Is Law 132/2025 Italy's implementation of the EU AI Act?

Law 132/2025 is national legislation that must be interpreted and applied consistently with Regulation (EU) 2024/1689. Article 3(5) says it creates no new obligations beyond the EU AI Act for AI systems and general-purpose AI models. The law also contains Italian sector provisions and amendments covering employment and healthcare. Other provisions cover intellectual professions and public administration, as well as copyright and criminal law. Build the control register against both instruments and avoid presenting the national law as a replacement for the EU AI Act.

Which evidence should a compliance owner request first?

Start with the artifact created by the person who owns the legal act. For employment, that may be the worker notice and delivery record. A professional engagement needs the client communication and professional approval. Healthcare needs patient information and the clinical decision record. System verification must also be retained. Public administration needs the procedure file and named responsible person. Request telemetry comes after those artifacts and should be joined to them by a stable use-case or event identifier.

Can an AI gateway satisfy the worker-information duty in Article 11?

A gateway can show activity on HTTP model routes deliberately sent through it. The record may identify the caller and approved destination. It can also preserve the classification and policy version, along with the outcome and time. Article 11 assigns an information duty to the employer or commissioning party by reference to Legislative Decree 152/1997. A runtime event cannot deliver or validate that notice. Keep notice content and delivery evidence with the employment owner, then use request records only as corroborating operational evidence.

Does a human approval button prove compliance in healthcare or professional work?

A button proves little without the surrounding record. The evidence should name the qualified person and show what information that person reviewed. It should capture the decision and time. It must also preserve the authority to reject or change the AI-assisted result. Article 7 reserves medical decisions to medical professionals. Article 13 requires intellectual work to remain predominant in the professional service. The application record must prove those facts; a gateway can only corroborate the routed model interaction.

Does Law 132/2025 require every prompt to be logged?

The law does not state a universal requirement to retain every prompt. Article 14 requires traceability of AI use in public administration, and other provisions create duties whose evidence depends on the context. Design retention around the applicable legal basis and necessity. Apply data minimization and security, as well as sector record rules. A classified metadata event may support a control test without creating a second archive of sensitive prompt content. Counsel should approve the evidence design for each use case.

Italy AI Law 132/2025: Ownership and Evidence Duties

Italy's Law 132/2025 took effect on October 10, 2025. Its sector provisions put responsibility on identifiable people and institutions. Employers must handle worker information duties, while professionals must preserve the primacy of intellectual work. Doctors retain clinical decisions. Public officials remain responsible for administrative measures. For each routed model call, a request log can preserve the caller, destination, policy and timestamp. The legal file also needs records held by the responsible employer, professional, doctor or public official.

The practical compliance artifact is an owner-to-evidence register. Each applicable provision gets one responsible owner and a legal artifact. It also gets an operating artifact and a boundary statement for any gateway record.

TL;DR

Law 132/2025 has been in force since October 10, 2025 and must be read alongside the EU AI Act.
Assign employment and professional duties to the people who already own those decisions. Do the same for medical and public-administration duties.
Keep legal artifacts such as notices and client communications separate from runtime request evidence. Clinical approvals and administrative records also belong outside that telemetry.
Gateway logs support only HTTP model traffic deliberately routed through the gateway. They leave copyright analysis and criminal provisions outside scope, along with human authority and off-route activity.

Law 132/2025 adds national sector rules

The official Normattiva text of Law 132/2025 describes principles for the research and development of AI systems and models. It also covers their adoption and use. Article 1 requires those provisions to be interpreted and applied consistently with Regulation (EU) 2024/1689. Article 3(5) also says the Italian law creates no new obligations beyond the EU AI Act for AI systems and general-purpose AI models.

Law 132/2025 still contains national rules directed at employers and intellectual professionals. Other rules apply to healthcare participants and public administrations. Separate provisions change copyright and criminal law. The Gazzetta Ufficiale record confirms publication in General Series No. 223 on September 25, 2025 and entry into force on October 10, 2025.

AgID and the National Cybersecurity Agency are designated as national AI authorities under Article 20, subject to the roles reserved for sector regulators and independent authorities. An enterprise register should therefore cite the exact sector provision. It should also name the authority relevant to the matter instead of treating Law 132/2025 as one generic AI control.

Build the owner-to-evidence register

Use one record per AI use case. Put the business purpose and accountable owner in the AI governance framework, then add the applicable Law 132/2025 provision. The record should identify who makes the legally consequential decision. It should also identify which artifact proves that authority was exercised.

For each provision, retain two evidence classes. The legal class includes the worker notice and client communication. It can also include the clinical decision or administrative measure, with the rights analysis or incident file where applicable. The operating class includes the approved model route and policy version. Record the user or agent identity, along with the request classification. Preserve the decision outcome and timestamp where those facts exist.

Join the two classes with a stable use-case and event identifier. Do not collapse them into one vendor dashboard. A gateway event can show that a classified prompt from a named role reached an approved endpoint. It can also show that policy version 17 governed the request. That event says nothing about the legal sufficiency of an employment notice or the doctor responsible for the final therapeutic choice.

The register should also name excluded paths. Browser sessions and local models need evidence from their own control points. The same applies to direct provider calls and applications that hide their model traffic.

Employment duties belong to the employer or commissioning party

Article 11 says workplace AI must be safe and reliable. It must also be transparent and consistent with human dignity and personal-data confidentiality. It places the information duty on the employer or commissioning party in the cases and manner provided by Article 1-bis of Legislative Decree 152/1997. Workers' inviolable rights and the listed protections against discrimination apply to the organization and management of the employment relationship.

