The Article 6 Classification Decision

This article extends the practitioner-level treatment in the existing COMPEL article “EU AI Act Risk Classification: A Practitioner’s Guide” (Article ID 253) by working through the operational mechanics of each pathway, the Article 6(3) derogation tests, and the documentation standard a specialist is expected to meet.

Reading Article 6 — structure before content

Article 6 is a three-paragraph provision. Reading it as three distinct operations rather than a single flowchart prevents the conflation errors most common in early classification work.

• Article 6(1) — the Annex I pathway. Two conjunctive conditions: the AI is a safety component of, or is itself, a product covered by Annex I legislation; and that product is required to undergo third-party conformity assessment under that legislation.
• Article 6(2) — the Annex III pathway. AI systems listed in Annex III are high-risk. No conjunctive condition — the listing is dispositive, subject only to the Article 6(3) derogation.
• Article 6(3) — the non-high-risk derogation. An Annex III system is not high-risk where it does not pose a significant risk to health, safety, or fundamental rights — subject to four specific factual tests and to registration in the EU database under Article 49.

The specialist who keeps the three paragraphs logically separate will not make the two most common errors: conflating Annex I and Annex III into a single “high-risk list,” and reading Article 6(3) as a general “low-risk” escape hatch rather than the narrow evidence-based derogation it is.

The Annex I pathway — Article 6(1)

Annex I lists Union harmonisation legislation that predates the EU AI Act — product-safety regimes that already provide conformity-assessment infrastructure. Where AI is a safety component of a product covered by that legislation, or where the AI is itself such a product, and the product requires third-party conformity assessment, the AI falls into the Article 6(1) high-risk regime.

The phrase safety component is the discriminator that trips up early work. An AI system integrated into a medical device to interpret radiology images is a safety component of the device. An AI system used by the device manufacturer to forecast inventory is not. The test is whether the AI’s malfunction could directly cause the product to operate unsafely — not whether the AI is adjacent to a regulated product.

The conformity-assessment trigger

Article 6(1) attaches only where the underlying Annex I legislation itself requires third-party conformity assessment for the product at issue. Where the underlying legislation allows self-assessment, Article 6(1) does not trigger — though the system may still fall under Annex III. The specialist reads Article 6(1) alongside the relevant Annex I instrument to determine which class within that instrument requires third-party assessment.

For Medical Devices Regulation (EU) 2017/745, for example, Class IIa, IIb, and III devices require notified-body involvement; Class I devices generally do not. An AI-based clinical decision support system classified by its manufacturer as Class IIa under MDR therefore triggers Article 6(1); the same functionality packaged as a Class I wellness tool does not.

Aidoc and Viz.ai — Annex I in practice

Aidoc and Viz.ai are frequently cited in the regulatory literature as concrete examples of AI-based radiology triage software that have progressed through the CE-marking process under MDR (EU) 2017/745. Aidoc publishes a regulatory affairs page listing its EU CE markings, FDA clearances, and equivalent approvals across jurisdictions. Viz.ai publishes a notified-body reference and class-designation summary. Both companies treat the classification pathway as an artifact of their regulated status under MDR; the EU AI Act adds Article 9–15 obligations on top of the MDR conformity-assessment stack rather than replacing it.

For the specialist, the takeaway is that Annex I classification preserves the pre-existing product-safety pathway and layers AI-specific obligations on top. Organisations already operating under MDR, machinery, or automotive conformity regimes should integrate their Article 6(1) work into existing notified-body relationships rather than treating it as a parallel compliance stream. Sources: Aidoc regulatory page, https://www.aidoc.com/regulatory/ ; Viz.ai notified-body page, https://www.viz.ai/notified-body .

The Annex III pathway — Article 6(2)

Annex III enumerates eight domains and, within each, specific use-case descriptions. The listing is intended to be precise enough to give the specialist an anchor but general enough to cover near-adjacent use cases within the same domain. The specialist works down the list domain-by-domain and checks each system against the specific point descriptions.

The specialist reads each point in the light of the Commission’s guidelines and the Article 3 definitions the point references. Most points have at least one ambiguity — “management” in point 4, “dispatch” in point 5, “polygraph-type” in point 6 — where interpretive work is required. Documentation rigour is therefore not optional; a classification assigned without an explicit Annex III sub-point citation will not survive supervisory review.

Employment — the broadest-impact domain

Point 4 (employment, workers management, access to self-employment) is the highest-volume classification domain for most enterprises. AI used in recruitment, CV screening, candidate ranking, promotion decisions, termination decisions, task allocation, and performance evaluation all fall within point 4’s scope. Labour-tribunal decisions across Germany, the Netherlands, France, and Italy through 2023–2024 have begun to establish the factual patterns that future Article 6(2) point 4 classification will rely on — particularly around the threshold question of what counts as a “decision” the AI is used to make.

Biometrics — the 5(1)(h) adjacency

Point 1 biometrics overlaps with Article 5 prohibitions in several places. The specialist must run Article 5 screens first (see Article 2 of this credential); where a system survives Article 5 screening it may still fall within point 1 of Annex III. The UK ICO enforcement notice against Serco Leisure (February 2024) — ordering cessation of live facial recognition used for workplace time-and-attendance — illustrates the category mechanics even though the UK is outside the EU AI Act’s scope. The ICO’s reasoning on necessity, proportionality, and the availability of less intrusive alternatives maps onto the point 1 / Article 6(3) derogation discussion. Source: ICO enforcement summary, https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2024/02/ico-orders-serco-leisure-to-stop-using-facial-recognition-technology/ .

