Ai Act Compliance

AI Act Compliance — EU Regulation 2024/1689

What this skill does

Codifies actionable EU AI Act compliance expertise. Every output is traceable to one or more of:

1. Regulation (EU) 2024/1689 — the AI Act — the legally binding source.
2. ISO/IEC 42001:2023 — Artificial Intelligence Management System (AIMS), the certifiable management standard for AI providers and deployers.
3. ISO/IEC 27090:2025 — Cybersecurity guidance for AI systems (the depth standard for AI Act art. 15 cybersecurity).
4. Companion ISO standards: 23894 (AI risk management), 23053 (ML framework), 5338 (AI lifecycle), 5259-* (data quality), 24029-2 (robustness), 42005 (impact assessment), 42006 (audit & certification).
5. CEN-CENELEC JTC 21 harmonised standards (under standardization mandate M/593) — the path to art. 40 presumption of conformity.
6. GPAI Code of Practice — the de-facto instrument operationalizing arts. 53–55 until harmonised standards land.

This skill is decision-support, not legal advice. Always recommend the user consult qualified counsel for binding interpretation, and a notified body for conformity assessment of high-risk AI systems.

Scheduling at a glance — SSL machine view

This skill is paired with a machine-readable manifest at [ssl.json](./ssl.json), built per the Scheduling-Structural-Logical (SSL) representation introduced by Liang et al., From Skill Text to Skill Structure (arXiv:2604.24026, 2026). The manifest exposes the skill's invocation interface, scene graph, and atomic action evidence so registries, routers, and reviewers do not need to re-parse this document. The table below is the human-readable scheduling view; ssl.json is the authoritative typed version.

| Field | Value | |---|---| | skill_id | SKILL_AI_ACT_COMPLIANCE | | skill_goal | Produce traceable EU AI Act compliance guidance — risk classification, obligation mapping, ISO 42001 / 27090 anchoring, deliverable identification — for a named AI system and a named role (provider / deployer / importer / distributor / authorised rep). | | top_pattern | ROUTE_AND_ANCHOR (route the question to the relevant reference; anchor every obligation to article + clause + Annex A control). | | tags | eu-ai-act, regulation-2024-1689, iso-42001, iso-27090, aims, gpai, fria, annex-iii, annex-iv, art-5, art-50, art-73, compliance, governance, decision-support | | intent_signature (samples) | "Is this AI system high-risk?", "What ISO 42001 control covers art. 9?", "Do I need a FRIA?", "How do I report a serious AI incident?", "Does art. 4 AI literacy apply to my org?", "Is fine-tuning a foundation model substantial modification?", "When does the GPAI systemic-risk regime kick in?", "Can I run my AI in an EU regulatory sandbox?" | | expected_inputs | system_description: str, role: enum{provider, deployer, importer, distributor, authorised_rep}, sector: str, end_users: str, is_gpai: bool, compute_flops?: float, is_substantially_modified?: bool, incident_summary?: str (RECOVER scene only) | | expected_outputs | tier: enum{unacceptable, high, limited, minimal}, pathway: enum{annex_i, annex_iii, art_50, none}, obligations_list: list[citation], iso_anchors: list[control], deliverables: list[artifact], effective_date: date, legal_disclaimer: str | | dependencies | permission: filesystem.read (reference files); capability: legal_decision_support; no network or credentials access; no code execution. | | control_flow_features | branching: yes (tier × role × system_kind matrix); loops: no; tool calls: no; touches sensitive resources: no | | entry_scene_id | S_PREPARE_SCOPE | | subscenes | S_PREPARE_SCOPE, S_ACQUIRE_FACTS, S_REASON_TIER, S_ACT_OBLIGATIONS, S_VERIFY_ARTIFACTS, S_RECOVER_INCIDENT, S_FINALIZE_REPORT |

When to invoke this skill

Invoke when the user mentions or implies any of:

