On 19 May 2026 the European Commission opened a five-week consultation on its draft Guidelines for the classification of high-risk AI systems under Article 6 of the AI Act, closing on 23 June.
The same draft confirms that the AI Omnibus has postponed application of Article 6(2) to 2 December 2027 and Article 6(1) to 2 August 2028. Brussels has produced the rigorous interpretive guidance industry has been asking for in the same week that it has moved the deadlines those rules attach to by sixteen and twelve months respectively. The two facts have to be read together to understand the present condition of European AI regulation.
The Guidelines are structured around the two statutory routes to high-risk classification under Article 6.
- Article 6(1) captures AI systems that are themselves products, or safety components of products, under the sectoral harmonisation legislation listed in Annex I, including the Machinery Regulation, the Toys Safety Regulation, the Radio Equipment Directive, medical devices and the automotive and aviation regimes.
- Article 6(2) captures the stand-alone use cases listed in Annex III, covering biometrics, critical infrastructure, education and vocational training, employment, access to essential private and public services, law enforcement, migration and border control, and the administration of justice and democratic processes.
The Guidelines are issued under Article 6(5) of the AI Act and address each route in turn with practical examples.
The Commission has resisted pressure to read the high-risk perimeter expansively. It has reaffirmed, by reference to Recitals 47 to 51, that classification is to be proportionate, targeted and safety-based, and that the AI Act does not itself extend the scope of sectoral legislation to products that legislation does not already cover. Most consumer smart-home appliances will therefore fall outside the high-risk category where their intended purpose is convenience, comfort or efficiency optimisation and where failure would produce discomfort or higher energy bills rather than safety harm.
Article 6(1): products and safety components
For Article 6(1) the central interpretive choice concerns the autonomous definition of safety component in Article 3(14), which the Commission treats as independent of the safety-component language used in the sectoral regimes. The definition contains two alternative criteria: an AI system intended to fulfil a safety function, and an AI system whose failure or malfunctioning would endanger the health or safety of persons or property. Intent and consequence operate as parallel routes to qualification. A combustion-efficiency optimiser in a household gas appliance is captured on the failure limb even where the provider has described its purpose in efficiency terms, because failure could lead to carbon monoxide formation, explosion or fire. The same logic captures lift door-timing systems, vehicle lane-assistance systems and agricultural spraying systems.
The most operationally significant clarification concerns conformity assessment. Relying expressly on Recital 15 of the new Toys Safety Regulation, and extending the same reasoning to the Machinery Regulation, the Commission confirms that a manufacturer’s procedural choice of internal control under Module A, where harmonised standards have been applied, cannot displace high-risk classification under Article 6(1). Some industry participants had hoped to keep open an interpretive route under which reliance on harmonised standards plus internal control would have lifted the product out of the high-risk category for AI Act purposes. This route is now closed.
Article 6(2): Annex III use cases and the role of intended purpose
For Article 6(2) and Annex III the provider’s articulated intended purpose is the primary classification lens, as expressed in instructions for use, technical documentation and promotional materials. Where a system is presented as broadly applicable across many contexts, and where its functionalities and capabilities make high-risk uses feasible and reasonably foreseeable, the intended purpose is deemed to encompass those high-risk uses. Disclaimers in terms of service that purport to exclude high-risk uses are insufficient on their own where the overall product positioning, examples and marketing effectively provide for or promote those uses. Limitations of use must, in the Commission’s words, be clearly, concretely and coherently described across all materials. The message for providers of general-purpose AI systems and enterprise copilot products is that contractual carve-outs without coherent operational substance will not insulate them from provider-level obligations under Article 25(1).
The Article 6(3) derogation mechanism and anti-circumvention
The Article 6(3) filter mechanism, by which a system that falls within an Annex III use case can be exempted from high-risk classification, is treated with appropriate restraint. The four conditions, comprising (i) narrow procedural task, (ii) improving the result of a previously completed human activity, (iii) ex post detection of decision-making patterns or deviations, and (iv) preparatory task, are exhaustive and to be interpreted narrowly as an exception to a fundamental rights protective regime. The exemption rests on a self-assessment by the provider with registration in the EU database, and profiling within the meaning of Article 4(4) GDPR removes the filter in every case.
Just as importantly, the Guidelines confirm in language directly aimed at modular and agentic systems that split architectures will be assessed as a single AI system. Where modular components combine to influence a high-risk decision, the configuration is treated as one system and individual modules cannot rely on the filter. The Commission expressly extends this principle to agentic AI systems whose linked actions together serve a high-risk purpose. For public-authority use cases the phrase “on behalf of” captures private entities performing delegated activities, closing the obvious route to circumvention through outsourcing.
The Omnibus postponement
Set against this interpretive clarity, the postponement of the operative dates under the AI Omnibus is striking. The original timetable would have brought Article 6(2) into application on 2 August 2026 and Article 6(1) on 2 August 2027. The dates are now 2 December 2027 and 2 August 2028, a delay of sixteen and twelve months respectively. The Commission has framed the delay as alignment with the readiness of harmonised standards, the establishment of regulatory sandboxes by 2 August 2026, the maturity of notified-body infrastructure with credible AI competence, and the publication of interpretive guidance of which this draft is a leading example. That justification is defensible on its own terms and reflects an honest recognition that the operational machinery required to enforce the high-risk regime is not yet in place across the Union.
The signal to providers and deployers is mixed. Compliance programmes built to the original timetable now have additional runway, while the substantive obligations attached to high-risk classification have not moved in any material respect. The substance of the regime, including risk management systems, data governance, technical documentation, logging, transparency, human oversight, accuracy, robustness and cybersecurity, requires months of design, testing and embedding into existing enterprise risk frameworks. An organisation that begins serious implementation work in 2027 for an August 2028 deadline will find that the deadline arrives sooner than the work can be completed, and that supervisory authorities, by then equipped with the sandboxes and standards the Omnibus is intended to enable, will have less patience for late starters than they had in 2024.
Consultation and next steps
Stakeholders can submit feedback through the online questionnaire on the Commission’s EU Survey platform until 22:00 CET on 23 June 2026. The draft Guidelines, an explorer of the use cases, and the consultation questions are available on the AI Act Single Information Platform. The Commission has indicated that responses received outside the questionnaire will not be reflected in the final summary report. The Commission has done useful and disciplined work. The Guidelines are technically defensible, properly anchored in the existing architecture of the New Legislative Framework and appropriately restrained in their interpretive ambition.
The Omnibus postponement is a more ambiguous signal. It buys time on the precondition that the time is used to build the infrastructure of an enforceable regime. Whether that condition will be satisfied is now the central question of European AI governance. The Commission has approximately twenty-seven months in which to provide an answer that is credible to industry, to civil society and to the wider world that is watching how the Union’s flagship technology regulation is operationalised.
For further information, please contact Barry Scannell or any member of the William Fry Technology Group.