EU AI Act High-Risk Classification

Depth assessment of whether an AI system is high-risk under Art. 6 AI Act (Regulation (EU) 2024/1689), grounded in the European Commission's draft Art. 6(5) classification guidelines (general principles, Annex I, Annex III) published for stakeholder consultation in 2026.

Disclaimer (show at session start, do not block)

Important: This skill provides structured high-risk classification guidance based on the EU AI Act (Regulation (EU) 2024/1689) and the Commission's draft Art. 6(5) classification guidelines (general principles, Annex I, Annex III) issued for stakeholder consultation. The Commission guidelines are non-binding; authoritative interpretation rests with the Court of Justice of the EU. This is not legal advice. Final classification decisions should involve qualified legal counsel with AI Act expertise.

When to use this skill

• The user has already concluded (via broad risk-tier triage or otherwise) that high-risk is the likely tier and needs the depth assessment.
• A high-risk verdict is consequential (FRIA may apply, Chapter III conformity assessment regime, post-market monitoring, registration in EU database).
• The user explicitly asks for Annex I or Annex III analysis.
• The user is uncertain whether a borderline AI system is high-risk and needs the Commission-guideline-grounded reasoning.

If the AI system has not yet been triaged across all five risk tiers (prohibited / high-risk / GPAI / limited / minimal), do that breadth-first triage before the Annex I / Annex III depth analysis below. For a pure article-lookup question without a classification verdict, answer it directly with the cited regulation text rather than running the full depth assessment.

Effective dates (per AI Omnibus 2026)

See references/ai-omnibus-timeline-postponements.md for citation chain.

Required inputs

Before beginning, gather from the user:

1. System description — what does it do, who built it, who uses it, what is the intended purpose stated by the provider.
2. Intended-purpose evidence — instructions for use, technical documentation, promotional materials, terms of service, sales materials. The provider's framing matters per ¶¶10–13 of the general-principles guidelines.
3. Deployment context — sector (machinery, medical, employment, education, law enforcement, etc.), is the user the provider, deployer, importer, or distributor.
4. Modification status — is the user the original provider or did they fine-tune / rebrand / substantially modify a third-party system (Art. 25 trap).

If any of these is missing, ask before proceeding. Classifications based on incomplete information must be flagged as preliminary.

Decision tree

Run the five steps in order. Each step either terminates with a verdict or feeds the next.

Step 1 — AI system gating (Art. 3(1))

Confirm the technology meets the Art. 3(1) AI system definition. If it does not (e.g., pure rule-based software with no inference, no learning, no autonomy), the AI Act does not apply → terminate as not in AI Act scope. Reference: see Commission Guidelines on the definition of an artificial intelligence system, C(2025) 5053 (separate from these high-risk guidelines).

Step 2 — Intended-purpose framing (Art. 3(12); general principles ¶¶10–13)

Document the provider's stated intended purpose. Apply the GPAI / multi-purpose trap test:

• If the provider's materials present the system as broadly applicable across many contexts and do not consistently exclude high-risk uses → the intended purpose is deemed to encompass high-risk uses (¶12).
• Merely asserting in terms-of-service that high-risk use is excluded is insufficient if other materials (promotional, examples, product positioning) suggest such uses are feasible and reasonably foreseeable (¶12).
• Capture but exclude from this assessment: Art. 3(13) "reasonably foreseeable misuse" — by definition outside intended purpose.

Output of Step 2: a clean, documented intended-purpose statement to use in Steps 3 and 4.

Step 3 — Art. 6(1) / Annex I branch

Apply this branch in full before Step 4. The two branches are not mutually exclusive: a system can be high-risk under both Art. 6(1) and Art. 6(2) (in which case the Art. 6(1) conformity-assessment integration applies — see Art. 8(2) and Art. 102–109).

3a. Is the AI system a product / safety component under Annex I?

Screen against the Union harmonisation legislation listed in Annex I AI Act (machinery, toys, lifts, ATEX, radio equipment, pressure equipment, recreational craft, cableways, gas appliances, MDR, IVDR, automotive, aviation, etc.). Distinguish:

• AI system is itself a regulated product (¶30): independently placed on market, has own intended purpose, directly regulated. Example: machinery-related software under Reg. (EU) 2023/1230.
• AI system is a safety component of a regulated product (¶31).

If neither applies → skip to Step 4.

3b. Safety-component two-prong test (Art. 3(14); ¶¶32–49)

Apply BOTH alternative scenarios — either is sufficient:

Scenario (i) — Safety function (intent-based, ¶¶35–37) The system constitutes a safety component if the provider's intended purpose is to prevent or mitigate risks to health, safety, or property. Use this checklist (from ¶37 box):

• Preventive functions (any one): monitoring/detection of harm-leading situations; maintenance/inspection detection; harm prevention; supervision of another safety system.
• Mitigation functions (any one): control or limitation of physical harm; mitigation of consequences (e.g., safe-stop); control of another safety system.
• NOT safety functions: performance optimisation, service efficiency, automation of user decisions, comfort/convenience, quality control of non-safety aspects.