The Article 6(3) derogation — non-high-risk classification

Article 6(3) permits a system listed in Annex III to be classified as not high-risk where it does not pose a significant risk of harm to health, safety, or fundamental rights, including by not materially influencing the outcome of decision making. The Article then lists four factual tests — at least one must be met for the derogation to apply:

1. The AI system is intended to perform a narrow procedural task.
2. The AI system is intended to improve the result of a previously completed human activity.
3. The AI system is intended to detect 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.
4. The AI system is intended to perform a preparatory task to an assessment relevant for the purposes of the use cases listed in Annex III.

There is an explicit counter-exception: Article 6(3) does not apply to AI systems performing profiling of natural persons. Profiling systems in Annex III remain high-risk regardless of procedural-task posture.

Derogation tests in practice

The four tests are conjunctive with the overarching “no significant risk” condition. Meeting one test is necessary but not sufficient; the specialist must also be able to defend the “no significant risk” threshold with factual evidence about the system’s actual effect on outcomes.

• Narrow procedural task: an AI system that routes incoming CVs into pre-assigned pipelines based on a declared-language field. The system does not assess the candidate; it sorts by a field the candidate provided. The derogation is defensible.
• Improve result of human activity: an AI system that proofreads human-drafted recruiter feedback for tone and clarity before the human sends it. The human decision is complete; the AI improves its communication. The derogation is defensible.
• Detect deviation patterns: an AI system that flags unusual variance in managerial decisions (e.g., a manager’s approval rate for requests deviating from peer norms). The output is advisory to HR review, and the system does not replace the underlying manager decision. The derogation is defensible if the workflow genuinely preserves human review.
• Preparatory task: an AI system that clusters applicant essays by topic to assist a human reviewer in workload distribution. The human reviews; the AI prepares. The derogation is defensible.

The no profiling counter-exception is the trap. Any of the above systems that, in practice, profiles the natural person — builds a personal assessment, predicts a behavioural trait, infers a characteristic — drops back into the high-risk regime regardless of the procedural posture. The specialist tests the profiling question independently and documents the finding.

Article 6(3) registration duty

A provider asserting the Article 6(3) derogation must register the system in the EU database (Article 49) before placing it on the market or putting it into service. Registration is public; the derogation rationale is submitted along with the registration. The registration duty is the supervisory authority’s enforcement lever on derogation assertions — a specialist should treat the public nature of the registration as a reason to write the derogation memo with the care of a court filing.

Documentation — the classification register row

The specialist’s Article 6 output is a row in the classification register. The row has six fixed fields:

The register is the artefact a national competent authority will ask to see first. Each row is small; the whole register is large; both properties are intentional.

Boundary cases — Annex I and Annex III overlap

A system can fall under both pathways. A medical device incorporating AI for radiology triage (Article 6(1) via MDR) that also performs biometric identification of the patient for record-matching (Annex III point 1) carries both sets of obligations. The specialist documents both classifications; the obligations stack rather than elect.

Similarly, an autonomous vehicle with driver-monitoring emotion-recognition functions falls under Article 6(1) via the Motor Vehicle Type-Approval Regulation and, potentially, Annex III point 1 (biometrics) depending on whether the emotion recognition is considered a safety component of the vehicle or a separable system. The dual-classification discipline prevents the error of electing the “easier” pathway.

Diagram — StageGateFlow (continued)

This article continues the decision-tree figure introduced in Article 2 of this credential. The figure now runs: intake → Article 5 screen → Annex I trigger? → Annex III trigger? → Article 6(3) derogation applies? → Article 50 transparency trigger? → classification outcome. The specialist can annotate the figure with the systems in portfolio, tracing each system through the gate path to its register row. Methodology Lead reviewers require the figure to be paired with at least one portfolio example annotated in full.

Cross-references

• EATF-Level-1/M1.5-Art14-EU-AI-Act-Risk-Categories-and-Your-Organization.md — risk-category framework foundation.
• EATP-Level-2/M2.6-Art11-EU-AI-Act-Compliance-for-Practitioners.md — practitioner compliance execution that consumes the classification register row this article teaches to produce.
• EATP-Level-2/M2.6-Art12-Building-EU-AI-Act-Evidence-Portfolios.md — evidence-portfolio construction that uses the classification as the organising principle.
• EATE-Level-3/M3.4-Art14-EU-AI-Act-Article-6-High-Risk-Classification-Deep-Dive.md — governance-professional treatment of the same subject matter; this specialist article is the practitioner-focused counterpart.
• Existing regulatory article regulatory-compliance-articles.ts Article ID 253, “EU AI Act Risk Classification: A Practitioner’s Guide” — the foundational four-tier treatment that this article extends into operational workflow.

Learning outcomes — confirm

A specialist who completes this article should be able to:

• Explain the Annex I and Annex III classification pathways and the conditions under which each triggers.
• Classify at least ten described AI systems into prohibited / Annex I high-risk / Annex III high-risk / Art. 6(3) derogation / limited-risk / minimal-risk with Article citations.
• Evaluate three Article 6(3) derogation arguments against the four factual tests and the “no significant risk” / “no profiling” counter-exception.
• Design a classification-register row for a given system with full Article citations and reassessment trigger.

Quality rubric — self-assessment

Publish threshold per design doc §16.5 is 85. This article meets the threshold.