EU AI Act Weekly Radar: Brussels Opens High-Risk Classification Consultation as Security and Omnibus Questions Converge

The biggest EU AI Act development this week is not a new enforcement action or a fresh wave of delegated rules. It is something more foundational: the European Commission has now put forward draft guidelines on how to determine whether an AI system is high-risk under the Act, and it has opened a feedback window that runs until 23 June 2026. For providers, deployers, product counsel, and compliance teams, that makes this week less about abstract governance and more about operational scoping. (European Commission, “Commission seeks feedback on the draft guidelines for the classification of high-risk artificial intelligence systems”; European Commission, “Targeted consultation on the draft guidelines for the classification of high-risk artificial intelligence systems”)

That shift matters because classification is where a large share of AI Act compliance effort begins. If a system falls inside the high-risk perimeter, the downstream obligations can be extensive. If it does not, a very different compliance pathway may apply. The Commission’s draft therefore speaks directly to one of the market’s most persistent implementation questions: how Article 6 should be applied in practice, especially across the split between product-safety regulated systems and standalone Annex III use cases. (European Commission, “Draft Commission guidelines on the classification of high-risk AI systems”)

1) The central development: draft high-risk classification guidance is finally on the table

On 19 May 2026, the Commission published draft guidance and practical examples aimed at helping actors determine when an AI system should be classified as high-risk under the AI Act. The Commission’s own framing is explicit: the material is intended to support both providers and deployers, and the draft has also been published on the AI Act Single Information Platform. Feedback is open until 23 June 2026. (European Commission, news item)

The underlying draft text is especially important because it addresses both sides of the high-risk architecture:

• Article 6(1) / Annex I cases tied to product-safety legislation; and
• Article 6(2) / Annex III cases involving standalone high-risk use cases.

That structure comes directly from the Commission’s library publication describing the draft, which says the guidelines set out the Commission’s interpretation of Article 6 and include examples of systems that should, and should not, be classed as high-risk. (European Commission, draft guidelines publication)

In practical terms, this is the kind of document teams have been waiting for to answer questions such as:

• Is the system captured because it is embedded in a regulated product context?
• Is it a standalone AI use case that falls under Annex III?
• What facts are likely to matter when documenting why a system is in or out of scope?
• How should borderline systems be assessed when the technical function is similar, but the deployment context changes?

The Commission’s consultation page shows that Brussels is not treating this as a purely academic exercise. The consultation specifically asks whether the draft guidance and examples help actors determine whether a system falls into the high-risk category. That signals a real attempt to test whether the guidance is usable in live compliance settings, not just whether it is legally elegant. (European Commission, consultation page)

2) Why this matters now: scope decisions drive nearly every other AI Act workstream

For many organizations, the hardest part of AI Act readiness is not listing obligations. It is deciding which systems actually trigger them.

That is why the timing of the draft guidelines is so significant. Before teams can map transparency obligations, provider and deployer responsibilities, technical documentation, or internal governance controls, they need a defensible answer on classification. A weak scoping decision can create two opposite problems:

• under-classification, which may leave a genuinely high-risk system outside required governance; or
• over-classification, which can impose unnecessary cost, delay, and control overhead on systems that may not belong in the high-risk category.

The Commission’s move therefore has broader significance than the title of the document might suggest. It affects procurement triage, product design reviews, contract allocation between provider and deployer roles, and how companies prioritize legal and engineering resources over the coming implementation phases.

For startups and SMEs, this is especially relevant. Even without new SME-specific relief announced in the sources this week, clearer classification logic can materially reduce uncertainty costs. When the high-risk threshold is unclear, smaller teams often spend scarce time preparing for the strictest case. More usable guidance may not eliminate that burden, but it can make compliance planning less speculative.

3) What the market readout adds: pressure had been building around delays and the Omnibus timeline

A secondary report from IAPP says the Commission’s 19 May draft guidance presents a three-part framework for classifying high-risk AI and arrived after earlier delays had intensified pressure around the AI Omnibus timeline. The same report emphasizes the divide between Article 6(1)/Annex I and Article 6(2)/Annex III cases. (IAPP, “European Commission delivers draft high-risk AI guidelines after delays”)

Even allowing for the fact that this is a secondary characterization rather than the primary legal text, the point is important: implementation guidance is landing in a political and operational environment where companies have already been recalibrating expectations around timing, sequencing, and overlap with other rulebooks.

That context helps explain why this week’s Commission draft feels larger than a routine publication. It is arriving at a moment when the market is trying to answer three questions at once:

1. What is high-risk?
2. When do the relevant obligations bite?
3. How much overlap will remain between AI Act compliance and sector-specific product regulation?

The new draft guidelines do not answer every one of those questions, but they do move the first one forward in a concrete way.

4) The Omnibus backdrop: timing and overlap remain central, but caution is required

Another item in this week’s radar is a law-firm summary of the provisional AI Omnibus agreement. Osborne Clarke says the deal includes delayed high-risk deadlines, mechanisms aimed at reducing overlap with sectoral product rules, and a new ban covering AI used to generate non-consensual intimate content or child sexual abuse material. (Osborne Clarke, “Provisional Agreement on the AI Omnibus: Key Changes at a Glance”)

From a governance perspective, the most important connection is not any single reported change. It is the combined signal:

• Brussels is refining how high-risk status should be determined;
• market observers are focused on whether deadlines shift; and
• regulators appear sensitive to friction caused by overlapping frameworks.

