Italy has finished designating the national authorities for the EU AI Act. Legislative Decree No. 179 of 7 October 2026 was published in Gazzetta Ufficiale n. 234 on 8 October and takes effect on 23 October 2026. It is the second implementing decree under Law 132/2025, the national AI statute that entered into force on 10 October 2025, according to Normattiva's record of that law. Most of the decree is a faithful transposition of the AI Act. One provision on employment goes beyond it.
What the decree does
According to noze's reading of the text, the 54-article decree makes AgID the notifying authority. AgID assesses, designates and monitors conformity-assessment bodies. The ACN, the national cybersecurity agency, becomes the market-surveillance authority and the EU single point of contact. It also keeps a national register of high-risk critical-infrastructure systems. Banca d'Italia, CONSOB and IVASS supervise high-risk systems in banking, financial services and insurance. The Garante for data protection keeps its role within the limits of the AI Act.
The penalty tiers mirror the AI Act. Under Article 99 of the Regulation, prohibited practices carry fines of up to €35 million or 7% of worldwide annual turnover, whichever is higher. Other operator obligations carry up to €15 million or 3%. Misleading information to authorities carries up to €7.5 million or 1%. The same article caps fines for SMEs and start-ups at the lower of the two figures. noze reports that the Italian decree repeats these tiers in its Article 23 and adds a lower tier of up to €1 million or 0.5% for certain other obligations.
The decree also creates an Italian AI regulatory sandbox in its Article 26. Its stated goals, per noze, include evidence-based regulatory learning and faster market access, especially for SMEs and start-ups.
The strongest case for the decree
The case for this design is serious. The AI Act is a regulation with direct effect, but it leaves enforcement to member states. Without designated authorities, there is no one to receive notifications, accredit assessment bodies or fine a bad actor. A single point of contact avoids the patchwork of regulators that businesses fear. Splitting conformity assessment (AgID) from surveillance (ACN) also separates the body that promotes AI adoption from the one that polices it. That is a sound institutional choice. Linklaters' analysis of Law 132/2025 notes that the parent statute was built to be read alongside the AI Act and not to add obligations beyond it.
Where proportionality gets tested
The first concern is the fine headline. A ceiling of 7% of turnover is the EU's number, not Rome's, and Italy deserves credit for not gold-plating it. The test is how the ceiling is applied. Fines up to that level for prohibited practices make sense. Applied to a small developer who misclassifies a system as limited risk, they would chill precisely the experimentation Italy says it wants. The SME cap in the AI Act is the safeguard, and regulators should state publicly how they will apply it.
The second concern is the sandbox. Sandboxes help only if participation actually reduces legal risk. A sandbox that offers supervised testing but no assurance on enforcement will attract mainly firms that were already well resourced. Italy should publish admission criteria, timelines and what protection a participant receives, and report outcomes so other regulators can learn from them.
The third concern is the employment rule. As noze reads Article 41, final decisions on hiring, changing or ending employment cannot rest solely on automated processing. That includes disciplinary actions and performance reviews. A natural person with real and autonomous decision power must make the final call. Worker consent does not lift the prohibition. Dismissals and decisions in breach are void. Screening and candidate search are not treated as final decisions, even if they exclude someone from later stages.
This is a bolder choice than the AI Act itself, which treats employment AI as high-risk and subject to oversight duties but does not void the resulting decisions. The rationale is clear: opaque scoring can quietly end a career, and a human with authority is a meaningful check. The risk is practical. If "real, autonomous decision power" is read strictly, a manager who rubber-stamps a model's output will expose the employer to nullity claims. Voiding the decision is a heavy remedy. It creates litigation risk well beyond a fine, and the burden falls hardest on mid-sized employers without legal departments.
The screening carve-out is therefore the most important sentence in the provision. It leaves room for AI tools that rank and filter applicants, which is where most recruitment automation sits. Courts and the labour inspectorate will decide how far it stretches.
What to watch
Three things will show whether Italy gets the balance right. First, whether AgID and ACN publish guidance before the first enforcement action, not after. Second, whether the sandbox admits real products within months, not years. Third, whether Italian labour courts read Article 41 as a requirement for meaningful human judgment or as a paperwork exercise.
The decree is a competent transposition. Its success will depend on the quality of regulator guidance and on restraint in applying the heaviest remedies. An evidence-based approach would pair the nullity rule with a clear safe harbour for documented human review, and would review both after the first year of enforcement.