EU draws the line for high-risk AI as Article 6 guidance goes to consultation

The European Commission has published draft guidance on how to classify high-risk AI systems under the AI Act. That sounds like a legal technicality. It is not. This is the map many organisations have been waiting for before they can decide which AI systems are ordinary tools, and which must be treated as regulated operational infrastructure.

The Commission says the guidance is meant to help providers, deployers and market surveillance authorities assess whether an AI system falls into the high-risk category under Article 6. It also includes practical examples of systems that should, and should not, be classified as high-risk. The consultation is open until 23 June 2026.

For Norwegian and European companies, this matters well beyond the legal department. The AI Act will shape the control framework for companies selling into the EU, buying European systems, or using AI in processes that touch employees, customers, credit, safety, education or public administration. CIOs, CISOs, HR, legal and boards need to agree on the first question: Which AI systems do we actually run?

Two routes into high-risk

The draft follows the structure of Article 6. An AI system can be classified as high-risk through two main routes.

The first applies when AI is used as a safety component of a product, or is itself a product, covered by EU harmonisation legislation listed in Annex I and subject to third-party conformity assessment. That matters for industry, health, transport, machinery and other environments where AI can affect physical safety. The question is not whether the model looks impressive. The question is whether it sits inside something that must be safe, documented and testable.

The second route runs through the use cases listed in Annex III. Those include biometrics, education, employment, worker management, access to essential services, law enforcement and certain public-sector decision processes. These are areas where AI can affect rights, opportunities and power relationships. An HR model, a case-handling tool or a risk-scoring system can therefore land in a very different risk class from an internal writing assistant.

The key point is that classification cannot be delegated to a vendor slide deck. The organisation must understand the use case, the data, the decision point and the level of human control. The same technical component can be low-risk in one context and high-risk in another. Boards will miss the issue if they treat the AI portfolio as a list of IT tools rather than operational risk.

The deadline moved, but the work starts now

IAPP notes that the guidance arrives after delays. The toughest high-risk rules have also been given more time through the Digital Omnibus track: the deadline for standalone high-risk systems has moved to 2 December 2027, while high-risk systems embedded in products point to 2 August 2028.

That is not a reason to wait. It is a warning with a date attached. Organisations that wait for suppliers to send complete compliance packages will lose control of their own assessments. They will have to accept the supplier’s definition of risk, documentation and responsibility sharing. That is a weak position once the systems are already connected to data, workflows and decisions.

The practical work should start with a portfolio inventory. Which AI systems are used today, including bought tools, internal systems and AI features embedded in larger platforms? Which systems affect people directly? Which are connected to safety, workforce management, credit, case handling or product responsibility? Which vendors can document the model, data, logging, test regime and human oversight?

This is also a procurement issue. New AI contracts should require suppliers to disclose expected AI Act classification, documentation for risk assessment, change notifications, logging, incident handling and audit rights. A generic line saying the supplier “complies with applicable law” is not enough. That line has the structural strength of wet cardboard.

Boards should ask for a classification log

The most useful move now is simple: ask management for an AI classification log. It should show purpose, owner, supplier, data types, user group, decision type, expected AI Act category, rationale, controls and next review date.

The log does not need to be perfect in week one. It needs to exist. Without it, the AI Act becomes a legal clean-up exercise at the end of a project. With it, the regulation becomes part of portfolio governance.

The Commission’s draft is still under consultation, and the examples may be updated. But the direction is clear. AI systems that shift power over people, safety or rights should not be treated as ordinary software. They should be classified, documented and governed.

For leaders, the conclusion is practical: find the systems, classify the use, require documentation from suppliers and place risk ownership inside normal governance before the deadlines become operational. It is cheaper now than after the first supervisory question.

Sources and media

Primary source: European Commission, “Commission seeks feedback on the draft guidelines for the classification of high-risk artificial intelligence systems”, published 19 May 2026: https://digital-strategy.ec.europa.eu/en/news/commission-seeks-feedback-draft-guidelines-classification-high-risk-artificial-intelligence-systems

Background: European Commission, “Draft Commission guidelines on the classification of high-risk AI systems”, published 19 May 2026: https://digital-strategy.ec.europa.eu/en/library/draft-commission-guidelines-classification-high-risk-ai-systems

Corroboration: IAPP, “European Commission delivers draft high-risk AI guidelines after delays”, published 19 May 2026.

Thumbnail: OpenAI Image 2 / hogby.ai.