Enforcement, Penalties, and the Obligation-to-Control Crosswalk

This article closes the credential by walking the crosswalk Article-by-Article, by setting out the enforcement architecture that will exercise the obligations, and by placing the penalty bands of Article 99 in context. Foundational context is in the existing COMPEL article “EU AI Act Risk Classification: A Practitioner’s Guide” (Article ID 253); this article extends that foundation into the operational control-design workflow.

Enforcement architecture

Three institutional tiers exercise the Act’s enforcement:

1. National competent authorities — Article 70. Each Member State designates at least one market-surveillance authority and one notifying authority. Market-surveillance authorities enforce Chapters II (prohibitions) and III (high-risk) at national level. Notifying authorities designate and oversee the notified bodies performing third-party conformity assessments.
2. AI Office and AI Board — Article 64 and the Chapter VII institutional provisions. The AI Office sits within the European Commission and coordinates GPAI enforcement, supports the AI Board, and produces guidance. The AI Board brings together Member-State representatives and coordinates cross-border enforcement.
3. Notified bodies — designated under Article 29 and following. Notified bodies perform third-party conformity assessment under Annex VII for biometric systems and integrate with existing product-safety notified bodies under Annex I legislation.

National competent authority designations progressed unevenly through 2024–2025. Spain led with the formal establishment of the Agencia Española de Supervisión de la IA (AESIA) — the first dedicated national AI supervisory authority in the Union, operational from September 2024 under Royal Decree 729/2023. AESIA is instructive as a template: a dedicated authority with mixed technical and legal expertise, operational independence, and explicit mandate to cooperate with the AI Office. Other Member States adopted variations — Italy designated AgID and ACN; Ireland extended its Data Protection Commission mandate; Germany structured around the Federal Ministry of Economic Affairs and Climate Action with Länder coordination. Source: Spanish State Gazette BOE-A-2023-18911, https://www.boe.es/buscar/doc.php?id=BOE-A-2023-18911 .

Article 99 penalty bands

Article 99 sets three administrative-fine bands for infringements of the Act:

The turnover-percentage cap is the binding constraint for large organisations; the fixed-sum cap binds for smaller. SMEs and start-ups face a cap at the lower of the two alternatives in each band per Article 99(6) — the inverse of the large-entity rule.

The top-band alignment with Article 5 prohibitions is the enforcement signal that the prohibitions are the regulator’s highest priority. The middle-band covering the bulk of substantive obligations (Articles 9–15, 26, 50–55, 72–73) is the band within which most classification work will generate exposure; a classification error that leads to non-application of Article 9 risk management or Article 10 data governance is a middle-band event.

The obligation-to-control crosswalk — method

The crosswalk proceeds Article-by-Article. For each Article obligation, the specialist identifies:

1. The NIST AI RMF function and subcategory that most closely corresponds (GOVERN, MAP, MEASURE, MANAGE at function level; numbered subcategories within each).
2. The ISO/IEC 42001:2023 clause and Annex A control reference (Clauses 4–10 for the AMS structure; Annex A controls organised under eight objectives).
3. The operational control artefact the organisation produces as evidence.
4. The stored-record location and retention requirement.

The crosswalk is populated once per system, maintained alongside the classification register and the obligation register, and reviewed at each substantive change. Below is the working crosswalk for the substantive obligations Articles 9–15 and Article 26, followed by crosswalks for Articles 50, 53, 72, and 73.

Article 9 — risk management system

Article 10 — data and data governance

Article 11 and Annex IV — technical documentation

Article 12 — record-keeping

Article 13 — transparency and information to deployers

Article 14 — human oversight

Article 15 — accuracy, robustness, cybersecurity

Article 26 — deployer obligations

Article 50 — transparency to natural persons

Article 53 — GPAI provider baseline

Article 72 — post-market monitoring

Article 73 — serious incident reporting

The Italian Garante ChatGPT €15M fine — architectural teaching case

The Italian Garante’s ChatGPT enforcement decision of 30 December 2024 (Provvedimento 551 of 2024, €15M fine against OpenAI) is the highest-profile post-AI-Act-publication European AI enforcement action. The decision is formally under the GDPR rather than the AI Act; at the time of the decision, the AI Act’s operative provisions for GPAI had not yet come into effect. Architecturally, however, the enforcement pattern previews how the AI Act will be enforced:

• The Garante framed failures across transparency to users, lawful basis for training-data processing, child-safety measures, and age-verification design. Each maps to a future AI Act obligation: Article 50(1) transparency to users; Article 10 data governance; Article 53 training-content disclosure; and Article 5 / Article 50 age-sensitive protections.
• The €15M headline figure is the kind of number Article 99 middle-band infringements will produce. The Garante’s decision signals that EU supervisors will treat AI-specific obligations with the same quantitative seriousness they have applied to GDPR.
• The procedural pattern — provisional order, cooperation, compliance changes, final fine — is the template the specialist will see repeated under the AI Act’s Chapter VII procedures.

