare, creditworthiness assessments, insurance risk assessment and pricing, and emergency call triage  

• Law enforcement: risk assessments, deception detection, evidence evaluation, and profiling  

• Migration, asylum, and border control: visa and asylum risk assessments, document verification, and irregular migration risk assessments  
• Administration of justice and democratic processes: assisting judicial authorities and influencing voting behaviour or election outcomes 

Each route is independent. Either one can trigger the full Chapter III high-risk obligations of the EU AI Act. 

Key Takeaways from the Draft Guidelines 

1. Failure modes matter for product-safety systems 

Under Annex I, an AI system is not high-risk simply because it is part of a regulated product. The key question is whether it is a safety component: either because it performs a safety function, or because its failure or malfunction could endanger health, safety, or property. In other words, an “efficiency” or “optimisation” tool may still be high-risk if its failure creates real safety risks. 

2. Combined AI systems may be assessed as one 

Where several AI components operate together and their combined output materially influences a high-risk decision, the configuration may be treated as a single AI system. This is particularly relevant for agentic AI systems, orchestrators, sub-agents, retrieval tools, scoring models, and decision-support modules.

3. Marketing claims matter 

Contractual exclusions or terms of service saying “no high-risk use” will not be enough if the product’s marketing, documentation, examples, or sales materials effectively promote such uses. The Commission will look at the full picture, meaning marketing and product positioning are now part of the compliance surface. 

4. Human involvement is not a safe harbour 

A human in the loop does not automatically prevent high-risk classification. The key question is whether the AI system’s intended purpose and output materially influence a high-risk decision. 

5. The Article 6.3 exception is narrow 

Annex III systems may avoid high-risk classification only if they do not pose a significant risk and fit one of the four Article 6.3 conditions: a narrow procedural task, improvement of a completed human activity, detection of decision-making patterns without influencing the assessment, or a preparatory task. If the system performs profiling of natural persons, the exception cannot apply. Providers relying on the exception must, of course, document and register their assessment. 

6. Compliance timelines have shifted 

Following the political agreement on the AI Omnibus, the relevant high-risk AI compliance dates are now expected to be: 

• August 2, 2028, for Annex I product-safety systems 

• December 2, 2027, for Annex III high-risk systems 

It would be wise to begin classification assessments now, rather than waiting for the final guidelines. 

What Should Organizations Do Now? 

Organizations should definitely begin by reviewing their AI portfolio against both high-risk classification routes: Annex I product-safety legislation and Annex III use cases. An AI system may fall within either route, depending on its intended purpose and regulatory context. Providers relying on the Article 6.3 exception should carefully document their assessment before placing the AI system on the market or putting it into service. 

Organizations should also align product marketing, sales materials, instructions for use, and terms of service with their classification position. Contradictions between contractual restrictions and promotional materials could undermine a compliance argument and should be prevented. 

Furthermore, for agentic and composite AI systems, organizations should assess the combined system as a whole where multiple components work together to serve a high-risk intended purpose. Classifying each module in isolation may not be enough.  

Finally, organizations affected by the draft guidance may also consider participating in the Commission’s consultation, open until June 23, 2026, via EU Survey (accessible here). The AI Act Single Information Platform provides the guidelines in a user-friendly format, where organisations can search and review only the areas and use cases relevant to them. 

Looking Ahead 

These draft guidelines are the Commission’s clearest statement yet on the EU AI Act’s high-risk classification framework. Although not legally binding, they can clearly serve as a practical baseline for compliance planning. Organizations should begin classification assessments now, align product and marketing teams, and engage with the consultation process while they can. 

Have questions about how the high-risk AI classification applies to your specific AI systems? Reach out to us for a tailored assessment.