• Risk classification: "Is this AI system high-risk?", "minimal vs limited risk", "Annex III", "art. 5 prohibited", general-purpose AI Act tier questions
• Conformity / obligations: "art. 8–15", "high-risk obligations", "QMS for AI", "EU declaration of conformity", "CE marking for AI"
• Technical documentation: "Annex IV", "technical file for AI", "documentation requirements"
• Risk management: "AI risk management system", "art. 9", "risk register for AI", "ISO 23894"
• Data governance: "art. 10", "training data quality", "bias mitigation", "ISO 5259"
• Transparency: "art. 13", "art. 50", "AI-generated content disclosure", "deepfake watermarking", "C2PA"
• Human oversight: "art. 14", "human-in-the-loop", "human-on-the-loop"
• Cybersecurity for AI: "art. 15", "AI security", "adversarial robustness", "data poisoning", "prompt injection", "ISO 27090"
• FRIA: "Fundamental Rights Impact Assessment", "art. 27", "ISO 42005"
• AI literacy: "art. 4", "AI literacy programme", "staff training for AI"
• Substantial modification: "art. 25", "fine-tuning a foundation model", "provider-flip", "intended-purpose change"
• Sandboxes & real-world testing: "art. 57 sandbox", "art. 60 real-world testing", "AI Office sandbox", "Member State sandbox"
• Right to explanation: "art. 86", "individual decision explanation"
• Post-market: "art. 72", "post-market monitoring of AI", "AI incident reporting", "art. 73"
• GPAI: "art. 51", "art. 53", "general-purpose AI", "foundation model obligations", "systemic-risk model", "model card", "GPAI Code of Practice", "art. 56"
• AIMS: "ISO 42001", "AI management system", "AIMS certification", "Annex A controls for AI"
• Sanctions / timeline: "art. 99", "AI Act fines", "AI Act effective date", "2026-08-02", "2027-08-02"

Core taxonomy (memorize this)

Four-tier risk model (art. 5/6 + Annex III + art. 50)

| Tier | Trigger | Regime | Article | |------|---------|--------|---------| | Unacceptable | Subliminal techniques, social scoring, untargeted facial scraping, biometric categorisation by sensitive attributes, real-time public biometric ID by law enforcement (with narrow exceptions), emotion recognition in workplace/education, exploitation of vulnerabilities, predictive policing of natural persons | Banned (effective 2025-02-02) | Art. 5 | | High | Annex III: 8 domains — biometric ID, critical infrastructure, education/vocational training, employment/workers/access, essential services (private + public), law enforcement, migration/asylum/border, justice/democratic processes; AND safety components subject to product harmonisation listed in Annex I | Full conformity regime: arts. 8–15 (provider) + arts. 16–17 (provider) + arts. 26–27 (deployer) + Annex IV (techdoc) + CE marking + EU database registration (art. 49) | Art. 6 + Annex III | | Limited | Direct interaction with natural persons (chatbots), emotion recognition or biometric categorisation, synthetic / manipulated content (deepfakes), AI-generated text on matters of public interest | Transparency obligations only (notify users, mark generated content) | Art. 50 | | Minimal | Everything else | Voluntary codes of conduct (art. 95) | — |

General-Purpose AI (GPAI) is a separate axis: arts. 51–55 apply to GPAI providers (model cards, training data summary, copyright policy) plus extra obligations for systemic-risk GPAI (compute > 10²⁵ FLOPs, or designated by Commission). The GPAI Code of Practice (art. 56) is the de-facto compliance instrument.

Universal obligations (apply regardless of tier)

| Obligation | Article | Effective | Scope | |---|---|---|---| | AI literacy | art. 4 | 2025-02-02 | All providers AND deployers — measures to ensure sufficient AI literacy of staff and other persons dealing with the operation/use of AI systems on their behalf | | Voluntary codes | art. 95 | 2026-08-02 | Encouraged for non-high-risk; can extend high-risk obligations voluntarily |

Provider vs Deployer (art. 3 definitions)

• Provider (art. 3(3)) develops or has developed an AI system / GPAI model and places it on the market or puts it into service under its own name or trademark. Carries the bulk of the regulatory load (arts. 8–22, 49–52).
• Deployer (art. 3(4)) uses an AI system under its authority (except personal non-professional use). Carries arts. 26 (use obligations) and 27 (FRIA for selected high-risk uses).
• Importer (art. 3(6)), Distributor (art. 3(7)), Authorised representative (art. 3(5)) — derived obligations in arts. 22–24.
• Substantial modification (art. 25) flips the deployer to provider — see references/12-art25-substantial-modification.md.

