On 7 May 2026, shortly before the publication of the Draft Guidelines, as part of a wider agreement in relation to the EU Digital Omnibus Package, the European Commission agreed in principle to substantially defer the compliance deadlines for obligations under the AI Act relating to high-risk AI systems. The revised timeline has been confirmed in the Draft Guidelines:
| AI system category | Original deadline | New deadline |
|---|---|---|
| Annex III systems (AI systems in any of the following areas: biometrics; critical infrastructure; education and vocational training; employment and workers; access to essential services; law enforcement; migration; asylum and border control; and the administration of justice and democratic practices) | 2 August 2026 | 2 December 2027 |
| Annex I systems (AI systems intended to be used as a safety component of a product, or the AI system is itself a product, covered by specific EU legislation set out at Annex I of the AI Act) | 2 August 2026 | 2 August 2028 |
Key takeaways for businesses
The Draft Guidelines are substantive and span three documents: a general principles document addressing horizontal concepts applicable to all AI systems classified as high-risk, and two annex-specific documents covering Annex I systems and Annex III systems respectively.
We have extracted some key takeaways from the Draft Guidelines below:
Human involvement does not remove high-risk status
The Draft Guidelines clarify that since human involvement cannot change the purpose and area in which a system is intended to be used, it has no effect on high-risk classification. Rather, human oversight is a prerequisite for compliance with the rules for high-risk AI systems pursuant to Article 14 AI Act and a necessary requirement for systems classified as high-risk.
Intended purpose is a decisive classification indicator
If the intended purpose of an AI system does not align with one of the use cases listed at Annex III, the system is not ‘intended to be used’ for such a use case and cannot be classified as high-risk. The Draft Guidelines clarify that a provider of an AI system must assess the intended use of the system prior to placing the system on the market or putting it into service. Actual use of the AI system is not required in order to be classified as high-risk.
Complex and agentic systems are assessed as a whole
If multiple AI systems form part of a more complex system whose combined intended purpose or joint outputs materially influence an individual decision, the combined configuration is treated as a single AI system for classification purposes, and split architectures are assessed as a whole to prevent circumvention. The Draft Guidelines clarify that this principle extends to interconnected and agentic AI systems coordinating linked actions where those actions serve a high-risk purpose.
Clarification on meaning of ‘safety component’ in relation to Annex I systems
The Draft Guidelines emphasise that the definition of ‘safety component’ under Article 3(14) AI Act is an autonomous definition of the AI Act that has its own meaning, independent of definitions of ‘safety component’ included in other EU harmonisation legislation listed in Annex I. Therefore, in assessing whether an AI system is a safety component for the purposes of high-risk classification under Article 6(1) AI Act, only the definition in Article 3(14) AI Act is relevant.
While the Guidelines remain in draft form and are subject to further consultation and potential revision, they nonetheless provide a useful early indication of the Commission’s supervisory approach to high-risk AI classification. In particular, the inclusion of detailed examples and structured interpretative guidance offers greater clarity on how the legal thresholds in Article 6 and Annexes I and III are likely to be applied in practice. For providers and deployers of high-risk AI systems grappling with AI Act compliance projects, this extension is significant and provides additional time to absorb the interpretive guidance offered by the Draft Guidelines ahead of the revised implementation deadlines.
The authors would like to thank Seoda Smyth for her contribution to this article.
