The European Commission has published a draft of guidelines on the classification of high-risk AI systems pursuant to Art. 6 of the AI Regulation (Regulation (EU) 2024/1689, hereinafter: the AI Act) for public consultation. The guidelines specify the conditions under which an AI system is to be classified as “high-risk” and thus subject to the extensive obligations under Chapter III of the AI Act. They are addressed primarily to providers and deployers of AI systems as well as to the competent market surveillance authorities.
Classification as a high-risk AI system is of central practical importance: it triggers a comprehensive set of obligations ranging from risk inventory and technical documentation to conformity assessment procedures, registration requirements and ongoing monitoring obligations. The following article outlines the key contents of the draft guidelines and their significance for compliance practice.
- Background: Art. 6 AI Act and the High-Risk Classification
The AI Act entered into force on 1 August 2024 and follows a risk-based approach: the requirements applicable to an AI system depend on the level of risk it poses. The most extensive obligations apply to so-called high-risk AI systems within the meaning of Art. 6 AI Act. The high-risk classification is therefore the pivotal determination upon which it depends whether a provider or deployer must comply with the full requirements regime of Chapter III of the AI Act – including a risk management system, technical documentation, conformity assessment, logging, transparency and human oversight.
Art. 6 AI Act provides for two independent classification pathways. The first pathway under Art. 6(1) read in conjunction with Annex I to the AI Act concerns AI systems that are embedded in regulated products as safety components or that themselves constitute such products – for example in the fields of machinery, medical devices or motor vehicles. The second pathway under Art. 6(2) read in conjunction with Annex III to the AI Act covers stand-alone AI systems deployed in particularly sensitive fields of application, such as biometrics, law enforcement or human resources management. The guidelines to be issued by the Commission pursuant to Art. 6(5) AI Act are intended to ensure uniform interpretation and application for both pathways. They are not legally binding – an authoritative interpretation falls exclusively within the jurisdiction of the Court of Justice of the European Union.
2. Classification Pathway I: AI Integrated into Products under Art. 6(1) read with Annex I AI Act
Art. 6(1) AI Act covers AI systems that, as safety components of a product or as products themselves, fall within the scope of one of the harmonisation acts listed in Annex I to the AI Act. These include, inter alia, the Machinery Regulation (Regulation (EU) 2023/1230), the Medical Devices Regulation, the Toys Safety Regulation and the Type Approval Regulation for motor vehicles. According to the draft guidelines, the high-risk classification under this pathway requires two cumulative conditions to be satisfied: the AI system must either itself be a product falling within the scope of the relevant harmonisation legislation or constitute a safety component of such a product; and the product in question (or the AI system itself) must be subject to third-party conformity assessment under the applicable harmonisation act.
Importantly, the Commission makes clear that not every AI system integrated into a regulated product is automatically classified as high-risk. The decisive criterion is always whether the system possesses safety-functional relevance within the meaning of Art. 3(14) AI Act. The AI Act does not alter the risk profile of a product – rather, it builds upon the sectoral risk classification already established under the applicable harmonisation legislation.
3. The Concept of “Safety Component”
Art. 3(14) AI Act defines a safety component as a component of a product that performs a safety function for that product or the failure or malfunctioning of which endangers the health and safety of persons or property. The draft guidelines distinguish two alternative scenarios in this regard:
Under the first scenario (intent-based), a safety component exists where the AI system’s intended purpose – as determined by the provider – is directed at preventing or mitigating safety risks, for example through monitoring and detection of hazardous conditions, initiation of protective measures or supervision of a safety-critical component. The determining factor is the intended purpose as documented by the provider in the instructions for use, technical documentation and promotional materials. Functions serving purely optimisation, efficiency or comfort purposes do not satisfy this criterion.
Under the second scenario (consequence-based), an AI system is to be classified as a safety component even where its primary intended purpose is not safety-related, provided that a failure or malfunction could lead to endangerment of persons or property. The draft guidelines cite as an illustrative example an AI system designed to optimise combustion efficiency in gas appliances: where the product is designed such that a system failure could lead to carbon monoxide formation, explosion or fire, the AI system qualifies as a safety component – irrespective of the fact that energy efficiency is its primary intended purpose. By contrast, an AI system that merely optimises heating schedules based on household routines, the failure of which would at most result in higher energy costs or inconvenience, does not qualify as a safety component.
