Guidelines on how to comply with high-risk AI finally published

On 19 May the Commission published draft guidelines on the classification of high-risk AI systems. With the Omnibus package promising organisations more time to comply with the high-risk rules (see our blog), these draft guidelines now start to provide some of the much-needed practical guidance to help providers and deployers alike to understand when the high risk rules are triggered.

What do the Guidelines cover?

The guidelines are designed to help providers and deployers of AI systems decide whether an AI system is high-risk. They discuss the two types of high-risk AI (namely AI systems satisfying the Annex I or Annex III criteria) and provide information and practical examples for both areas. Split across three documents, the draft guidelines cover:

• General principles for the classification of high-risk AI systems. Key takeaways include:
  • Providers must assess whether their AI systems are high-risk and describe their use: Perhaps unsurprisingly, the guidelines confirm that providers must carry out this self-assessment. They discuss how the AI must first satisfy the definition of AI systems, referencing the Commission’s 2024 Guidelines on this, and also highlight the importance of the intended purpose(s) or use of the AI system. Providers must also clearly describe the system’s envisaged use (or uses) in its materials, as this will impact whether or not the AI is high-risk.
  • General purpose AI may be caught, and a simple exclusion in the terms of service may not prevent this: Those providing broadly applicable AI systems may want to revisit their sales materials and terms of use given that:
    • the way the provider describes the intended use of the system matters for systems with multiple purposes and for GPAI systems. If the relevant provider materials (the instructions for use, contractual arrangements or terms of service, usage policy, sales/promotional materials, or the technical documentation) present the AI system as ‘broadly applicable across a generality of contexts and functions’, and do not consistently exclude high-risk uses or otherwise limit its application, the system’s intended purpose will be deemed to include high-risk use cases and it will therefore qualify as high-risk. The guidance goes on to add that this will particularly be the case where high-risk uses are feasible and reasonably foreseeable in light of the system’s functionalities and capabilities;
    • any exclusions or limitations on the use must be “clearly, concretely and coherently” described across all materials; and
    • simply excluding high-risk use in the terms of service will not be sufficient where the provider’s product positioning effectively provides or promotes high-risk use cases.
  • Timings and exemption: The guidelines discuss the timeframes for the high-risk rules, which will be pushed back following the AI Omnibus. They will now apply from 2 December 2027 for Annex III high-risk AI and 2 August 2028 for Annex I high-risk AI. There is an exemption available for high-risk systems that are already on the market, provided they have not been subject to significant design changes. This was not, however, discussed as part of the Omnibus and the provisions (and guidelines) therefore still reference 2 August 2026 as the relevant date for the system to be placed on the market (which was the original application date for the Annex III high-risk systems). In other words, the grandfathering provisions for those systems don't extend to protect systems put on the market between the original 2026 implementation date and the revised 2027 implementation date - which dilutes the benefit of the delay for businesses planning to launch new systems during that period.
  • Reviews expected: While the guidelines provide some additional detail on how the high-risk rules will be interpreted by those enforcing them, further guidance is expected. The high-risk AI use cases themselves may also be reviewed and updated – for example, the AI Act provides a mechanism to future proof the list of high-risk use cases in Annex III as new risks emerge and the market develops. It is therefore important for potential providers and deployers to keep this area under review.
• The scope of the Annex I high-risk AI systems: Annex I covers AI systems which are products regulated under the EU’s product safety legislation or are used as safety components in such products. Given the wide range of products covered (lifts, toys, medical devices etc.) the guidelines discuss the main elements of the assessment, including the classification rationale. They also consider the changes introduced by the AI Omnibus, including the change to the definition of a safety component.
• The scope of the Annex III high risk AI systems: Annex III contains a list of AI systems that can significantly affect people's health, safety, or fundamental rights. It covers use cases such as recruitment and credit scoring which could impact a wide range of organisations. The guidelines provide details and examples to show when an AI system will, and will not, fall within each use case and also look at some horizontal issues applicable across all of these use cases, such as the role of human involvement. In our next blog, we will examine the Annex I and Annex III guidelines.

Next steps:

The guidelines are part of a targeted consultation and feedback can be submitted via a dedicated portal until 23 June 2026. There is no formal timeline for their adoption, although the plan is for this to be agreed in advance of the date on which the respective high-risk rules take effect.

We regularly publish blogs on AI developments. Please keep an eye out for our upcoming blogs on the Annex I and III guidance contained in these guidelines, and the ICO’s blog on AI-powered cyber threats.