EU Commission drafts guidelines on classifying high-risk systems under the AI Act
Draft guidelines from the European Commission aim to assist companies in determining whether their products, including medical devices that incorporate artificial intelligence (AI) components, are classified as high-risk under the EU's Artificial Intelligence Act.
For devices that are subject to the AI Act, these requirements will be layered on top of existing obligations under the Medical Device Regulation (MDR) and In Vitro Diagnostic Regulation (IVDR) rather than replacing them.
Nils Lölfing, an attorney specializing in AI at the law firm Bird & Bird in Germany, recently discussed the implications of the AI Act and its impact on medical devices with Focus. He also highlighted the ongoing efforts to prevent devices from being regulated under multiple frameworks.
The AI Act creates a legal framework to encourage the development of "human-centric and trustworthy" artificial intelligence for medical devices, including in vitro diagnostics (IVDs) and other products. (RELATED: EU publishes regulation governing use of AI in medical devices and IVDs, Regulatory Focus 15 July 2024)
The AI Act will take effect on 2 December 2027 for stand-alone high-risk AI systems and 2 August 2028 for high-risk AI systems embedded in products.
According to a Commission announcement, “the Guidelines set out the Commission’s interpretation of certain concepts that are relevant for classification purposes and, in accordance with Article 6(5) AI Act, contain practical examples of AI systems that should or should not be classified as high-risk. The examples listed in these Guidelines strive to cover all areas and use cases, but they are not to be considered as exhaustive and may be updated over time.”
The Commission published three different guidelines to shed some light on its interpretation of what is meant by high-risk products. These include a set of general principles on determining whether an AI system is high risk; the second guidance discusses Annex 1 of the act, and the third addresses Annex III.
The guidelines clarify the classification process under Article 6, but they do not specify device types, which is why they appear as broad, said Lölfing.
The principles guideline states that “the intended purpose of an AI system plays an important role in its classification as high-risk, as well as for assessing compliance with the requirements for high-risk AI systems laid down in Section 2 of Chapter III AI Act.”
It adds that “Under Article 6 AI Act, an AI system shall be considered high-risk in two scenarios. First, if the system is intended to be used as a safety component of a product, or the AI system itself is a product, covered by the Union harmonisation legislation listed in Annex I, and the product whose safety component is the AI system or the AI system itself is required to undergo a third-party conformity assessment, the system will be classified as high-risk pursuant to Article 6(1) AI Act.”
The guidance on Annex III of the EU AI Act, which is 148 pages, discusses high-risk classifications according to Article 6(1) and Annex I of the AI Act, and Article 6(2).
It specifies that certain AI systems are considered "high-risk" due to their potential impact on human health, safety, and fundamental rights. These systems are subject to strict compliance obligations, which include independent conformity assessments, risk management systems, and registration requirements.
It states that “Emergency healthcare patient triage systems may qualify as medical devices in so far as they fulfil the definition of a medical device in Article 2(1) of the Medical Devices Regulation. If such systems meet the conditions in Article 6(1) AI Act, they will be classified as high-risk pursuant to that provision and Annex I AI Act and require streamlined and consistent compliance with sectoral laws.”
The guidance on Annex I of the AI Act notes that while the AI Act does not provide a list of safety components or safety functions, based on the definition of safety component in Article 3(14) AI Act and the clarifications provided in Recitals 47 and 48 AI Act, “the concept of ‘safety function’ may include, in particular, the following preventive or mitigation functionalities: Preventive functions: – monitoring and detection of situations which may lead to physical harm or damage to people or property (e.g. AI system detecting abnormal system behaviour); where failure to perform such actions may lead to physical harm to people or property (e.g. an AI system detecting whether safety-related parts, or parts whose failure may lead to harm to people or property , are worn down and may need replacement or maintenance).”
The guidance also outlines the types of products covered under the AI Act. These include safety components of machinery, toys, lifts, equipment, and protective systems designed for use in potentially explosive atmospheres. Additionally, the Act covers radio equipment, pressure equipment, recreational craft, cableway installations, appliances that burn gaseous fuels, medical devices, in vitro diagnostic medical devices, and products in the automotive and aviation sectors.
Efforts underway to stop dual regulations
Lölfing discussed the interplay of the AI Act with the MDR and the IVDR in an email shared with Focus.
He said that “the starting point is that AI in medical devices already falls under the EU's device rules, the MDR and IVDR. The AI Act adds a second layer on top, but only for what it calls high-risk systems. An AI medical device is high-risk where it is either a device in its own right or a safety component of one, and where that device needs a notified body (an independent assessor) to sign it off. In practice that is roughly Class IIa and above under the MDR, and most IVDs.”
He noted two examples to clarify the type of device that would be subject to the AI Act. This would include pure AI software, such as an algorithm that analyzes scans to flag or diagnose diseases, where the software itself functions as the device; and second, AI integrated into hardware, like AI that regulates dosing in an infusion pump or controls a surgical robot, serving as a safety component. Simple, self-certified Class I devices fall outside the high-risk category.
Lölfing observed that two reforms are currently underway, attempting to prevent these devices from being regulated twice.
This includes the Digital Omnibus on AI, which includes provisions that keep medical devices under the AI Act while postponing the high-risk obligations until a final deadline of August 2028. Additionally, it allows the Commission to pass follow-up legislation, or delegated acts, that can exempt certain AI Act high-risk requirements if the MDR or IVDR already mandates the same standards. This means that within the AI Act, there is now a mechanism to eliminate redundant requirements. While an agreement has been reached, it has not yet been formally adopted into law and awaits final legal review.
Another proposal is incorporated in the planned revision of the MDR and IVDR issued in December 2025. “This plan would make the device rules the main rulebook, so AI requirements are handled within the normal device approval rather than in a separate AI Act process, with the AI-specific obligations likely written into rules sitting under the MDR itself. This mirrors how the EU already treats AI in cars and in machinery and robotics, where the sector law, not the AI Act directly, carries the AI requirements.”
Lölfing said that under these proposals, “this means medical AI stays regulated, but companies get more time, and the focus is shifting to the device rules.”
He added that “the sensible approach is to evidence the AI-specific points, mainly data governance, robustness and human oversight, through MDR and IVDR documentation rather than as a separate AI Act exercise. The detail will depend on the Commission's follow-up acts and on how the MDR revision lands.”
According to a 2024 blog post from the consultancy IQVIA, a large segment of AI’s use in healthcare would be classified as ‘high-risk’ under the Act and thus subject to multiple requirements.
The deadline for submitting comments is 23 June. Comments can be submitted here.
