Skip to content

Commission draft worked examples — high-risk classification

This page paraphrases illustrative examples from the Commission's draft Annex I and Annex III classification guidelines (May 2026) organised by Annex III sub-point and Annex I sectoral law. For the classification framework — the Article 6 routes, the Article 6(3) filter mechanism, the profiling override, the timeline — see the companion page High-risk classification.

Status — DRAFT

Commission interpretive guidance, not binding law. The EU AI Act text and any Court of Justice of the European Union (CJEU) interpretation prevail in case of conflict.

This page reflects the draft Commission Guidelines published in May 2026. The consultation closes 23 June 2026 and the text may change before formal adoption. Each example is illustrative — the Commission's draft warns at paragraph (118) that classification ultimately turns on the specific architecture, functionality, and intended purpose of the system at issue.

Quick decision

  • You want to find an example for your Annex III use case → jump to the relevant H2 below.
  • You want the legal framework → see High-risk classification.
  • You want to know whether profiling is involved → the profiling override at the third sub-paragraph of Article 6(3) bars the filter mechanism for any system that performs profiling within the meaning of Article 4(4) GDPR.
  • None of the Annex III categories fits your system → check Annex I product-safety route (medical devices, machinery, lifts, motor vehicles, etc.); see the Annex I sectoral examples below.

TL;DR

  • The Commission's worked examples are illustrative, not binding (paragraph (118) Annex III annex).
  • Classification ultimately turns on architecture, functionality, intended purpose — relabelling a system as another category does not change the analysis (paragraph (118) Annex III; paragraph (12) general principles).
  • Profiling within Article 4(4) GDPR is the most-likely overlooked element; many Annex III systems perform profiling and therefore cannot use the Article 6(3) filter.
  • Most "high-risk" / "not high-risk" examples in the Commission's draft turn on a small number of recurring distinctions: per-individual output vs aggregate output; influences a substantive decision vs purely administrative/procedural; processes personal data vs anonymised data; sits in the same context vs cross-context use.
  • For the Annex I route, the test is two cumulative conditions — Annex I legislation applies AND third-party conformity assessment is required.

Primary source

Draft Commission Guidelines on the classification of high-risk AI systems under Article 6 of the EU AI Act for stakeholder consultation, May 2026 — three companion documents: general principles, Article 6(1) Annex I annex, Article 6(2) Annex III annex. © European Union. Draft guidelines issued pursuant to Article 6(5) AI Act; not yet adopted. Consultation closes 23 June 2026; text may change.

How to read this page

Each table row is a paraphrase of a Commission worked example, not a verbatim quote. The "Section" column points back to the relevant Annex III annex section so you can read the Commission's full treatment.

The Commission's examples typically come in three buckets per sub-point:

  • (a) Practical examples of AI systems falling within the use case (high-risk by default).
  • (b) Practical examples of AI systems falling outside the use case (not high-risk for that reason).
  • (c) Practical examples of AI systems falling within the use case but exempted under the filter mechanism of Article 6(3) (high-risk by default but a possible filter applies).

Where the Commission identifies a profiling override, the system stays high-risk regardless of (c).

Annex III(1) — Biometrics (section 3.1)

The Commission treats Annex III(1) as three use cases: (a) remote biometric identification (post-RBI); (b) biometric categorisation by sensitive attributes; (c) emotion recognition.

ScenarioCommission verdictSection
Post-RBI system for police evidence — identifying a suspect from prior CCTV footageHigh-risk under Annex III(1)(a)3.1.2
Voluntary biometric enrolment for access to a service (e.g. payment with consent)Not Annex III(1)(a) because the individual actively presents themselves; not "remote"3.1.2
Biometric categorisation to infer attributes that are not on the Article 5(1)(g) sensitive listHigh-risk under Annex III(1)(b)3.1.3
Biometric categorisation inferring sensitive attributes on the Article 5(1)(g) listProhibited under Article 5(1)(g), not Annex III high-risk3.1.3
Workplace or education emotion recognitionProhibited under Article 5(1)(f), not Annex III(1)(c) high-risk3.1.4
Emotion recognition outside workplace/education (e.g. consumer research)High-risk under Annex III(1)(c)3.1.4

Annex III(2) — Critical infrastructure (section 3.2)

The Commission treats Annex III(2) as safety components of management/operation of critical digital infrastructure, road traffic, supply of water, gas, heating, electricity.