4. Classification Pathway II: Stand-Alone High-Risk AI under Art. 6(2) read with Annex III AI Act
Art. 6(2) AI Act covers stand-alone AI systems whose intended purpose falls within one of the eight areas listed in Annex III to the AI Act: biometrics; critical infrastructure; education and vocational training; employment, workers’ management and access to self-employment; access to and enjoyment of essential private and public services and benefits; law enforcement; migration, asylum and border control management; and administration of justice and democratic processes. The list of relevant use cases in Annex III is exhaustive and may only be amended by delegated acts of the Commission. The sole determining factor for classification is the intended purpose as defined by the provider – not the actual deployment scenario.
The draft guidelines clarify that human oversight during deployment does not, as a rule, eliminate the high-risk classification – it is rather a prerequisite for compliance with the requirements applicable to such systems under Art. 14 AI Act. Providers who demonstrably market their systems for high-risk use cases without clearly restricting the stated intended purpose must expect a corresponding classification. Mere liability disclaimers in terms of service are insufficient to avoid high-risk classification where the system, by virtue of its functionalities and market positioning, is evidently designed for such applications. This is of particular relevance for general-purpose AI systems whose terms of use do not clearly and consistently exclude high-risk deployment scenarios.
5. The “Filter” Mechanism under Art. 6(3) AI Act
Art. 6(3) AI Act provides for a so-called filter mechanism, enabling providers to exempt an AI system that would otherwise fall within the scope of Annex III from high-risk classification, provided that the system is exclusively directed at one of the following limited functions: (a) performance of a narrow procedural task; (b) improvement of a previously completed human activity; (c) detection of decision-making patterns or deviations therefrom, without replacing or influencing a human assessment; or (d) performance of a preparatory task relevant to the use cases listed in Annex III. The filter mechanism applies exclusively to the classification pathway under Art. 6(2) AI Act – not to Annex I systems.
The draft guidelines emphasise that the filter mechanism is to be interpreted narrowly. It does not apply where the AI system performs profiling within the meaning of the General Data Protection Regulation. It likewise does not apply where the system forms part of a complex system whose overall configuration materially influences the outcome of a decision in a high-risk use case – for example in agent-based AI architectures involving the interaction of multiple components. Providers invoking the filter mechanism bear the burden of proof for the relevant conditions; the corresponding self-assessment must be documented and substantiated.
6. Dates of Application: Current Position Following the AI Omnibus
The draft guidelines reflect the revised dates of application introduced by the AI Omnibus (political agreement of 7 May 2026). Under the original schedule, the high-risk obligations were to apply to Annex III systems as of 2 August 2026 and to Annex I systems as of 2 August 2027. These dates have been deferred: for AI systems under Art. 6(2) read with Annex III AI Act, the high-risk obligations now apply as of 2 December 2027. For AI systems under Art. 6(1) read with Annex I AI Act, the applicable date is 2 August 2028.
Nonetheless, the draft guidelines are already of considerable practical relevance: the question of whether an AI system is to be classified as high-risk is determined by the intended purpose at the time of placing on the market – not from the date of application. Providers currently developing or placing AI systems on the market must therefore carry out a classification analysis now, in order to avoid costly retrofit obligations and liability exposure.
Practical Note
The draft guidelines already provide valuable guidance for the classification analysis under Art. 6 AI Act. Businesses should actively monitor the consultation process and consider submitting their own responses, as the guidelines, once adopted, will be authoritative for regulatory market surveillance and conformity assessment. Above all, the additional time afforded by the AI Omnibus must not be mistaken for licence to delay: high-risk classification is a preliminary assessment that must be carried out at an early stage – ideally during the design and development phase of an AI system.
Recommended Actions
- Conduct a classification analysis: Systematically assess all AI systems you develop or deploy to determine whether they fall within Art. 6(1) or Art. 6(2) AI Act. Draw on the draft guidelines as an interpretive tool – in particular the definitions relating to safety components and the illustrative examples for both classification pathways.
- Document the intended purpose consistently: The intended purpose of the AI system is the central classification criterion. Ensure that instructions for use, promotional materials, contractual terms and technical documentation contain a coherent, complete and unambiguous description of the intended purpose and do not imply unintended high-risk use cases.
- Evaluate the filter mechanism: For AI systems with a connection to Annex III, assess whether the exception under Art. 6(3) AI Act applies. The exception is to be interpreted narrowly, and the self-assessment must be documented and substantiated. Systems with a profiling function or forming part of agent-based architectures cannot benefit from this exception.
- Follow the consultation process: The draft guidelines are open for consultation. Make use of the opportunity to submit comments, particularly where your products or systems fall within the categories addressed in the draft. The final version of the guidelines will significantly shape the regulatory approach to market surveillance and conformity assessment.
Contact:
Jens Borchardt