If that summary proves directionally accurate, it would reinforce a pattern that has become increasingly visible in EU AI governance: the core architecture of the AI Act remains intact, but implementation is being shaped by practical concerns about legal certainty, supervisory capacity, and cumulative compliance load.

For lextrace readers, the main takeaway is straightforward. Do not treat classification guidance and timeline debates as separate topics. They are tightly linked. A delay or recalibration in one area changes how organizations prioritize the other. If high-risk scoping becomes clearer while key dates move, companies may gain more time, but they also lose excuses for weak internal triage.

5) Parliament’s cyber-capable AI debate shows where enforcement attention may intensify next

Alongside the classification guidance, the European Parliament scheduled a plenary debate on EU readiness for emerging cyber-capable AI models. According to Parliament’s agenda note, the discussion concerns advanced cyber-capable AI models, how EU authorities are testing them, and how this interacts with the AI Act, the NIS2 Directive, and the Cyber Resilience Act. (European Parliament, “EU readiness for emerging cyber-capable AI models”)

This does not create new AI Act obligations by itself. But as a governance signal, it matters.

First, it shows that the EU conversation is not limited to classic high-risk enterprise or public-sector use cases. Policymakers are also focused on advanced model capability risk, especially where cybersecurity consequences could be severe.

Second, it highlights the growing reality of cross-framework supervision. The Parliament note explicitly places the AI Act alongside NIS2 and the Cyber Resilience Act. That suggests organizations working on advanced or security-relevant AI should avoid overly siloed compliance programs. The relevant governance question is no longer just, “Are we inside the AI Act?” It is increasingly, “How will AI-specific obligations interact with cyber resilience and incident-preparedness expectations?”

Third, the debate adds texture to the broader enforcement outlook. Even where formal AI Act enforcement pathways are still maturing, political attention often points to where oversight scrutiny may become sharper. Cyber-capable models, testing access, and preparedness planning appear to be moving up that agenda.

6) The practical significance of Article 6 and Annex III is becoming more concrete

The Commission’s draft guidance is especially consequential because Article 6 and Annex III have long been among the most difficult parts of the AI Act to operationalize.

At a high level, the challenge has been that classification is not purely technical. It depends on legal context, intended purpose, deployment setting, and the interaction between AI functionality and regulated use case categories. That means teams cannot solve classification with a model card alone, nor with a legal memo written in isolation from product facts.

This week’s publication matters because it appears designed to make those boundary questions more concrete through examples. The Commission expressly says the draft includes practical examples, and that alone is meaningful. In AI regulation, examples often do more compliance work than broad principles, because they help organizations test analogies against real product configurations. (European Commission, news item; draft guidelines publication)

For providers, the immediate implication is better support for product-by-product scoping decisions. For deployers, the implication is equally important: deployers frequently inherit classification assumptions made upstream, but they still need enough understanding to assess whether the intended use and deployment environment align with those assumptions.

7) What organizations should do with this week’s update

This roundup is not legal advice, but the operational implications are clear from the published materials.

Re-open your high-risk inventory

Any company that built a preliminary AI Act inventory before the new draft guidance should revisit it. Systems previously tagged as “uncertain,” “borderline,” or “awaiting guidance” are the obvious priority set.

Separate product-safety and standalone use cases

Because the Commission draft explicitly addresses both Annex I and Annex III pathways, internal review processes should avoid collapsing them into one generic “high-risk” bucket. The classification logic may differ materially depending on which route is in play. (European Commission, draft guidelines publication)

Review provider/deployer assumptions

The Commission’s news item makes clear that the guidance is aimed at both providers and deployers. Organizations should therefore check whether contracts, procurement flows, and governance documentation rely on role allocations that remain too simplistic for real-world deployment chains. (European Commission, news item)

Consider participating in the consultation

The consultation closes on 23 June 2026. Teams facing recurring uncertainty around concrete examples, especially in borderline or hybrid use cases, now have a live opportunity to push for clearer final guidance. (European Commission, consultation page)

Do not isolate AI Act work from cyber governance

The Parliament debate on cyber-capable AI models is a reminder that AI oversight is increasingly tied to broader digital resilience expectations. Organizations dealing with advanced models or security-sensitive deployments should make sure AI governance, cybersecurity, and product assurance teams are not working from incompatible assumptions. (European Parliament, agenda note)

8) The bigger picture: implementation is moving from headline law to classification discipline

This week’s radar points to a mature phase of EU AI Act implementation. The questions are becoming less theoretical and more administrative, evidentiary, and supervisory.

That is often the real turning point in regulation.

When policymakers publish draft guidance on classification, open targeted consultations on whether examples are usable, debate cyber-capable model preparedness in plenary, and continue refining the timing and overlap story through Omnibus discussions, they are doing more than filling gaps. They are shaping how the AI Act will function in practice.

For the market, the message is equally clear. The next compliance advantage will not come from repeating that the AI Act is “risk-based.” It will come from building a repeatable method for deciding, documenting, and defending where a system sits inside that risk structure.

This week, Brussels gave the market a new draft tool for exactly that job. The consultation window is open, the political focus on advanced-model risk is rising, and the broader implementation timeline remains closely watched. For AI governance teams, that makes now the right moment to tighten classification workflows before those interpretations harden into supervisory expectations.