EU AI Office Enforcement Priorities for 2026: What Signals Say

The EU AI Office, established within DG CNECT and operational since mid-2024, has not published a formal enforcement plan. Its priorities can be inferred from staffing decisions, technical-paper publications, member-state coordination, and the few public statements its leadership has made. For organizations preparing for active AI Act enforcement, those inferences are the most useful signal available. This is the working read as of May 2026.

What the AI Office actually enforces

The AI Office has direct enforcement authority over general-purpose AI (GPAI) models, including those with systemic risk. Article 88 designates the AI Office as the supervisory authority for GPAI models. Member-state market surveillance authorities (designated under Article 70) handle high-risk AI systems and most Article 50 transparency violations.

For most organizations, the relevant enforcement path is through the member-state authority. The AI Office’s direct path matters primarily for foundation model providers.

Signal 1: Staffing emphasizes evaluation and technical assessment

The AI Office’s staffing profile in 2025-2026 has heavily weighted technical evaluation roles: AI safety engineers, evaluation methodologists, red-teamers. This signals that early GPAI enforcement will turn on substantive technical assessment of model capabilities rather than purely documentary review.

The practical implication: providers of GPAI models with systemic risk should expect the AI Office to run its own evaluations, not just review provider-submitted documentation. The provider’s evaluation methodology, model cards, and risk assessments will be tested against AI Office findings.

Signal 2: Code of Practice as soft enforcement

The AI Office has invested heavily in the GPAI Code of Practice (published in stages 2024-2026). The code is voluntary, but adherence creates a presumption of compliance under Article 56(7).

Conversely, non-adherence is not a violation but invites scrutiny. The AI Office has signaled that providers that have declined to adhere to the code will be subject to additional information requests and assessment. This is soft enforcement: not formal action, but resource-intensive engagement that few providers will want.

Signal 3: Transparency obligations are the early target

Article 50 obligations are the most likely site of early enforcement action by member-state authorities. The reasoning:

• Article 50 obligations are objective (disclosure present/absent) rather than substantive (was the risk assessment good enough).
• Many deployments demonstrably lack the required disclosures.
• The political case for transparency-focused enforcement — protecting users from undisclosed AI interaction — is the strongest.
• Sanctions under Article 99 are graduated, allowing authorities to pursue settlements rather than maximal fines.

Expect the first publicly-reported Article 50 enforcement actions in the second half of 2026, primarily from larger member-state authorities (France’s CNIL acting in coordination, Germany’s BfDI, the Netherlands’ Autoriteit Persoonsgegevens, Spain’s AEPD, Italy’s Garante).

Signal 4: Coordination with national DPAs

The EDPB-EDPS joint opinion on the AI Act and subsequent EDPB statements have framed AI Act enforcement as inseparable from GDPR enforcement. Member-state DPAs that already have AI Act competence have signaled they will use existing GDPR investigative tools to address AI Act issues.

Practical implication: AI Act compliance is unlikely to be assessed in isolation. An organization in regulatory engagement on GDPR grounds should expect AI Act questions to be added to the same investigation.

Signal 5: The codes of conduct under Article 95

Article 95 contemplates codes of conduct for non-high-risk AI systems, addressing voluntary application of high-risk requirements. The AI Office and DG CNECT have indicated interest in promoting sectoral codes (financial services, healthcare, education, employment).

Implication: organizations in those sectors should expect either to participate in code development or to be assessed against codes when they emerge.

Substantive priorities, as inferred

Combining the signals, the inferred substantive priorities for 2026 are:

Priority 1: GPAI compliance for the largest model providers

The AI Office’s first wave of substantive engagement targets the providers of the most-deployed GPAI models. Expectations: detailed risk assessments, evaluation suites, watermarking implementations, content-provenance commitments, copyright disclosures.

Priority 2: Article 50 disclosure compliance across deployers

Across all member states, the second wave targets deployers of chatbots, synthetic-content generators, and emotion-recognition systems that lack the required user-facing disclosures.

Priority 3: High-risk classification and conformity assessment

For systems classified as high-risk under Annex III (employment, education, credit, law enforcement, etc.), conformity assessment compliance becomes the focus through 2026-2027. The Notified Body landscape is still maturing; expect enforcement to focus on the conformity-assessment process itself before moving to substantive technical compliance.

Priority 4: Foundation models with systemic risk

For GPAI models meeting the systemic-risk threshold (Article 51), substantive obligations under Articles 55-56 (risk evaluation, adversarial testing, cyber-protection, incident reporting). The AI Office’s technical-staff capacity will be the gating factor on how many such providers can be assessed in detail.

What this implies for compliance prioritization

For organizations now budgeting AI compliance work, the prioritization implied by these signals:

1. Article 50 disclosure is the most likely early enforcement contact and the easiest to address comprehensively. Should be the highest near-term priority for most organizations.
2. DPIA / FRIA quality is the most likely topic if you’re subject to regulatory engagement on GDPR grounds; investing in substantive (not boilerplate) assessments pays off here.
3. Conformity assessment readiness for high-risk systems is the longest-lead-time item; planning should start now even where deployment is 12+ months out.
4. GPAI provider relationships matter — your provider’s compliance posture affects your own. Choose providers that publish credible evaluations and engage with the Code of Practice.

What is unlikely to be enforcement-relevant in 2026

The areas the AI Office is unlikely to prioritize in 2026, based on signaling:

• Prohibited practices under Article 5 — these are rare in commercial deployment and will be addressed when they arise.
• Detailed substantive review of every high-risk system — capacity does not exist.
• Enforcement against small deployers absent harm — political and resource constraints disfavor this.

The areas where enforcement may surge:

• A widely-publicized incident triggers fast-track action against the responsible party
• A regulator coordination action against multi-jurisdictional deployments
• An EU Court referral establishes interpretive precedent

Cross-references

For the Article 50 obligations that are most likely to be the early enforcement target, see Article 50 transparency obligations. For the Article 52 disclosure obligations for foundation models, see the Article 52 disclosure checklist. For the engineering-side controls that operationalize transparency obligations, mlobserve.com’s observability patterns ↗ cover the relevant infrastructure. For the incident-reporting context that will inform AI Office enforcement decisions, aiincidents.org’s source verification ladder ↗ explains how the public record is built.

The shape of the engagement

The AI Office’s early enforcement will be measured. It is staffing for the long game — building evaluation capacity, coordinating with member states, establishing doctrine through engagement with the largest providers. For organizations now preparing, the signal to take seriously is that compliance is becoming a body of substantive practice, not a paperwork exercise. The earliest enforcement actions will set expectations for years.

For more context, AI policy watch ↗ covers related topics in depth.