When is AI “High-Risk” under the EU AI Act? Key Takeaways from the Commission’s Draft Guidelines

Introduction

The EU Artificial Intelligence Act introduces a risk-based framework for regulating AI systems. A central concept in that framework is the category of “high-risk AI systems”, which triggers enhanced regulatory obligations. To support stakeholders in interpreting this concept, the European Commission has published draft guidelines on the classification of high-risk AI systems under Article 6 AI Act. These guidelines aim to facilitate consistent application of the rules and include practical examples to support interpretation and enforcement.¹

The draft guidance provides useful insight into how the Commission expects the high-risk threshold to be applied in practice. This article outlines the key takeaways and highlights the aspects most likely to be relevant for organisations assessing whether their AI systems fall within scope.

The starting point: when is an AI system “high-risk”?

The starting point is that the AI Act does not define high-risk AI in abstract terms. Instead, classification follows two clearly defined pathways.

The first route is linked to product safety regulation. An AI system will qualify as high-risk where it forms part of a regulated product — either as a product in its own right or as a safety component — and that product is subject to third-party conformity assessment under EU harmonisation legislation.² The rationale is straightforward: where EU law already imposes enhanced safety scrutiny on certain products, AI components affecting that safety should be subject to a comparable level of regulatory oversight.³

The second route focuses on the context of use. AI systems are also classified as high-risk where they are intended for specific use cases listed in Annex III AI Act, such as employment decision-making, education, law enforcement or access to essential services.⁴ The guidelines indicate that the scope of high-risk AI is intended to remain limited and proportionate, focusing on systems that are considered to present significant risks to health, safety or fundamental rights.⁵

The decisive factor: intended purpose

A central theme running through the guidelines is the importance of intended purpose. Classification does not depend solely on technical capabilities, but also on how the system is described, marketed and documented.

The intended purpose encompasses, among other things, instructions for use, contractual arrangements, promotional materials and technical documentation.⁶ Where an AI system is presented as broadly applicable across multiple contexts, without clearly excluding certain sensitive applications, those applications may be taken into account when determining its regulatory classification.⁷

This has tangible consequences in practice, particularly for organisations developing or marketing AI systems across multiple use cases. The guidelines make clear that generic disclaimers are insufficient to narrow the intended purpose where the system’s functionality or positioning suggests broader use.⁸ In other words, organisations cannot rely on formal wording alone if the practical reality points the other way.

For organisations deploying AI, an important practical implication is that limiting the use of a system at the deployment stage does not, in itself, alter its classification. Where a system is intended by the provider to be used in a high-risk context, that classification will generally continue to apply regardless of whether a particular deployer makes more limited use of the system in practice. This approach differs from a GDPR-style notion of purpose limitation: as a general rule, the classification of an AI system is not determined by how it is used in a particular deployment scenario, but by the broader intended purpose defined by the provider.

This is reinforced by the allocation of responsibilities under the AI Act. The assessment of the intended purpose — and therefore of whether a system is high-risk — is primarily carried out by the provider, subject to oversight by competent authorities. At the same time, other actors in the value chain should be mindful that they may assume provider obligations, or even trigger high-risk classification, where they modify a system or its intended purpose.⁹

Annex I: AI as part of regulated products

The Commission’s guidance on AI embedded in regulated products (the Annex I route) provides helpful clarification on the notion of a “safety component”, which is central to the classification exercise.

Not every AI system integrated into a regulated product will qualify. Rather, the assessment hinges on whether the system either performs a safety function — for example by preventing or mitigating risks — or whether its failure or malfunction could endanger the health and safety of persons or property.¹⁰ Many AI systems used for optimisation, performance improvement or efficiency will fall outside the definition, as their intended purpose is not safety-related.¹¹ At the same time, those same systems may still be caught where a failure could create a safety risk, depending on the product design and operating context.¹² This also means that organisations will need to look not only at what an AI system is intended to do, but also at how it interacts with the broader product and what risks may arise if it fails.