ScenarioCommission verdictSection
AI-driven grid-stability management for an electricity transmission operatorHigh-risk under Annex III(2)3.2.7
AI optimisation of office HVAC in a single buildingNot Annex III(2); the building is not critical infrastructure3.2.7
AI traffic-signal control system in a major urban networkHigh-risk under Annex III(2)3.2.3
AI demand-forecasting for a water utility, not directly controlling supplyNot Annex III(2) if it does not act as a safety component of the management system3.2.4

Annex III(3) — Education and vocational training (section 3.3)

Four sub-points: (a) access / admission; (b) evaluating learning outcomes; (c) appropriate level of education; (d) monitoring prohibited behaviour during tests.

ScenarioCommission verdictSection
AI essay-scorer used to make admission decisions for a higher-education institutionHigh-risk under Annex III(3)(a)3.3.2
AI essay-scorer used only for formative feedback during a course, no impact on gradesNot Annex III(3); not affecting admission, learning-outcome evaluation, or level assignment3.3.2
AI grading of standardised tests where the result affects passing/failingHigh-risk under Annex III(3)(b)3.3.3
AI placement test for assigning students to language proficiency levelsHigh-risk under Annex III(3)(c)3.3.4
AI proctoring system flagging suspected cheating from webcam + audio + keystrokes during online examHigh-risk under Annex III(3)(d)3.3.5
AI proctoring system that also infers student emotional state during the examProhibited under Article 5(1)(f) (emotion recognition in education institutions)3.3.5

Annex III(4) — Employment and worker management (section 3.4)

Two sub-points: (a) recruitment / selection; (b) decisions affecting work-related relationships, promotion/termination, task allocation based on personal traits, monitoring / performance evaluation.

ScenarioCommission verdictSection
AI CV-screening tool that filters applications and ranks candidatesHigh-risk under Annex III(4)(a); typically performs profiling → filter unavailable3.4.2
AI tool that targets job advertisements to specific demographics based on inferred attributesHigh-risk under Annex III(4)(a); profiling override applies3.4.2
AI tool that helps a recruiter format and post a job ad (no candidate evaluation)Not Annex III(4) on its face; potential 6(3)(d) preparatory-task candidate3.4.2
AI worker-performance scoring affecting promotion or terminationHigh-risk under Annex III(4)(b); profiling override applies3.4.3
AI workplace emotion-recognition affecting worker treatmentProhibited under Article 5(1)(f), not Annex III(4)(b) high-risk3.4.3
AI task-allocation system that assigns shifts based on stated availability only (no profiling)Not Annex III(4)(b) where there is no personal-trait-based allocation3.4.3

Annex III(5) — Essential services (section 3.5)

Four sub-points: (a) public-assistance eligibility; (b) creditworthiness / credit scoring of natural persons (excluding fraud detection); (c) life and health insurance risk assessment and pricing; (d) emergency-call dispatch / triage.

ScenarioCommission verdictSection
AI system evaluating eligibility for unemployment benefitsHigh-risk under Annex III(5)(a); profiling override likely3.5.2
Consumer credit-scoring model for personal-loan applicationsHigh-risk under Annex III(5)(b); profiling override applies3.5.3
AI transaction-anomaly detector intended for fraud detection (no creditworthiness / credit-score output)Not Annex III(5)(b) under the explicit fraud-detection carve-out; AML/CFT use is outside 5(b) only where the system is not intended for creditworthiness evaluation3.5 overview + 3.5.3
Life-insurance underwriting AI that prices premiums based on individual riskHigh-risk under Annex III(5)(c); profiling override applies3.5.4
AI insurance fraud-detection systemNot Annex III(5)(c) where the system targets fraud, not pricing3.5.4
Emergency-call triage AI that ranks 112 calls for dispatch priorityHigh-risk under Annex III(5)(d)3.5.6

The Annex III(5)(b) and (c) interaction with sectoral financial law (Capital Requirements Regulation, Insurance Distribution Directive) is addressed at section 3.5.5.

