EU AI Act Update: First Look at Draft Guidelines on High-Risk AI Classification

In recent months, the European Commission has continued to advance the EU AI Act framework and develop additional guidance around its implementation. Following progress on the Omnibus legislative package, the Commission has now published draft guidelines on the classification of high-risk AI systems.

While still subject to consultation, the draft guidelines offer companies an early view of how the AI Act’s risk-based framework is expected to work in practice. They also provide a useful indication of how the Commission is likely to approach high-risk classification.

The guidelines set out the Commission’s current interpretive approach and include practical examples illustrating when AI systems are likely to be classified as high-risk. Their purpose is to reduce uncertainty and help organizations assess whether their systems are likely to fall within the high-risk category, including situations where a system may fall outside that classification under the draft guidelines’ filtering logic.

A high-risk classification may carry significant legal, operational, and financial implications for companies. This update provides an initial overview of the key aspects of the draft guidelines and our recommendations for organizations preparing for compliance.

AI Act High-Risk Classification

Under the EU AI Act, an AI system may be classified as high-risk through one of two primary routes:

• The AI system itself is a regulated product, or is intended to be used as a safety component of a regulated product (such as a medical device or vehicle); or
• The AI system falls within one of the specific high-risk use cases defined by the AI Act. 

These use cases span a wide range of sectors, including biometrics, critical infrastructure, education and vocational training, employment, and access to essential private and public services. They also extend to areas more commonly associated with public authorities, such as law enforcement, migration, asylum and border control, and the administration of justice and democratic processes.

This update focuses on the high-risk use cases most relevant to commercial activity.

Key Aspects of Classification

• Intended purpose – The draft guidelines emphasizes that classification depends not only on how an AI system operates technically, but also on its intended purpose. This may be reflected in a provider’s instructions for use, technical documentation, contractual materials, and promotional content. As a result, the way a system is described, positioned, and presented can influence its classification. For example, an AI system used primarily to detect fraud in credit applications may fall outside the high-risk category. However, if the same system is intended to evaluate creditworthiness or generate credit scores, it is more likely to be classified as high-risk.

• Human involvement – The draft clarifies that the presence of a human reviewer does not, by itself, remove an AI system from high-risk classification. If the system’s intended purpose falls within a high-risk use case, human involvement will not change that classification. Human involvement may be relevant only where the AI is genuinely limited to a narrow procedural task, a preparatory function, the enhancement of a previously completed human activity, or the detection of decision-making patterns without replacing meaningful human review. For example, an AI system that summarizes and organizes applicant data for a human underwriter, who then conducts an independent assessment, may fall outside the high-risk category because it performs only a preparatory role. By contrast, the same system is more likely to be considered high-risk if the human reviewer’s role is limited to approving or rejecting the AI output.

• Profiling – The draft guidelines treat profiling as an important factor in determining whether a system falls within the high-risk category. Drawing on existing EU data protection guidance, the Commission adopts a broad interpretation of profiling as the automated processing of personal data for the purpose of evaluating personal aspects of an individual. In practice, systems that analyze, predict, score, rank, or categorize individuals based on personal characteristics will often constitute profiling, even where human oversight is retained. Examples in the guidelines include systems that aggregate applicant data to generate composite risk scores or categories such as “low,” “medium,” or “high risk,” as well as systems that assign students to schools based on factors such as home address, sibling attendance, and school capacity.

• Biometric verification – While the AI Act excludes biometric verification from the high-risk use case relating to remote biometric identification, the draft guidelines provide important clarification regarding the concept of “active involvement.” The key question is whether users are actively and knowingly participating in the identification process. For example, a camera installed at the entrance to a restricted area, where employees actively present themselves for identity verification, would generally be considered biometric verification rather than remote biometric identification. By contrast, cameras installed on the ceiling of a metro station for surveillance purposes are more likely to be treated as remote biometric identification because individuals are not actively involved in the identification process. The guidelines also distinguish between a system used solely to verify a traveler’s identity against the image stored in a passport and a system that compares biometric data against a criminal database.

• Emotions and high-risk classification – The draft guidelines clarify that AI systems intended to identify or infer emotions or intentions from biometric data will generally be classified as high-risk. For example, a crowd-management system that analyzes audience mood through cameras at a concert venue to assess aggression levels is likely to be considered high-risk. By contrast, a driver-monitoring system designed to detect fatigue or loss of concentration is less likely to be treated as high-risk on this basis, because physical conditions such as pain and fatigue are distinguished from emotions under the draft guidelines.

• Limits of the filtering mechanism – The draft guidelines clarify that the filtering mechanism offers less protection from being classified as high-risk than the text of the AI Act alone might initially suggest, but this is not limited only to purely mechanical functions. The mechanism may apply where an AI system is genuinely limited to a narrow procedural task, a preparatory function, the enhancement of a completed human activity, or the detection of decision-making patterns without materially influencing the outcome of a decision. However, systems that perform profiling or generate case-specific recommendations, rankings, credibility assessments, or other evaluative outputs that shape the substance of a decision are less likely to benefit from the filtering mechanism. Conversely, functions such as translation, document sorting, form-field verification, refining human-written performance evaluations without changing their substance, or summarizing medical reports for human review may remain outside the high-risk category, provided they are genuinely supportive and do not replace meaningful human judgment. Organizations relying on this filtering logic to avoid high-risk classification should therefore reassess that assumption carefully.

Practical Implications for Companies

The draft guidelines provide a practical roadmap for how regulators are likely to evaluate specific products, features, and use cases. As such, they represent an important tool for legal and operational risk assessment at the product design stage.

Practical Steps for Companies

Companies should consider taking the following steps:

• Map AI use cases against the categories most likely to be treated as high-risk under the AI Act.
• Review products in light of the interpretive approach and practical examples contained in the draft guidelines.
• Carefully assess whether any aspect of a system may fall outside the high-risk category under the draft guidelines’ filtering logic.
• Conduct a gap analysis against the AI Act’s requirements for systems that are, or may become, high-risk, and evaluate current compliance levels and associated risks.
• Adopt an organizational governance framework covering the development, procurement, deployment, and oversight of AI systems to ensure that relevant uses can be identified and assessed appropriately.
• Implement a phased compliance plan aligned with the relevant application dates, including a clear work plan and internal milestones for execution and rollout.

Although the AI Act entered into force in 2024, recent legislative developments, including the Omnibus package, have resulted in revised application dates for certain high-risk obligations. According to the draft guidelines:

• High-risk use cases under the standalone AI system route: December 2, 2027 (originally August 2, 2026).
• High-risk systems under the product safety route: – August 2, 2028 (originally  August 2, 2027)

Organizations should begin preparing now by familiarizing themselves with the AI Act’s requirements and assess their position within the evolving risk framework so they are prepared when the relevant obligations take effect.

 ***

The firm’s Privacy, Cyber & AI Department is available to assist with EU AI Act readiness, including assessments of the AI Act’s applicability and risk classification criteria, development of internal policies and governance frameworks, and alignment with regulatory requirements in Israel and internationally.

Dr. Avishay Klein is a partner and head of the firm’s Privacy, Cyber and AI Department.

Adv. Masha Yudashkin is an associate in the firm’s Privacy, Cyber and AI Department.