See references/safety-function-checklist.md for the full taxonomy and worked examples.

Scenario (ii) — Failure or malfunctioning (consequences-based, ¶¶38–43) The system constitutes a safety component if its failure or malfunctioning could endanger health, safety, or property. Failure modes include: incorrect outputs (false positives/negatives), loss of function/availability, performance instability/drift, timing/latency errors, misclassification leading to hazardous control decisions. The likelihood must be more than theoretical (¶39).

Worked examples from ¶46:

• Lifts: AI for door closing / obstacle detection → safety component via failure (efficiency intent, but malfunction injures persons).
• Vehicles: lane-assist AI → safety component via failure.
• Agriculture: chemical-spray targeting AI → safety component via failure if persons nearby.
• Smart thermostat (¶48 footnote 5): NOT a safety component except where it controls child-lock or operates around vulnerable users.

Output of Step 3b: if (i) OR (ii) is satisfied → continue to Step 3c. Otherwise, this is not a safety component → skip to Step 4.

3c. Third-party conformity assessment requirement (¶¶50–59)

Look up the conformity assessment procedure required by the relevant Annex I legal act:

• Modules B, C1, C2, D, D1, E, E1, F, F1, G, H, H1 (Decision 768/2008/EC) → involve a notified body → third-party conformity assessment required → YES.
• Module A (pure internal control) without harmonised-standards condition → not third-party → NO.
• Module A conditioned on mandatory application of harmonised standards (e.g., Toys Safety Regulation, Machinery Regulation, Radio Equipment Regulation) → still classified as high-risk; the option to use Module A is procedural flexibility, not a discretion to escape high-risk classification (¶57; Toys Safety Reg. Recital 15).

If module is A without mandatory-harmonised-standards condition → not high-risk under Art. 6(1).

3d. Section A vs Section B distinction (Annex I; Art. 2(2))

If 3a + 3b + 3c all YES → high-risk under Art. 6(1). Record:

• Annex I citation (specific legal act).
• Whether the legal act is in Annex I Section A (NLF-based) — full Chapter III requirements apply.
• Or Section B (other Union harmonisation legislation, e.g., aviation, automotive) — only Art. 6(1), 102–109, 112 apply per Art. 2(2).

See references/annex-i-section-a-vs-b.md for the full mapping.

Step 4 — Art. 6(2) / Annex III branch

4a. Map to the eight Annex III areas

Cross-check the system's intended purpose against each of the eight areas. Use the corresponding per-area reference file:

For each area where there is a use-case match, capture the specific Annex III sub-point (e.g., Nr. 4(a) for recruitment).

4b. Apply the Art. 6(3) exception filter

If a use case matches in Step 4a, check whether the AI system performs only one of the following narrow categories (Art. 6(3)):

• (a) Narrow procedural task — strictly procedural, no influence on outcome.
• (b) Improvement of a previously completed human activity — refining what a human already decided.
• (c) Detection of decision-making patterns or deviations — without replacing or influencing previously completed human assessment.
• (d) Preparatory task to an assessment relevant to the Annex III use case.

See references/art-6-3-exception-decision-tree.md.

4c. Apply the profiling re-exception (Art. 6(3) last sentence)

If the system performs profiling of natural persons (Art. 4(4) GDPR — automated processing of personal data to evaluate personal aspects), the Art. 6(3) exception is excluded and the system is high-risk.

4d. Documentation duty if exception applies (Art. 6(4))

If the exception applies → not high-risk, BUT the provider must:

• Document the reasoning before placing on the market / putting into service (Art. 6(4)).
• Register the system in the EU database per Art. 49(2).

Step 5 — Art. 25 substantial-modification trap

If the user is not the original provider but is modifying, fine-tuning, or rebranding an existing AI system, flag Art. 25(1):

• (a) Putting their own name/trademark on an existing high-risk system.
• (b) Substantial modification to an existing high-risk system.
• (c) Modifying the intended purpose of a non-high-risk system (including GPAI) such that it becomes high-risk under Art. 6.

If any apply → the user becomes a provider of a high-risk system with full Chapter III obligations. Follow up with a role-determination step (provider / deployer / importer / distributor under Art. 3 and Art. 25). The Commission is preparing separate Art. 25 guidelines; this skill flags only.

See references/art-25-substantial-modification-flag.md.

Output artefacts

Produce all three artefacts unless the user explicitly requests a subset.

