Italy readies fines of up to €35 million as the EU delays its AI deadlines
The EU's Digital Omnibus pushed high-risk AI deadlines to December 2027 and August 2028, while Italy has already built out its enforcement apparatus with fines of up to €35 million.
Delayed, but not dismantled
In summer 2026, the EU's Digital Omnibus on artificial intelligence was adopted, revising the implementation timeline of the AI Act. Requirements for stand-alone high-risk systems are postponed to December 2027, while high-risk systems embedded in regulated products face an August 2028 deadline. The regulatory framework itself, however, remains intact (The National Interest).
That means the delay is a time shift, not a teardown. The categories, the obligations attached to them, and the allocation of responsibility along the value chain all survive; only the calendar moves. For companies hoping for a more sweeping loosening, the result is the same framework — just with more time. Law.com warns in a standfirst (the article body is behind a paywall) that American legal officers who assume they have plenty of time may be underestimating the compliance pressure — not least because the regulation applies wherever AI outputs are used in the EU, regardless of where the company is established. Since only the summary is visible, the article's other arguments cannot be verified here.
How the classification mechanics work
The AI Act is a risk-based framework that sorts systems into tiers. At one end sit prohibited practices, such as certain forms of social scoring and real-time biometric identification. At the other are minimal-risk systems, which are largely exempt. Between those poles lies the "high-risk" category: systems used in sensitive contexts such as law enforcement, migration management, employment screening, creditworthiness assessments, and access to public benefits. These are permitted, but must meet strict requirements for data governance, risk management, transparency, human oversight, and post-deployment monitoring (The Regulatory Review).
The framework also assigns obligations by role: providers, deployers who put the systems into use, importers, distributors, and other actors bear responsibility according to the system or model in question (Unite.AI). Classification and applicable deadlines thus depend on concrete facts, role, jurisdiction, and the official guidance that gradually develops — not on industry affiliation alone.
Italy builds the enforcement apparatus
With the publication of Legislative Decree No. 179 of October 7, 2026 and Decree No. 160 of September 9, 2026 in Italy's official gazette, the government has completed the implementation of the delegation under Article 24 of Law No. 132 of September 23, 2025, write the lawyers Claudio Chiarella, Francesco Gaggioli, Daniela Martini, and Veronica Rossetti of A&O Shearman (JDSupra).
The apparatus is concrete: AgID is designated as the notifying authority, while the national cybersecurity agency ACN becomes the market surveillance authority, contact point, and competent authority for national registration of the Annex III, point 2 high-risk systems — that is, systems used as safety components in the management and operation of critical infrastructure. Penalties are tiered up to €35 million, or 7 percent of global annual turnover, for prohibited practices.
This makes Italy an early test case for how the AI Act is enforced in practice. While the EU's common deadlines for high-risk requirements lie years ahead, the national authorities, contact points, and penalty levels already stand ready in at least one large member state.
The practice gap: No one accounted for fundamental rights
While the framework hardens on paper, findings from the EU Agency for Fundamental Rights (FRA) show that it has not yet reshaped practice. In a 2025 report, based on interviews across five member states, the FRA found that none of the interviewed providers or deployers of potential high-risk systems conducted risk assessments that comprehensively accounted for fundamental rights. The assessments instead emphasized privacy, technical metrics, or business and legal risk (The Regulatory Review).
The FRA also points out that high-risk systems are often rolled out in policy areas that hit certain groups disproportionately hard — migrants, persons with disabilities, and recipients of social benefits. Because these groups already occupy vulnerable positions, weaknesses in risk assessment or oversight can reinforce existing inequalities.
The finding strikes at the core of the framework's logic: the AI Act rests on the assumption that risky use cases can be identified in advance and regulated through documented assessments. If the actors who actually develop and deploy such systems do not include fundamental rights in those assessments, the link between the classification on paper and the protection it was meant to provide dissolves.
What compliance costs
In April 2026, the European Commission estimated €23,850–€40,400 in additional costs for certain high-risk systems already covered by existing product safety and quality management requirements (The National Interest). The figure therefore concerns the marginal cost for systems that already have a compliance regime in place — not the total cost for all types of systems, and it should be read with that limitation in mind.
At the same time, the extraterritorial reach means the costs do not stop at the EU's borders. According to the Law.com article's standfirst, the regulation applies wherever AI outputs are used in the EU, regardless of place of establishment. For American companies supplying systems into the European market, the delay may therefore be tempting to read as a delay of liability — but the timeline shifts; it does not disappear.
Open questions
Three uncertainties should be kept open. First, the primary sources are not available in this evidence base: the FRA report, the Digital Omnibus text, and the Italian decrees are known through secondary coverage, not first-hand documents. Second, it is unclear to what extent the Commission's own classification guidance exists in final form, and what it says if so. Third, the FRA finding is based on interviews in five member states; it says something about practice among those interviewed, but not necessarily about the entire market.
What can nonetheless be established is a temporal tension: while the deadlines slide toward 2027 and 2028 and national authorities such as Italy's are already in place, the FRA evidence suggests that the very core of the framework — rigorous risk assessment with fundamental rights at the center — is not yet part of how actors actually work. The delay buys time to build that practice. It does not replace it.