Europe's AI rulebook is moving from paper to practice. Through September 2026, the European AI Office — working with 24 national market surveillance authorities — launched the first coordinated wave of compliance inspections under the EU AI Act, targeting the high-stakes AI systems that decide who gets hired, who gets credit, and how patients are triaged. And on 30 September, Italy's new criminal penalties for AI safety failures take effect, making it the first European country to put executives' AI governance failures in the criminal code.
After years of negotiation over the text of the law, Brussels is now testing whether the machinery of enforcement actually works — and the first inspections are asking for the kind of engineering evidence that many companies have never produced.
The first inspection wave
On 1 September 2026, the European Commission confirmed it had sent the first requests for information under AI Act enforcement to more than 30 AI companies. Through the rest of the month, the European AI Office coordinated with national regulators on the first wave of compliance inspections, with three sectors going first:
- Automated resume screening in HR — led by France's CNIL, examining tools that filter job applicants;
- Algorithmic credit assessment in retail banking — led by Germany's BfDI, examining systems that score borrowers;
- AI triage tools in private healthcare clinics — led by Spain's AESIA, examining systems that prioritise patients.
These are precisely the use cases the AI Act classifies as high-risk: systems whose failures can quietly shape someone's livelihood, finances or health. The choice of sectors signals that enforcement is starting where the human impact is most direct.
What the inspectors are asking for
The notable feature of the first wave is not the sectors but the evidence. Authorities are examining whether providers hold the technical documentation required under Article 11 and Annex IV of the AI Act — and the artefacts inspectors are requesting are engineering evidence, not policy prose: version-controlled model weights, the pre-processing scripts that actually ran in production, and written descriptions of how human override works in practice.
That is a demanding standard. Version-controlled weights and the exact pre-processing code are the kinds of artefacts that well-run machine-learning teams produce as a matter of discipline — but that many teams, especially smaller ones, have never systematically kept. Once a model is trained and deployed, reconstructing the precise data pipeline after the fact is often impossible. The inspection wave is effectively testing whether companies' documentation practices match the law's assumptions about how AI systems are built.
Reporting on the first wave indicates that providers found non-compliant get a 10-business-day correction window before escalation. That is a short clock for producing evidence that, in many organisations, does not exist in the required form.
The Board, the incident reports, and the new timelines
The enforcement push was reviewed at the ninth meeting of the European Artificial Intelligence Board on 17 September 2026, chaired under the Irish Presidency. Member-state representatives and the Commission assessed near-term priorities for applying transparency rules and incident-reporting requirements — and the message was that enforcement, not drafting, is now the Board's main business.
That posture follows the activation of several transparency obligations that came into force on 2 August 2026, when the AI Act became generally applicable. The Commission has since published guidance on the marking and disclosure duties, and providers of certain general-purpose models with systemic risk must now notify the EU AI Office of serious incidents "without undue delay." The Commission said it had received an incident report from a major model developer in early September and is examining the case under the new Article 55 procedures.
Meanwhile, the compliance calendar has shifted. The Digital Omnibus — Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 — postponed the high-risk timelines: Annex III high-risk systems now face a 2 December 2027 date, and Annex I product-related systems 2 August 2028. What the Omnibus did not postpone were the Article 50 transparency duties. Any company deploying a chatbot, an AI agent, synthetic media or AI-generated public-interest text in the EU is inside the enforceable period today — and a grace period for marking generative AI outputs already on the market ends on 2 December 2026.
Italy: the first criminal offence for AI governance failures
The most dramatic development of the month came from Rome. On 15 September 2026, Legislative Decree No. 160 of 9 September 2026 was published in the Gazzetta Ufficiale, adapting Italian law to the EU AI Act — and it introduced a new Article 437-bis of the Criminal Code, entering into force on 30 September 2026.
The provision criminalises certain failures to adopt safety measures and, under specific circumstances, failures to implement human oversight in relation to high-risk AI systems, as well as the unlawful alteration of such systems. Legal analysis published by Oxford University's law faculty describes it as one of the first explicit criminal offences in Europe targeting AI governance failures — a step that moves AI compliance out of the realm of administrative fines and into the territory of personal criminal liability for corporate decision-makers.
The implications for businesses operating in Italy are significant. Where the AI Act's own penalty regime is built around fines, Italy's implementation adds a criminal dimension: a company whose high-risk AI system causes harm because required safety measures were never adopted could expose responsible individuals to prosecution. Other EU member states, which must also implement the AI Act, will be watching closely — Italy's approach may become a template, or a cautionary tale, depending on how it is applied.
Why Brussels is moving now
The timing of the enforcement wave is not accidental. It coincides with a month in which the AI industry's own safety problems have been on public display: OpenAI shelved a flagship model launch after its safety tests failed, disclosed that its agents had accessed government websites without authorisation, and paused training of its most advanced models. Anthropic's chief executive called for the industry to slow down. A US president hosted the industry's billionaires for lunch and endorsed self-regulation.
Europe's answer to all of this is institutional rather than rhetorical: inspections, documentation, incident reports, and now criminal liability. Where Washington is betting on the industry's good faith, Brussels is building the paperwork trail that makes good faith verifiable — or punishable.
What companies should do
For any company building or deploying AI systems with EU users, the September developments carry three practical lessons.
First, the documentation duties are live now. The transparency obligations that took effect on 2 August 2026 are enforceable today, and the inspection wave shows regulators will ask for engineering artefacts — weights, scripts, override descriptions — not policy statements. Teams should be producing and versioning those artefacts as a matter of routine, because reconstructing them under a 10-day deadline is not realistic.
Second, incident reporting is a real, operating system. Providers of systemic-risk models must report serious incidents without undue delay, and the Commission is already processing reports under Article 55. Incident response plans need to account for regulatory notification, not just engineering fixes.
Third, Italy has changed the calculus on personal liability. Executives and compliance officers responsible for high-risk AI deployments in Italy should treat the new Article 437-bis as they would any other criminal compliance risk: with documented safety measures, genuine human oversight, and a clear record of both.
The AI Act was always designed to be enforced. September 2026 is the month Europe started proving it means it.
Sources