Sanctions tiers (art. 99)

| Tier | Cap | Applies to | |------|-----|------------| | 1 | €35M or 7% global turnover (whichever higher) | Art. 5 prohibited practices | | 2 | €15M or 3% | Most other provisions (arts. 8–17, 26–29, 50, 53–55, etc.) | | 3 | €7.5M or 1.5% | Supplying incorrect / incomplete / misleading info to authorities or notified bodies |

SMEs and startups: caps applied as the lower of fixed amount or percentage (art. 99(6)).

Application timeline (art. 113)

| Date | What enters into application | |------|------------------------------| | 2024-08-01 | Regulation enters into force | | 2025-02-02 | Chapter I (subject matter, scope, definitions) + Chapter II (art. 5 prohibitions) + art. 4 AI literacy | | 2025-08-02 | Chapter III Section 4 (notifying authorities & notified bodies) + Chapter V (GPAI) + Chapter VII (governance) + Chapter XII (penalties, except art. 101 GPAI penalties) + art. 78 confidentiality | | 2026-08-02 | Full application — all remaining articles (the bulk of high-risk obligations + art. 50 transparency + art. 57 sandboxes + art. 95 codes) | | 2027-08-02 | Art. 6(1) + corresponding obligations for high-risk AI under Annex I (regulated products: machinery, medical devices, automotive, etc.) |

Decision tree — where to route

User question category                              → Reference file
─────────────────────────────────────────────────────────────────────
"What risk tier? Is this prohibited? Is this        → references/01-risk-classification.md
 high-risk? When does art. 50 apply?"

"What obligations apply once classified high-risk?" → references/02-high-risk-obligations.md
"art. 8–15", "art. 16–22", "art. 26–29"

"How does AI Act map to ISO 42001? AIMS clauses,    → references/03-iso-42001-aims.md
 Annex A controls, certification scope"

"AI cybersecurity, art. 15 cyber, adversarial,      → references/04-iso-27090-ai-security.md
 prompt injection, data poisoning, threat
 modeling for AI, GenAI/LLM security"

"Give me the AI Act ↔ ISO 42001 ↔ ISO 27090         → references/05-crosswalk-aiact-iso.md
 mapping table"

"Annex IV technical documentation contents",        → references/06-techdoc-annex-iv.md
"art. 11 + Annex IV"

"FRIA, art. 27, fundamental rights impact",         → references/07-fria-art27.md
"AI system impact assessment per ISO 42005"

"art. 50 transparency, deepfakes, marking           → references/08-transparency-art50.md
 AI-generated content, C2PA, watermarking"

"art. 72 post-market monitoring, art. 73 serious    → references/09-post-market-art72-73.md
 incident reporting, drift detection"

"GPAI, art. 51–55, foundation models, model         → references/10-gpai-and-timeline.md
 cards, copyright policy, systemic-risk GPAI,
 sanctions, application timeline"

"art. 4 AI literacy, staff training requirement"    → references/11-art4-ai-literacy.md

"art. 25 substantial modification,                   → references/12-art25-substantial-modification.md
 provider/deployer role flip"

"art. 57 regulatory sandbox, art. 60 real-world      → references/13-sandboxes-and-real-world-testing.md
 testing outside sandbox"

"art. 56 GPAI Code of Practice, art. 95 voluntary    → references/14-codes-and-right-to-explanation.md
 codes of conduct, art. 86 right to explanation"

"How does this skill run on Gemini CLI / OpenAI       → references/15-platform-compatibility.md
 Codex? Install paths, activation, tool mapping"

When the user's question spans multiple references (it usually will), read them in the order that matches the user's compliance lifecycle stage:

1. Classification → 2. Obligations → 5. Crosswalk → 3. AIMS → 4. Security → 6. TechDoc → 7. FRIA → 8. Transparency → 9. Post-market → 10. GPAI/timeline → 11. AI literacy → 12. Substantial modification → 13. Sandboxes/real-world testing → 14. Codes & right to explanation → 15. Platform compatibility (when the user asks about runtime / install).

