After discussing liability for misleading statements made by an AI chatbot in our last AI Flash, we would like to continue providing you with regular legal insights on AI-related developments.
In today’s AI Flash, we would like to take a closer look at high-risk AI systems in light of recent developments at the European level. On 19 May 2026, the European Commission launched a consultation process and published three draft guidelines on the classification of high-risk AI systems. For the first time, these drafts provide detailed clarification on what is arguably the most consequential question under the AI Act (AIA): When does an AI system qualify as high-risk—and when does it not?
Given the complexity of the topic, particularly considering the different categories of high-risk AI systems, we will divide our comments on the draft guidelines into several AI Flash publications. This contribution is intended to provide an initial overview, enabling us to address specific aspects in greater detail in future editions.
Timeline: What the AI Omnibus Has Changed
Before addressing the substantive classification question, it is important to understand the applicable timeline. With the provisional trilogue agreement on the so-called “Digital Omnibus on AI” of 7 May 2026, the European Parliament and the Council substantially postponed the application deadlines for high-risk AI systems:
- AI systems serving as safety components in regulated products under Annex I of the AI Act (including AI used in medical devices, lifts, and toys)
Postponement: 2 August 2027 → 2 August 2028 (+12 months) - Standalone high-risk AI systems under Annex III of the AI Act (including AI used in human resources and critical infrastructure)
Postponement: 2 August 2026 → 2 December 2027 (+16 months)
Although the trilogue agreement has not yet been formally adopted and the original deadlines technically remain applicable until adoption, formal enactment within the relevant timeframe is considered highly likely.
Against this backdrop, the newly published draft guidelines constitute an important tool for companies. They provide an opportunity to use the additional preparation time in a structured and substantive manner and, for the first time on an official basis, assess whether and to what extent AI systems may be subject to high-risk classification in the future.
The Two Classification Pathways under Article 6 AI Act
Article 6 AI Act provides for two independent pathways leading to a high-risk classification.
Annex I (Product Safety)
An AI system qualifies as high-risk if it is itself a regulated product or serves as a safety component of such a product falling within specific EU harmonisation legislation and the product is subject to a third-party conformity assessment.
This pathway is particularly relevant for AI used in medical devices, toys, vehicles, and lifts.
The Commission appears to interpret the key concept of a “safety component” rather broadly. Even a system not expressly intended to perform a safety function may qualify as a safety component if its failure or malfunction could create a health or safety risk. Purely efficiency-, comfort-, or performance-related functions without safety relevance do not fall within this category.
Annex III (Use Case-Based Classification)
For most businesses, the practically more relevant pathway is Annex III of the AI Act. Here, the legislator identifies eight areas of application (with further subdivisions) in which the use of AI systems is generally considered high-risk:
- Biometrics
- Critical infrastructure
- Education and vocational training
- Employment and worker management
- Access to essential private and public services
- Law enforcement
- Migration, asylum and border control
- Administration of justice and democratic processes
This category will be the primary focus of the present contribution.
Intended Purpose as the Key to Classification
A central theme throughout the draft guidelines is the concept of the intended purpose of an AI system.
The decisive factor is how the intended purpose is presented externally in instructions for use, technical documentation, marketing materials, and other statements made by the provider.
The Commission makes it clear that providers cannot avoid a high-risk classification through disclaimers or exclusions in their terms of use if the system’s design, marketing, or overall positioning suggests a high-risk use case. Any limitation of the intended purpose must be communicated clearly, specifically, and consistently across all materials.
HR Example: Where Does the High-Risk Zone Begin?
Annex III, Section 4 of the AI Act expressly covers AI systems used in the fields of employment, worker management, and access to self-employment.
Specifically, this includes:
- “AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, analyse or filter applications, and evaluate candidates.”
- “AI systems intended to be used for decisions affecting the terms of employment relationships, promotions and termination of contractual employment relationships, for task allocation based on individual behaviour or personal traits or characteristics, or for monitoring and evaluating the performance and behaviour of persons in such relationships.”