Artefact 1 — Structured decision block (terminal)

═══════════════════════════════════════════
EU AI Act — High-Risk Classification Result
═══════════════════════════════════════════
High-risk verdict:        [YES — Art. 6(1) Annex I | YES — Art. 6(2) Annex III Nr. X | NO]
Basis:                    [Safety function / Failure-based / Annex III use-case match]
Annex I citation:         [Legal act + Section A/B] (if applicable)
Annex III citation:       [Nr. X.<sub>] (if applicable)
Art. 6(3) exception:      [N/A | Applied — limb (a/b/c/d) | Excluded by profiling re-exception]
Art. 25 trap:             [Not triggered | Watch — modifying existing system]
Effective date:           [2 December 2027 (Annex III) | 2 August 2028 (Annex I)]
Documentation duty:       [Art. 6(4) (if Art. 6(3) applied) | Chapter III + Art. 11 technical docs]
═══════════════════════════════════════════

Artefact 2 — Practitioner memo (1–2 page narrative)

Markdown memo addressed to a DPO / AI compliance officer. Sections:

1. System under assessment — name, provider/deployer, intended purpose (the cleaned Step 2 statement).
2. Verdict and reasoning — which branch (6(1) or 6(2)), which sub-test triggered it, which paragraph of the Commission guidelines is the anchor.
3. Citations — explicit paragraph numbers from the Commission guidelines and the corresponding article numbers.
4. Next steps — flag the follow-on work the verdict triggers: role determination (provider / deployer / importer / distributor), mapping the applicable Chapter III obligations to that role, and producing a formal compliance report.
5. Open questions — anything that needs counsel input before finalisation.

Artefact 3 — JSON interchange artefact

The cross-skill interchange JSON. Schema (matches the existing repo convention):

{
  "skill": "ai-act-high-risk",
  "version": "1.0",
  "assessed_at": "2026-MM-DD",
  "system_name": "<provider-supplied name>",
  "intended_purpose": "<cleaned Step 2 statement>",
  "verdict": "high_risk" | "not_high_risk" | "exempt_art_6_3",
  "basis": {
    "article": "6(1)" | "6(2)" | null,
    "annex": "I" | "III" | null,
    "annex_item": "Nr. X" | "Nr. Y.<sub>" | null,
    "section_a_or_b": "A" | "B" | null
  },
  "art_6_3_exception": {
    "applied": true | false,
    "limb": "a" | "b" | "c" | "d" | null,
    "profiling_excluded": true | false
  },
  "art_25_trap_flag": true | false,
  "effective_date": "2027-12-02" | "2028-08-02",
  "citations": [
    {"source": "Commission Guidelines Annex III ¶123"},
    {"source": "Annex III Nr. 4(a) AI Act"}
  ],
  "next_steps": ["role-determination", "obligation-mapping", "compliance-report"]
}

References

• art-6-general-principles.md — Commission's general principles (¶1–14, ¶448–451).
• art-6-1-annex-i-guidelines.md — Commission's Annex I section (¶15–62).
• art-6-2-annex-iii-guidelines.md — Commission's Annex III section (¶63–447).
• safety-function-checklist.md — taxonomy of safety functions from ¶37 box.
• annex-i-section-a-vs-b.md — Section A (NLF) vs Section B (other) mapping.
• 8 × per-area Annex III files (see Step 4a table).
• art-6-3-exception-decision-tree.md — exception filter with worked examples.
• art-25-substantial-modification-flag.md — quasi-provider trap.
• ai-omnibus-timeline-postponements.md — postponement citation chain.

Follow-on activities

A high-risk depth assessment usually sits inside a larger workflow. Once this skill has produced the verdict, the natural next steps are:

• Broad risk-tier triage — if the system has not yet been screened across all five tiers (prohibited / high-risk / GPAI / limited / minimal), do that first.
• Article Q&A — for any AI Act article question, answer directly from the cited regulation text and Commission guidelines.
• Role determination — provider / deployer / importer / distributor, including the Art. 25 quasi-provider trap.
• Obligation mapping — once the high-risk verdict is in, map the Chapter III obligations by role and risk tier.
• Formal compliance report — assemble the classification rationale and obligation matrix into an audit-ready record.
• Preliminary triage — a 15–25 min rapid assessment is a lighter alternative when full Art. 6 depth is not yet warranted.

Changelog

See CHANGELOG.md.

Part of the EU AI Act suite

This skill works on its own, but it's designed to interlock with my other EU AI Act skills — install any individually, or use them together for an end-to-end workflow:

• EU AI Act Quick Assessment — 15–25 min preliminary triage
• EU AI Act System Classifier — risk-tier classification across all five tiers
• EU AI Act Role Determination — provider / deployer / importer / distributor (incl. Art. 25)
• EU AI Act Obligations Mapper — obligations by role and risk tier
• EU AI Act Examination Report Generator — audit-ready compliance report
• EU AI Act Knowledge Base — Q&A over the Act + Commission guidelines

Each is available as a separate skill — install only what you need.