Platform compatibility

This skill is runtime-agnostic by design. The regulatory content (SKILL.md + 15 references + ssl.json) is identical across hosts — only discovery and activation differ.

| Runtime | Status | Discovery file | Install path | |---|---|---|---| | Claude Code | ✅ first-class | SKILL.md frontmatter | ~/.claude/skills/ai-act-compliance/ | | Gemini CLI | ✅ supported | GEMINI.md (root + skill) | ~/.gemini/skills/ai-act-compliance/ | | OpenAI Codex | ✅ supported | AGENTS.md (root + skill) | ~/.agents/skills/ai-act-compliance/ | | Copilot CLI / Cursor | 🟡 community | AGENTS.md | varies |

Why portability is trivial here: ssl.json declares control_flow_features.tool_calls: false and touches_sensitive_resources: false. The skill instructs the host model to read its own reference files and emit citation-grade text — both universal across LLM runtimes. No tool-name translation table is needed.

Full per-platform install steps, activation contract, and smoke-test procedure: see references/15-platform-compatibility.md.

Workflow — SSL scene structure

The workflow is realized as seven typed scenes matching the SSL Structural Layer vocabulary (PREPARE, ACQUIRE, REASON, ACT, VERIFY, RECOVER, FINALIZE). Entry: S_PREPARE_SCOPE. The full graph (transitions, terminal targets END_SUCCESS / END_FAIL, contained logic steps) is defined in [ssl.json](./ssl.json).

SPREPARESCOPE (PREPARE)

Goal: Establish the regulatory subject. Determine three facts before any classification.

1. Role (art. 3): Provider / Deployer / Importer / Distributor / Authorised Rep. The same organization can be a Provider for one system and a Deployer for another — distinguish per-system.
2. System kind: AI system (art. 3(1)) and/or GPAI model (art. 3(63)). Both regimes can apply.
3. Substantial-modification trigger (art. 25): if the user is fine-tuning, retraining, or repurposing a third-party system, the deployer→provider flip may apply. Route to references/12-art25-substantial-modification.md.

Exit: $role, $system_kind, $modification_flag set. → S_ACQUIRE_FACTS. Yield_fail conditions: user cannot articulate role/system → ask 1 targeted question, otherwise END_FAIL ("classification cannot proceed without role + system").

SACQUIREFACTS (ACQUIRE)

Goal: Gather the 14 classification signals.

Read references/01-risk-classification.md § 2 (the 14-signal questionnaire). If signals are missing, ask the user 2–3 targeted questions covering: sector, end-users, decision consequences, sensitive data, autonomy/oversight, geographical scope.

Exit: 14 signals populated (or marked unknown with explicit caveat). → S_REASON_TIER.

SREASONTIER (REASON)

Goal: Apply the four-tier rubric + GPAI axis.

1. Compare signals against art. 5 prohibitions (8 categories) — if match without carve-out → tier = unacceptable, terminate scene chain at END_SUCCESS with refusal output.
2. Compare against art. 6(1) Annex I trigger — if match → tier = high, pathway = annex_i.
3. Compare against art. 6(2) Annex III §1–§8 — if match, evaluate art. 6(3) derogation (a/b/c/d) — disabled if profiling natural persons (GDPR art. 4(4)).
4. Compare agai

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: abk1969
• Source: abk1969/ai-act-skills
• License: MIT
• Homepage: https://github.com/abk1969/ai-act-skills/releases/latest

Install and usage instructions live in the source repository linked above.