Italy's Council of Ministers gave definitive approval to the two decrees implementing Law 132/2025, settling who supervises AI and making a dismissal based only on automated processing null

What the Council of Ministers actually did

On 4 August 2026 the Council of Ministers approved, in esame definitivo, two legislative decrees adapting Italian law to the EU AI Act, exercising the delegation at art. 24 of Law 132/2025. Esame definitivo is the final cabinet stage. It follows the mandatory opinions, which in this case came from the parliamentary committees, the Conferenza unificata and the Garante per la protezione dei dati personali.

This is the same pair of decrees that received preliminary approval on 10 June 2026, not a new instrument. The government's own communique sets out what changed between the two stages, which is the useful part for anyone who read the earlier text.

One caution frames everything below. A cabinet communique is the government's summary of a decree, not the decree. The operative wording is what appears in the Gazzetta Ufficiale, and the two can differ. Treat the detail here as the government's account until the published text is available.

The governance map, which is the most useful thing in the update

The second decree settles who does what, a question Italian practitioners have been asking since the AI Act's supervision deadlines arrived. On the government's account, the Agenzia per l'Italia digitale is the notifying authority and the Agenzia per la cybersicurezza nazionale is the market surveillance authority and single contact point for the EU institutions.

Sector regulators keep their own lanes. Banca d'Italia, CONSOB and IVASS supervise high-risk AI used by financial intermediaries and directly connected to the supply of financial services. The Garante covers high-risk systems used in law enforcement, border management, justice and democratic processes.

For a compliance officer this answers the practical question of who to call, and it means the answer is not the same for every system you run. A bank deploying a high-risk credit tool and a police force deploying a biometric one are answerable to different authorities under the same decree.

Penalties, and why the commencement date is the harder question

The communique says the sanctions framework is graduated and proportionate, and that Italy used the option in the AI Act to set maximum limits lower than the European ones, calibrated to the degree of responsibility along the AI supply chain. It also describes non-pecuniary measures for breaches of low offensiveness.

The timing is conditioned rather than fixed. The decree was updated to reflect Regulation (EU) 2026/1744, with a referral that makes the entry into force of the penalties depend on the actual entry into force of the corresponding obligations and prohibitions. In plain terms, a penalty provision can sit dormant until the duty it punishes is itself live.

That is why we give no single national date for when sanctions bite. Any date would have to be assembled provision by provision from the AI Act phase-in, and we have not done that against a published Italian text.

The employment rule, which is the sharpest edge for ordinary businesses

The decree provides that decisions on the formation, modification or termination of an employment relationship, including disciplinary measures and dismissals, cannot be taken solely on the basis of automated processing. A dismissal issued in breach of that prohibition is null.

One clarification was added at the definitive stage and it matters. Candidate sourcing and selection are not final decisions on the formation of an employment relationship, even where they result in a candidate not advancing to later stages of the selection. Screening automation is therefore treated differently from the decision to hire, discipline or dismiss.

Nullity is a strong remedy. It is not a fine that a well-advised employer can price in, and it is the provision most likely to reach companies with no other AI exposure in Italy.

Civil liability, rebalanced toward the claimant

The first decree strengthens the position of a person harmed by an AI system without, on the government's account, imposing new substantive obligations on businesses. The tools are procedural: access to the system's technical documentation, a presumption of the causal link, an alternative forum near the residence of an injured natural person, and a direct action against the insurer.

The presumption of causation is the significant one. In litigation about an opaque system, the practical fight is usually whether the claimant can show the system caused the harm, and a presumption moves that burden without eliminating it.

Read alongside the direct action against the insurer, this is a meaningful change to how AI harm claims will be run in Italy, and it deserves attention from anyone pricing product or professional liability exposure.

Policing and the new criminal offence

The policing decree is narrower in audience but worth knowing. Real-time remote biometric identification is permitted only in exceptional cases, for limited periods and with prior authorisation from the judicial authority, and the creation of biometric databases through mass, untargeted collection of information from the web is prohibited.

The decree also introduces art. 437-bis into the criminal code, punishing the failure to adopt security measures in high-risk AI systems and their unlawful alteration, with the penalty graduated according to the legal interest exposed to danger. Liability can extend to the entity under the corporate liability regime.

For most commercial readers this is context rather than compliance. For anyone supplying biometric or surveillance technology to Italian public bodies, or serving as a professional deployer of a high-risk system, it is not.

The duties that already bind, with their statutory anchors

Two obligations under Law 132/2025 apply now and do not depend on the decrees. Both are more precisely drawn than general summaries suggest, so the anchors matter.

Art. 11 provides that the employer or principal must inform the worker of the use of artificial intelligence in the cases and by the methods set out in art. 1-bis of legislative decree 26 May 1997, no. 152. The duty is therefore channelled through the existing transparency regime for automated decision and monitoring systems rather than being a freestanding notice for any AI use whatsoever. If you use AI in hiring, scheduling, monitoring or evaluation, that is squarely the territory art. 1-bis addresses.

Art. 8 governs research and scientific experimentation in building AI systems in healthcare. It provides that the processing described at its first two paragraphs must be communicated to the Garante, and that it may begin once thirty days have elapsed from that communication, provided it has not been the subject of a blocking order issued by the Garante. That is a genuine waiting period, and it belongs in research and product timelines rather than being discovered late.

Art. 11 also establishes an Osservatorio on the adoption of AI in the world of work at the Ministry of Labour.

What we did not verify

We did not read the decrees. We read the Council of Ministers communique of 4 August 2026 and we say so throughout. Every description of what the decrees provide is the government's account of its own instrument, and the enacted wording may differ.

We did not locate a Gazzetta Ufficiale publication for either decree. We searched the Gazzetta Ufficiale, searched it again restricted to its own domain, and checked Normattiva's recent highlights, and none surfaced a citation. That is a failed search, not evidence that the decrees are unpublished, and we do not assert either way. We therefore state no entry-into-force date, no vacatio legis and no sanction commencement date.

We did not map the Regulation (EU) 2026/1744 phase-in provision by provision, so we cannot say which Italian penalty provisions are dormant and which are not.

We did read Law 132/2025 itself. Arts. 8 and 11 were verified word for word against the text published in Gazzetta Ufficiale n. 223 of 25 September 2025.

On the earlier version of this article: our 8 July text said the delegation had to be exercised by around October 2026 and advised readers to map their duties before then. That was accurate when written and is now superseded by the 4 August approval. We have not removed it quietly, and it is recorded here.

Related briefings

• Where the EU omnibus moved the AI Act deadlines
• The Garante's opinions on Italy's AI Act decrees
• High-risk delay and penalties under the EU omnibus
• The Garante's deepfake warning to RTI over Striscia
• AI Regulation News hub