The responsible owner should be the employer function accountable for the deployment, with employment counsel or the designated labor-compliance owner approving the notice analysis. Keep the issued notice and delivery evidence. Record the affected worker population and system purpose. Preserve the decision role and change history. Keep the discrimination assessment and escalation route in the same control file.

For a deliberately routed model request, an enforcement record can corroborate the caller identity and destination. It can preserve the content classification and policy decision, as well as the time. It can also show that a blocked route never reached the model. The HTTP-layer policy enforcement guide explains that narrow mechanism.

Employers still need separate evidence for Article 11 notice completeness, consultation duties elsewhere in employment law and any discriminatory effects.

Professional duties stay with the professional

Article 13 limits AI use in intellectual professions to activities that are instrumental and supportive of the professional service, with the intellectual work at the center of the engagement. The professional must communicate information about the AI systems used to the client in language that is clear and simple while remaining exhaustive, so the relationship of trust is preserved.

The professional or practice owner is responsible. The evidence file should contain the client communication with its version and delivery date. It should identify the system and task described. Keep the professional review and final work product. A matter-level record should identify the person who accepted responsibility before advice or other professional output reached the client.

Match the model call to the disclosed system and completed matter with the approved destination, active policy and initiating identity from the gateway record. The engagement file must establish that the professional's intellectual work remained predominant and the client received an exhaustive communication. It must also show that the final advice met the profession's standard of care.

A useful acceptance test takes one completed matter and asks the professional to reconstruct both records. The request evidence should match the disclosed system. The engagement file should independently show the review and final professional judgment.

Healthcare records must preserve medical authority

Article 7 gives the individual a right to be informed about the use of AI technologies in healthcare. It states that AI systems support prevention and diagnosis. They can also support treatment and therapeutic choice, while the decision always remains with medical professionals. Systems and the data used in healthcare must be reliable and periodically verified. They must also be updated to reduce error risk and improve patient safety.

The clinical service owner and responsible medical professional need separate artifacts. The service file should hold the patient-information mechanism and validation scope. Keep the periodic verification result and approved version. Record the update decision and safety escalation process. The patient record should identify the medical professional responsible for the actual clinical decision and preserve the clinical rationale under the applicable record rules.

For an investigation into clinical content sent to a prohibited destination, the routed request record supplies the authenticated role, endpoint, classification, policy result and timestamp. Model validation and data reliability require evidence from the clinical systems. Patient information and the doctor's exercise of authority require evidence from those systems and the responsible medical professional.

Research processing under Articles 8 and 9 has its own statutory conditions. Treat those provisions as a separate legal workstream rather than infer research authorization from a production gateway log.

Public administration needs procedural evidence

Article 14 directs public administrations to use AI to improve efficiency and service quality while ensuring that interested parties can know how it functions and that its use is traceable. AI remains instrumental and supportive. The person with decision-making authority stays solely responsible for the measures and procedures in which AI was used. Public administrations must also adopt technical and organizational measures for responsible use, backed by training.

The procedure owner should retain the administrative file and named decision-maker. Keep the notice or explanation supplied to the interested party and the AI-use record. Preserve the review performed and final measure. The administration's training owner should keep role-specific completion evidence. Technical owners need route configuration and access control. They should also retain release approval and test results.

A separately written request record can strengthen traceability for model traffic under the administration's control. The AI audit-log chain-of-custody guide covers the custody questions behind that record. Even a signed event leaves the administrative decision outside the gateway. Legal reasons for the measure and procedural fairness remain in the administration's records. So do official training and the named person's responsibility.

Article 15 uses a similar human-authority rule in judicial activity by reserving interpretation and evaluation to the magistrate. Decisions are also reserved to the magistrate. Judicial procedure requires its own evidence design.

Copyright and criminal provisions require different controls

Article 25 amends Italy's copyright law. It ties protected works to human intellectual creation, including works made with AI assistance when they remain the result of the author's intellectual work. Text and data extraction through AI systems remains subject to the conditions in Articles 70-ter and 70-quater of the copyright statute.

The rights owner should keep source and license records, along with reservations of rights. Preserve author contributions and prompts or drafts where appropriate. Keep the legal analysis for text and data mining. A request log can identify the model interaction associated with the work, while legal and creative records establish authorship, lawful access, rights reservations and the output's eligibility for protection.

Article 26 creates a new offense concerning harmful distribution without consent of falsified or altered images and videos, as well as voices, and changes other criminal and penalty provisions involving AI use. Criminal exposure belongs with legal and trust and safety functions, with fraud and incident-response functions involved according to the event. Preserve the underlying content and consent record. Keep the distribution facts and investigation custody, along with the referral decision.

Gateway logs cover only content that crossed the managed model route. Publication channels and local generation require evidence elsewhere. The same applies to stolen accounts, downstream editing, and user intent.

DeepInspect

DeepInspect can enforce and evidence selected controls on HTTP AI traffic deliberately routed between authenticated users or agents and LLM endpoints. It evaluates application-supplied identity and role before forwarding a request. It also evaluates prompt classification and destination, along with model authorization. The resulting per-decision record can show what the policy did for a specific routed event under a specific version.

Use that record for route approval, policy testing, incident reconstruction and traceability for managed model calls. Legal owners still retain worker notices and client communications. They also retain clinical authority and administrative procedure, along with copyright analysis and criminal-response evidence. Local models and personal browser sessions sit outside the gateway's view. The same applies to direct provider calls, hidden vendor integrations, and downstream actions.

Book a technical deep dive at deepinspect.ai.