Italy has now finished the institutional half of its AI Act implementation. Legislative Decree No. 179 of 7 October 2026 was published in the Gazzetta Ufficiale (General Series no. 234) on 8 October and entered into force on 23 October. It is the second of two decrees under the delegation in Law No. 132 of 23 September 2025, the first dedicated national AI statute in the EU. The decree settles who regulates AI in Italy. It does not yet settle how a start-up gets from a prototype to a compliant product.
What the decree does
The decree's official title covers the powers of national authorities and the use of AI in training. Its 54 articles split the work as follows, according to the noze analysis and the IAPP summary:
- AgID is the notifying authority. It assesses, designates, notifies and monitors conformity assessment bodies.
- ACN, the National Cybersecurity Agency, is the market surveillance authority for AI systems and Italy's single point of contact with EU institutions. It also handles national registration of high-risk systems in critical infrastructure.
- Banca d'Italia, CONSOB and IVASS supervise high-risk systems in lending, financial services and insurance respectively.
- The Garante supervises within the limits of Article 74(8) of the AI Act. That provision is the data-protection carve-out for law enforcement, border management, justice and democratic processes.
The decree also sets administrative sanctions. Per noze's reading of Article 23, the ceiling is €35 million or 7% of worldwide turnover for prohibited practices. Lower tiers run to €15 million or 3%, €7.5 million or 1% for misleading information to authorities, and €1 million or 0.5% for certain other obligations. Separately, Articles 36-51 cover AI use and training in schools, universities, public administration, employment and the professions.
The case for this design
The strongest argument for the decree is that the AI Act leaves enforcement architecture to member states. A country that fails to designate authorities leaves companies facing a regulation with nobody accountable for interpreting it. Italy chose existing institutions rather than a new AI super-regulator, and it kept sector supervisors where they already know the market. A bank's credit-scoring model is judged by people who understand credit. That is a defensible, low-overhead choice, and it is better than the Garante-only or new-agency models that other states considered. The statute itself is reachable in the Gazzetta Ufficiale.
Where it falls short
The difficulty is the number of doors. A developer building an AI product for banks must now think about AgID, ACN, Banca d'Italia, CONSOB, IVASS and possibly the Garante, held together by a coordination committee and inter-authority agreements. Coordination mechanisms are only as good as their first test case. If two supervisors disagree about a classification, the company waits, and waiting is the real cost for small firms.
The unfinished pieces matter more. Article 26 creates the Italian AI regulatory sandbox under Article 57 of the AI Act, run jointly by AgID and ACN under the direct supervision of the national AI authorities. But the access, management and operating rules are left to a joint decree of the two agencies' directors-general. The Gazzetta text also leaves the registration procedures for high-risk critical-infrastructure systems to later implementing acts, according to noze. A sandbox with no operating rules offers no relief. It is a promise.
This is the part that should concern anyone who wants Italy to be a place where AI companies are built. Sandboxes are the one tool in the AI Act that can turn proportionality into practice, with supervised testing, regulator feedback and a path to the market. The decree itself states the aim as "faster access to the EU market, in particular for SMEs and start-ups," per noze. The test is whether the joint AgID-ACN decree arrives quickly and is drafted for firms with ten employees rather than ten thousand. Teacher-training criteria, by comparison, carry a 90-day deadline under Article 38. The sandbox rules carry no such date in the text we could verify.
Choices beyond the EU baseline
Italy has also gone further than the AI Act in places. Article 41 says final employment decisions, including hiring, termination and disciplinary measures, cannot rest solely on automated processing. Consent is not an exception, and decisions in breach are null and void. That is a strong human-oversight rule. It is well intentioned, but a flat ban on solely automated employment decisions can catch low-risk tools such as scheduling and shortlisting that the AI Act would treat proportionately. Italy should watch whether it pushes HR-tech vendors to serve other markets first.
Other choices help. Article 52 extends trade-secret protection to "the data, algorithms and mathematical methods used to train" AI systems, per noze, which is a pro-innovation clarification. Article 37 earmarks up to €100 million from the 2021-2027 national schools and skills programme for the education effects of social media and AI environments. Skills spending is the kind of intervention that expands AI capacity rather than restricting it.
The broader framework traces to Law 132/2025, which entered into force on 10 October 2025. As Linklaters noted at the time, it designates AgID and ACN and leaves the Garante and AGCOM with their existing powers.
What to watch
Three things will decide whether this is a good framework or merely a complete one:
- The joint AgID-ACN sandbox decree. Its deadline, eligibility rules and treatment of SMEs.
- The ACN registration procedures. Whether registration is a light notification or a heavy gate.
- The first sanctions. Whether ceilings of up to 7% of turnover are reserved for prohibited practices, as the structure suggests, or reach into ordinary compliance gaps.
Italy deserves credit for building enforcement institutions on schedule. The delegation under Law 132/2025 was due to expire on 10 October. The next job is to make the permissive parts of the framework, the sandbox and a predictable registry, as operational as the punitive ones.