The final element of the assessment is whether the relevant product is subject to third-party conformity assessment under EU law. This reflects the legislator’s intention to limit high-risk classification under this route to cases where sectoral legislation already involves heightened regulatory scrutiny.¹³

Annex III: high-risk use cases

The Annex III pathway adopts a different logic, focusing on specific use cases rather than product categories. While this list includes areas that are frequently discussed in public debate, the guidelines underline that its scope remains precise and limited.

Only those applications explicitly listed in Annex III are relevant for classification, and the framework does not extend beyond them.¹⁴ That said, the practical reach of these provisions should not be underestimated. Many commonly deployed AI tools — particularly in HR, education or customer-facing contexts — may fall within scope where they materially influence decisions affecting individuals. In practice, this requires organisations to assess not only the functionality of their AI systems, but also the specific decision-making contexts in which they are deployed.

Not every system in scope is “high-risk”

An important counterbalance is the so-called “filter mechanism” under Article 6(3) AI Act. This allows certain systems to fall outside the high-risk category even where they are connected to Annex III use cases.

In particular, systems that perform purely preparatory, procedural or supportive functions, and do not materially influence decision-making, may be excluded from classification.¹⁵ This reflects an effort to ensure that the regime remains proportionate and does not capture AI systems with only a limited or indirect role in decision-making.

At the same time, the boundary between “supporting” and “influencing” decisions is not always straightforward. This is likely to become a key area of interpretation as the regime is applied in practice. However, determining whether a system merely supports or materially influences decisions will often require a case-by-case assessment.

Human involvement does not affect classification

Another important point is that the presence of human oversight does not, in itself, prevent a system from being classified as high-risk. Whether a human reviews or validates outputs is relevant for compliance with the AI Act’s requirements, but it does not affect the classification analysis itself, which remains grounded in intended purpose and use case. This is particularly relevant for organisations relying on “human-in-the-loop” models as part of their governance approach.

A shifting landscape: roles and future updates

Beyond the initial classification exercise, the guidelines also highlight that responsibilities may shift along the AI value chain. Parties other than the original provider may, in certain circumstances, assume provider obligations where they modify the system, place it on the market under their own name, or change its intended purpose in a way that leads to high-risk use.¹⁶

Finally, the framework is designed to evolve. The Commission is expected to review and update the list of high-risk use cases over time, including through delegated acts, ensuring that the regime can respond to emerging risks and technological developments.¹⁷

Practical takeaway

Taken together, the draft guidelines confirm that high-risk classification under the AI Act is neither as broad nor as open-ended as sometimes assumed. At the same time, it is highly fact-specific and dependent on how AI systems are designed, positioned and used.

For organisations, this means that classification should not be treated as a purely legal exercise, but as part of a broader assessment involving technical design, product positioning and actual use. It requires close alignment between technical design, product positioning and documentation, as well as a clear understanding of how systems are expected to be used and may reasonably be used.

 

1. Draft Guidelines (General Principles), p. 2, paras. (3)–(4).
2. Draft Guidelines (General Principles), p. 3, para. (7); Draft Guidelines (Annex I), p. 4–5, para. (27).
3. Draft Guidelines (Annex I), p. 2-3, paras. (18-22).
4. Draft Guidelines (General Principles), p. 3, para. (7).
5. Draft Guidelines (General Principles), p. 2, para. (2).
6. Draft Guidelines (General Principles), p. 4, para. (10).
7. Draft Guidelines (General Principles), p. 4, para. (12).
8. Draft Guidelines (General Principles), p. 4, para. (12).
9. Draft Guidelines (General Principles), p. 5, paras. (13)–(14).
10. Article 3(14) AI Act; Draft Guidelines (Annex I), p. 5–6, paras. (32)–(34).
11. Draft Guidelines (Annex I), p. 6–7, paras. (35)–(37).
12. Draft Guidelines (Annex I), p. 7–8, paras. (42)–(43).
13. Draft Guidelines (Annex I), p. 9–10, para. (50); p. 11–12, paras. (57)–(58).
14. Draft Guidelines (General Principles), p. 3, para. (7).
15. Draft Guidelines (General Principles), p. 2–3 (structure of Article 6 framework; targeted scope).
16. Draft Guidelines (General Principles), p. 5, para. (14).
17. Draft Guidelines (General Principles), p. 6, para. (451).