Annex III(6) — Law enforcement (section 3.6)

Seven sub-points (6(a)–(g)) covering victim-risk assessment, polygraph-style evidence reliability, crime detection / location, profiling for investigations, profiling under the Law Enforcement Directive, etc. Note that several practices that would have been Annex III(6) high-risk are instead prohibited under Article 5 — particularly real-time RBI in publicly accessible spaces (Article 5(1)(h)) and individual criminal-offence risk prediction based solely on profiling (Article 5(1)(d)).

ScenarioCommission verdictSection
AI tool assessing victim-vulnerability score for protection-order prioritisationHigh-risk under Annex III(6)(a); profiling override likely3.6.2
AI polygraph-style truthfulness inference for evidence reliability assessmentHigh-risk under Annex III(6)(b)3.6.3
AI tool detecting deep-fakes in evidentiary recordingsHigh-risk under Annex III(6)(c)3.6.4
Predictive-policing AI based solely on personality profilingProhibited under Article 5(1)(d), not Annex III(6) high-risk3.6 overview
Predictive-policing AI combining objective verifiable facts and profiling, supporting human assessmentOutside 5(1)(d) "solely" limb; Annex III(6)(d) high-risk applies; where profiling is performed, the Article 6(3) filter is unavailable3.6.5

Annex III(7) — Migration, asylum and border (section 3.7)

Sub-points 7(a) polygraph-style truthfulness inference; 7(b) risk assessment of persons seeking entry/stay; 7(c) examining asylum/visa/residence applications and associated complaints, including evidence reliability; 7(d) detecting, recognising or identifying natural persons in migration/asylum/border-control contexts, excluding travel-document verification.

ScenarioCommission verdictSection
AI risk-scoring at border control to recommend secondary inspection of named travellersHigh-risk under Annex III(7)(b); profiling override applies3.7.3
AI aggregating travel patterns to produce anonymous trend analytics (no per-individual flag)Not Annex III(7)(b) where output is not applied to identified individuals3.7.3
AI document-authenticity scanner for visa applicationsHigh-risk under Annex III(7)(c) (associated with examining applications); travel-document verification per se is carved out of 7(d)3.7.4
AI translation tool for asylum interviews producing literal translation onlyWithin Annex III(7)(c) on its face; candidate for Article 6(3)(d) preparatory-task filter if it does not materially influence the substantive assessment3.7.4
AI tool flagging credibility / inconsistency in asylum testimonyHigh-risk under Annex III(7)(c) (evidence reliability in examining applications); profiling override likely3.7.4
AI biometric identification system at border control (not travel-document verification)High-risk under Annex III(7)(d)3.7.5
Maritime / unmanned platform sensing used for search-and-rescue navigation safety onlyNot Annex III(7); outputs are for safety of navigation, not border-control response3.7 overview

Annex III(8) — Administration of justice and democratic processes (section 3.8)

Two sub-points: (a) research and interpretation of facts and law by judicial authorities; (b) influencing elections / referenda.

ScenarioCommission verdictSection
AI tool assisting a judge in legal-research synthesis that feeds the substantive analysis of a caseHigh-risk under Annex III(8)(a)3.8.2
AI tool used by court administrative staff to retrieve case law or schedule hearings (no judicial-analysis output)Not Annex III(8)(a); record-keeping / administrative use3.8.2
AI tool drafting procedural documents submitted to a judicial authority by a partyNot Annex III(8)(a) where the AI is used by a party, not by the judicial authority3.8.2
AI political-targeting tool used during an election campaignHigh-risk under Annex III(8)(b)3.8.3

Annex I sectoral examples

The Commission's Annex I annex (May 2026 draft) provides illustrative examples per sector. Note Article 6(1) requires both cumulative conditions — Annex I legislation applies AND third-party conformity assessment is required.