According to the draft guidelines, the specific use case and, subsequently, the degree to which the system influences a human decision are of particular importance.
The mere existence of human oversight is not sufficient to avoid classification as a high-risk AI system. Given the significant practical relevance of HR applications, we intend to address this topic in greater detail in a dedicated future AI Flash.
The Article 6(3) Filter: An Exception for Annex III Systems
For AI systems that fall within one of the Annex III areas but whose functions are comparatively limited, Article 6(3) AI Act provides a filter that may prevent classification as a high-risk system.
This filter applies exclusively to Annex III systems and not to product safety-related AI systems under Annex I.
The Commission explicitly emphasises that the conditions for applying the filter must be interpreted narrowly, as Article 6(3) constitutes an exception to rules primarily designed to protect fundamental rights.
The filter applies only where one of the following criteria is met:
- Narrow Procedural Task: The system performs only a clearly limited procedural task, such as sorting application documents into predefined categories. Systems that evaluate data or make qualitative assessments do not fall within this category.
- Improvement of a Completed Human Activity: The system improves the result of a previously completed human action but neither replaces nor revises it. Example: Highlighting inconsistencies in a decision already made by a human.
- Detection of Decision-Making Patterns without Influence: The system identifies patterns or deviations in decision-making processes but neither influences nor replaces human judgment.
- Preparatory Task with Limited Influence on Outcomes: The system prepares a decision without substantively steering it, for example by providing relevant guidelines without directing the outcome of the decision.
The operational key question therefore becomes: Does the system merely structure a process, or does it influence human judgment?
Systems that sort, convert, or detect duplicates may fall within the filter. Systems that rank, assess, score, or make clear recommendations will generally be classified as high-risk.
Furthermore, the filter is entirely unavailable where the AI system performs profiling, i.e., automated processing of personal data consisting of the use of personal data to evaluate certain personal aspects relating to a natural person.
This limitation is highly relevant in practice because many AI tools used in HR inherently involve profiling functionalities. Thorough assessment is therefore essential.
Providers seeking to rely on the filter must, prior to placing the system on the market, prepare a written self-assessment documenting:
- the intended purpose;
- the basis for the high-risk classification under Article 6(2) AI Act;
- the applicable filter condition; and
- the reasons why no profiling is involved.
The system must subsequently be registered in an EU database.
A Potential Pitfall for AI Deployers
As a first step, it is the provider of the AI system who is responsible for defining its intended purpose and, consequently, determining its risk classification.
From the deployer’s perspective, however, Article 25(1) AI Act is particularly important. It provides, among other things, that a deployer may effectively become a provider where the intended purpose of an AI system that was not originally classified as high-risk and has already been placed on the market or put into service is modified in such a way that the AI system subsequently becomes a high-risk AI system within the meaning of Article 6 AI Act.
Put simply:
A company that uses a generative AI system—which will often not qualify as high-risk—for candidate selection purposes may alter the system’s risk classification and simultaneously assume a different regulatory role.
Practical Recommendation: Participate in the Consultation
The draft guidelines are open for public consultation until 23 June 2026. Formal adoption by the Commission is expected thereafter, although no specific timeline has yet been announced. Under the current framework, the guidelines are not legally binding; authoritative interpretation remains the responsibility of the Court of Justice of the European Union (CJEU). Nevertheless, they represent the clearest indication to date of how the Commission—and, in practice, national market surveillance authorities—are likely to approach the classification question.
The additional preparation time resulting from the AI Omnibus should therefore not be viewed as a reason to delay action, but rather as an opportunity for structured preparation. The key classification question—Is my system high-risk, and if so, does the Article 6(3) filter apply?—should be addressed, or at least prepared, as early as possible.
Our experience shows that delayed engagement with these issues, or a late change of course, often results in considerable additional effort. We would therefore be pleased to support you with an initial assessment of your AI systems and with establishing a practical framework for documenting and routinely reviewing your compliance documentation.
'/%3e%3c/svg%3e){kind=small}
'/%3e%3c/svg%3e){kind=small}
