REFERENCE

The four conditions of Article 6(3), verbatim

What the regulation provides on this derogation. To read once the survey is done, not before.

Most companies in Annex III sectors accepted high-risk classification by default. Few have formally tested Article 6(3): the derogation that allows an Annex III system to exit high-risk where the provider demonstrates it does not pose a significant risk of harm. Post-Digital Omnibus (adopted 29 June 2026), the derogation was preserved by Parliament and Council against the Commission's proposal to remove it. But the mechanism changed: under the agreed Annex VIII Section B, providers self-classifying non-high-risk must register the system in the EU database with documented justification, even when they consider their system exempt from high-risk classification.

Take the six-question check See pricing

· Methodological assessment: not legal advice
· Article-by-article: Article 6(3) + Annex III §1–§8 + Annex VIII Section B + Article 49
· Documented justification deposable as a public artefact (post-Omnibus 7 May 2026)

Classification

When Article 6(3) becomes a path

Article 6(3) of Regulation (EU) 2024/1689 provides that a system listed in Annex III shall not be considered high-risk where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision-making. The derogation applies where ONE of four conditions is met. A profiling clause carve-out remains: a system that performs profiling of natural persons is always high-risk, regardless of the four conditions. A second structural anti-circumvention rule applies: where an AI system forms part of a complex or agentic configuration whose joint outputs materially influence an individual decision in a high-risk use case, the complete configuration is assessed as a single AI system. Splitting the system into modules to make each module fit a filter condition does not work (Draft Commission Guidelines on Article 6, §90, 19 May 2026).

Test: does your AI system fit one of the four conditions below AND avoid the profiling carve-out? If yes: Article 6(3) becomes a potentially eligible path, subject to documented self-assessment and public registration under Annex VIII Section B.

Condition	Article 6(3)	Annex VIII Section B
(a) narrow procedural task	Potentially eligible	Documented justification required
(b) improvement of prior human activity	Potentially eligible	Documented justification required
(c) detection of patterns without replacing human review	Potentially eligible	Documented justification required
(d) preparatory task to an Annex III assessment	Potentially eligible	Documented justification required
Profiling of natural persons	Always high-risk (carve-out)	Full Annex III registration required

Source: Reg. (EU) 2024/1689 Article 6(3) verbatim · profiling carve-out in last subparagraph of Article 6(3).

Four conditions

The four conditions of Article 6(3), verbatim

Article 6(3) lists four alternative conditions. One is sufficient; but the conditions are narrow, technical, and the defensibility of the self-classification depends on the documentation supporting the chosen condition.

(a) Narrow procedural task

The AI system is intended to perform a narrow procedural task. Examples in the Draft Guidelines (§92): sorting incoming school applications by level applied for (primary, secondary, high school), converting scanned documents to text for indexing, automatically filing items into predefined folders. IN: categorisation without evaluating suitability. OUT: any task that produces a value judgement (e.g., ranking applicants, labelling documents as ‘useful’ or ‘less useful’). The narrower and more mechanical the task, the stronger the position.

(b) Improvement of prior human activity

The AI system is intended to improve the result of a previously completed human activity. The Draft Guidelines (§96) require that the improvement does not change the rights, protection, or legal/economic position of the persons impacted. IN: quality-assurance systems that flag errors or contradictions in finalised human work; systems that convert human-validated content for accessibility or interoperability. OUT: any system that produces a substantially different result from the human decision. That is ‘replacement’, not ‘improvement’.

(c) Pattern detection without replacing human review

The AI system detects decision-making patterns or deviations from prior decision-making patterns and is not meant to replace or influence the previously completed human assessment, without proper human review. Three procedural limits (Draft Guidelines §§100–102): the human assessment must already be completed; the AI may only perform an ex-post comparative analysis; the AI must not infer criteria from past decisions to propose a new assessment. IN: an audit tool analysing past eligibility checks for quality reporting, without proposing outcomes on live cases. OUT: a system that re-evaluates current cases using patterns inferred from past decisions.

(d) Preparatory task to an Annex III assessment

The AI system performs a preparatory task to an assessment relevant for the Annex III use cases. The Draft Guidelines (§§107–108) draw the line: a system performs a preparatory task only if its output is a general input, for example supplementary information supporting the operator's reasoning. IN: a system referencing relevant legal provisions, jurisdictional information, or internal guidelines to support a human assessor. OUT: a system producing a specific recommendation or evaluation on the individual case. That output plays a decisive role in the assessment and is not preparatory.

Sources: Reg. (EU) 2024/1689 Article 6(3) subparagraphs (a), (b), (c), (d) verbatim · Commission draft guidelines on Article 6 (forthcoming consolidated text post-Omnibus).

Test your position

This page describes what Article 6(3) requires. It does not say where your own workflows stand. Six questions, two minutes, no account, nothing sent and nothing stored.