Sector + sectoral lawExample AI systemHigh-risk under Article 6(1)?Section
Medical devices (Regulation 2017/745) — Section A NLFAI imaging diagnostic for class IIa or higher devicesYes (third-party CA required)Annex I annex
Medical devices (Regulation 2017/745) — Section A NLFAI-enabled class-I medical device (self-declaration)Not high-risk on Article 6(1); third-party CA condition not metAnnex I annex
Machinery Regulation — Section A NLFAI safety function controlling collaborative robot motionYes where third-party CA applies under the Machinery RegulationAnnex I annex
Toy Safety Directive — Section A NLFAI-enabled connected toy with safety functions requiring CAYesAnnex I annex
Lifts Directive — Section A NLFAI-controlled lift safety functionYes (third-party CA required)Annex I annex
Motor-vehicle type-approval — Section BAI autonomous-driving component subject to UNECE / type-approvalYes (type-approval is third-party CA equivalent)Annex I annex
Civil aviation (Regulation 2018/1139) — Section BAI safety-critical component on a manned aircraftYes via EASA certificationAnnex I annex
Consumer smart-home appliance with AI features under Low Voltage DirectiveTypically not high-risk on Article 6(1); third-party CA not required for self-certification regimeAnnex I annex

Settled vs vague (per draft Commission guidance)

Settled by the draft examples:

  • Anti-cheat / behaviour-monitoring during exams in education institutions is Annex III(3)(d); combined with emotion inference, may flip into prohibited Article 5(1)(f).
  • Fraud-detection AI is carved out of Annex III(5)(b) where the system targets fraud, not creditworthiness.
  • CV-screening + recruitment-AI is Annex III(4)(a); profiling typically applies.
  • Post-RBI for law-enforcement evidence is Annex III(1)(a); real-time RBI for law-enforcement is prohibited under Article 5(1)(h).
  • Maritime navigation safety is outside Annex III(7); migration-related border AI is within.

Left vague:

  • The exact boundary of "preparatory task" under Article 6(3)(d) — many of the draft's "outside scope" examples turn on whether the AI sits in a substantive-decision step.
  • The handling of mixed-use systems that are both Annex III and Annex I (e.g. emotion-recognition AI embedded in a medical device).
  • Treatment of consumer-facing AI products that fall under multiple Annex I acts.
  • The interplay between Annex III(5)(b) credit scoring and sectoral CRR/IDD obligations is under section 3.5.5; the Commission's draft says they layer rather than conflict, but the operational mechanics are not always concrete.

How to operationalise worked examples in Modulos

Each Annex III sub-point is part of the Modulos system-classification questionnaire associated with MFF-1. The classification rationale (which sub-point, the intended-purpose statement, the profiling determination, any Article 6(3) filter assessment) is owner-authored documentation stored as evidence on MRF-38 (with MRF-111 carrying the documentation + EU-database registration evidence where the filter is invoked).

Use caseModulos requirementCode
Annex III sub-point classification (any of 1–8)AI System Classification (app)MRF-38
Article 6(3) filter self-assessmentAI System Classification Exemption (app)MRF-111
Profiling determination (Article 4(4) GDPR)AI System Classification (app)MRF-38
Article 49(2) EU-database registration (for non-high-risk Annex III determination)AI System Classification Exemption (app)MRF-111

For the full classification framework + decision flow see High-risk classification. For the Modulos rollout see Operationalizing the EU AI Act in Modulos.

Source attribution

Draft Commission Guidelines on the classification of high-risk AI systems under Article 6 of the EU AI Act for stakeholder consultation, May 2026 — worked-examples portions of the Annex I annex and Annex III annex. © European Union. Draft guidelines issued pursuant to Article 6(5) AI Act; not yet adopted. Section references on this page (e.g. "section 3.4.2") are to the numbered sections of the Annex III annex. Worked examples on this page are paraphrases of Commission examples, not verbatim quotations; the Commission's draft warns at paragraph (118) that examples are illustrative and classification turns on system-specific architecture, functionality and intended purpose. Draft — consultation closes 23 June 2026; text may change before formal adoption. The underlying Regulation (EU) 2024/1689 is published in the OJ L of 12 July 2024 under CELEX 32024R1689.

Disclaimer

This page is for general informational purposes and does not constitute legal advice.