For the specialist, the Garante decision is not an AI Act enforcement but is the closest predictive analogue. Source: Italian Garante, “ChatGPT: Garante privacy sanziona OpenAI,” https://www.garanteprivacy.it/home/docweb/-/docweb-display/docweb/10085455 .

The EDPB Opinion 28/2024 overlap

The EDPB’s Opinion 28/2024 on certain data-protection aspects relating to the processing of personal data in the context of AI models (December 2024) clarifies the interaction between GDPR and GPAI regulation. For the specialist, three points from the Opinion inform the AI Act obligation-to-control crosswalk:

• Whether a GPAI model “embeds” personal data is a factual question; the GDPR’s applicability to the model itself depends on the answer. This informs Article 10 data-governance obligations and Article 53 copyright-compliance duties.
• Lawful basis for training on personal data must be assessed independently of AI Act compliance. Article 10(5) allows special-category processing for bias correction under narrow conditions; GDPR lawful-basis analysis applies throughout.
• The controller of model training and the controller of model deployment may differ; the specialist’s role register (see Article 1 of this credential) should reflect this. Source: EDPB Opinion 28/2024, https://www.edpb.europa.eu/our-work-tools/our-documents/opinion-board-art-64/opinion-282024-certain-data-protection-aspects_en .

CEN-CENELEC harmonised standards

The CEN-CENELEC JTC 21 work programme produces harmonised standards that give the crosswalk its concrete benchmarks. Standards covering risk management, data and data governance, transparency, human oversight, accuracy/robustness/cybersecurity, and AI management system are progressively finalising through 2025–2027. As each standard is listed in the Official Journal, the specialist updates the crosswalk with the specific standard clause that provides the presumption of conformity for the relevant Article obligation. Until standards are listed, the specialist uses ISO/IEC 42001 Annex A controls as the interim reference, documenting which specific clauses are relied on.

Diagram — MatrixDiagram

This article is accompanied by a MatrixDiagram. Rows are EU AI Act obligations — Articles 9, 10, 11, 12, 13, 14, 15, 26, 50, 53, 72, 73. Columns are NIST AI RMF functions (GOVERN, MAP, MEASURE, MANAGE) and ISO/IEC 42001 clauses and Annex A control families (A.2 policies, A.6 impact assessment, A.7 data, A.8 information, A.9 security, A.10 human oversight). Cells show the primary mapping. The matrix compresses the eleven-obligation crosswalk into a single-image reference the specialist can annotate per-system. Methodology Lead reviewers require the matrix to be paired with the worked per-system crosswalk narrative in this article.

Cross-references

• EATE-Level-3/M9.1-Art01-NIST-AI-RMF-ISO-42001-Crosswalk.md — enterprise-level crosswalk treatment; this specialist article is the practitioner-applied version.
• EATP-Level-2/M2.6-Art15-NIST-AI-RMF-Alignment-with-COMPEL-Stages.md — COMPEL-stage alignment.
• EATP-Level-2/M2.6-Art14-ISO-42001-Implementation-Using-COMPEL.md — ISO/IEC 42001 implementation guide.
• EATF-Level-1/M1.5-Art18-The-Regulatory-Convergence-10-Requirements-Every-Framework-Shares.md — the convergence article situating the EU AI Act against NIST and ISO.
• Existing regulatory article regulatory-compliance-articles.ts Article ID 253, “EU AI Act Risk Classification: A Practitioner’s Guide” — four-tier foundation from which the specialist extends into enforcement and crosswalk.

Learning outcomes — confirm

A specialist who completes this article should be able to:

• Explain the enforcement architecture, naming the roles of the national competent authority, the AI Office, the AI Board, and notified bodies.
• Classify at least ten obligation clauses against corresponding NIST AI RMF subcategories and ISO/IEC 42001 Annex A controls, using the matrix in this article.
• Evaluate an organisation’s evidence pack for alignment to Article 11 + Annex IV, Article 72 post-market monitoring, and Article 73 incident reporting.
• Design a single-page obligation-to-control crosswalk for a high-risk system using ISO/IEC 42001 as the operational backbone.

Quality rubric — self-assessment

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