What Law No. 132/2025 Actually Changes
Italy's Law No. 132/2025, published in the Gazzetta Ufficiale and in force since October 2025, is the first comprehensive national AI law enacted by an EU member state. It was written to complement the EU AI Act (Regulation (EU) 2024/1689) rather than substitute for it. This distinction matters operationally: the law does not introduce a separate risk classification scheme. Enterprises continue to work within the EU AI Act's existing tiers (unacceptable, high, limited, minimal risk), and Italy's law adds obligations on top of that structure rather than requiring a parallel assessment framework.
The practical effect is that compliance teams already mapping systems to EU AI Act risk categories do not need to rebuild that work. What they do need is a second pass that checks whether any system touches a sector the Italian law singles out, and whether Italy's human-oversight and criminal-liability provisions create obligations the EU-level classification alone would not surface.
Two Authorities, Divided Responsibility
Law 132/2025 assigns oversight to two separate national bodies. The Agenzia per l'Italia Digitale (AgID) handles promotion and certification activity, while the Agenzia per la Cybersicurezza Nazionale (ACN) is responsible for cybersecurity and risk control. This division means a single AI deployment can fall under the purview of both agencies depending on which aspect of the system is in question, for example certification of a model versus the security controls protecting it at runtime.
For enterprise governance teams, this bifurcation has a direct organizational consequence. Compliance ownership needs to identify which authority governs which part of a given AI use case before an inquiry or incident arises, rather than treating Italian oversight as a single point of contact. Reporting structures and internal escalation paths should reflect this split rather than assume a unified regulator model familiar from other jurisdictions.
Sector-Specific Human Oversight Requirements
The law includes explicit provisions for AI use in healthcare, judicial proceedings, employment and labor relations, and public administration. Across these sectors, the consistent requirement is human oversight and accountability for decisions affecting individuals. This reinforces, rather than replaces, the human-oversight expectations already present in the EU AI Act for high-risk systems.
The operational implication is that any AI system used in these four domains needs a demonstrable point where a human reviewed, confirmed, or overrode an AI-generated recommendation. The law does not specify an exact technical format for this record. That absence is significant: enterprises should not wait for a prescribed schema before establishing the underlying practice of logging human review checkpoints, because building that capability after implementing decrees are published will take longer than adapting an existing logging practice to a new format.
Deepfakes, Criminal Liability, and Content Provenance
Law 132/2025 amends the Italian Criminal Code to introduce or extend penalties for illicit AI use, with specific attention to AI-generated deepfakes and synthetic media used to cause harm. The law also addresses copyright implications of AI-generated content, clarifying protections for works created with AI assistance.
These provisions create a practical intersection with any enterprise workflow that generates AI content for marketing, communications, or customer-facing material. While the law does not mandate a specific technical standard such as a watermarking protocol, the criminal provisions around deepfakes suggest that being able to demonstrate the provenance of AI-generated content, where it originated and how it was reviewed, is a reasonable operational safeguard even absent a codified requirement.
What Remains Undefined, Including AI Agents
A significant portion of Law 132/2025's operational detail, including sanctions schedules, certification procedures, and technical documentation standards, is delegated to implementing decrees that had not been fully published at the time the law entered into force. This means enterprises are currently working against a partially specified regulatory target.
The law also does not contain explicit statutory language defining "AI agent" as a distinct legal category. Autonomous or agentic AI systems are governed by the same general human-oversight and high-risk provisions inherited from the EU AI Act framework, applied case by case based on the sector and function involved. In practice, this means an AI agent taking healthcare-related or employment-related actions falls under the same human-oversight expectations described above, even though the statute does not name agents specifically. Enterprises deploying agentic systems in these sectors should treat the general oversight requirement as applicable now, rather than waiting for agent-specific language that may or may not appear in future decrees.
Where Runtime Governance Fits
Because Law 132/2025 reinforces human-oversight and accountability requirements without yet specifying exact technical formats, the practical compliance posture is to build durable, evidence-producing controls rather than wait for a finalized checklist. This is where runtime governance and monitoring for AI agents becomes operationally relevant: systems that enforce policy at the point of action, log decisions and overrides, and maintain agent identity and permission boundaries produce the kind of audit trail that sector-specific human-oversight provisions are likely to require, regardless of the exact schema future decrees adopt.
Trussed AI provides runtime governance and security for enterprise AI agents, including runtime policy enforcement, audit logging, agent identity and permissions, and tool approval workflows. These capabilities are relevant to the operational gap Law 132/2025 leaves open: the law establishes that oversight and accountability are required, but leaves the technical implementation to future guidance. Enterprises that already have auditable, least-privilege agent controls in place are better positioned to adapt to that guidance than those building from a compliance blank slate.