ATLAS
← Back to News and Insights

AgID requires Italian public administrations to classify AI by AI Act risk

3 min

The final framework applies to every AI system and model used by bodies covered by Italy’s Digital Administration Code.

LinkedIn Post
AgID requires Italian public administrations to classify AI by AI Act risk
Italy · Photo: Ouael Ben Salah / Unsplash

Italy’s Agenzia per l’Italia Digitale (AgID) adopted the final Guidelines for the Adoption of Artificial Intelligence in Public Administration through Determination No. 169 on 9 November 2026. The announcement describes them as an operational framework for safe, transparent and AI Act-compliant use of AI in public services.

The Guidelines apply to the public bodies covered by Article 2(2) of Italy’s Digital Administration Code (CAD), and to all AI systems and models defined by Article 3 of the EU AI Act and referenced by Italy’s Law 132/2025. They were issued through the procedure in Article 71 of the CAD and are marked as definitive, version 1.0. The drafting table included AgID, the National Cybersecurity Agency, the anti-corruption authority, Consip, central government departments, social-security bodies, ISTAT, PagoPA and representatives of regional and local government.

For compliance teams, the central change is that AI adoption now begins with a formal risk classification. Under the Guidelines, public administrations must classify systems according to the risk levels in the AI Act and apply transparency, human oversight and accountability requirements, with particular attention to high-risk systems.

A mandatory classification gate

The Guidelines make their own hierarchy of requirements explicit. They state that the Italian terms “DEVE” and “DEVONO” indicate a mandatory requirement, while “NON DEVE”, “NON DEVONO”, “NON PUÒ” and “NON POSSONO” indicate an absolute prohibition. “DOVREBBE” and “NON DOVREBBE” indicate a recommendation whose implications must be considered before an alternative is chosen. “PUÒ”, “POSSONO” and “OPZIONALE” indicate a choice.

That distinction helps readers separate requirements from recommendations throughout the document. Its legal references include the AI Act, Italy’s AI Law 132/2025, the 2026 EU Omnibus AI Regulation and Legislative Decree 179/2026 on national authority powers and AI use in training.

The Guidelines cover the full adoption process. Their organisational model addresses an AI strategy, assessment of the administration’s context, priority use cases, technical standards, performance indicators, governance, risk management, impact assessment, resources, implementation, monitoring and continuous improvement. They also address responsibility along the AI value chain, lifecycle monitoring, surveillance measures and document retention.

In Atlas’s view, that structure moves public-sector AI governance away from a one-time procurement or deployment decision towards a continuing control cycle. The document calls this the Plan-Do-Check-Act model and supports it with self-assessment tools for organisational and technological maturity. ISO/IEC 42001 is used as part of that adoption model.

Data and security become adoption prerequisites

AgID treats data management as a prerequisite for trustworthy AI. The Guidelines require administrations to address data quality through accuracy, consistency, completeness and timeliness, alongside processes and governance for managing data.

The cybersecurity section is directed at AI-specific attack categories: evasion attacks, poisoning attacks, privacy attacks and abuse attacks. It provides for integrated cyber-risk management for AI systems, alongside separate treatment of assets, threats and security objectives. Personal-data protection is also addressed in a dedicated section, with the GDPR and the EU rules governing law-enforcement data among the listed references.

This creates a combined assessment rather than a narrow model-performance test. A public administration adopting an AI system must consider its AI Act risk category, the quality and governance of its data, privacy implications and the security threats associated with the system’s lifecycle.

Ethical codes and specialist roles

The Guidelines promote a human-centric model and call for an ethical code of conduct. They also emphasise digital literacy and the development of specialist professional profiles within the Uffici per la Transizione al Digitale, or Digital Transition Offices.

The official guidance says those offices should be supported by stronger internal expertise, while the document’s training section provides the organisational setting for developing relevant skills. Accountability therefore sits not only with an individual project team or supplier, but within the administration’s governance and digital-transition functions.

AgID says practical support tools will be made available on its website and periodically updated to reflect changes in law, technology and emerging good practice.

Official source: agid.gov.it
AIAI FrameworksAI Governance

Permanent link to this Atlas analysis

Continue with coverage connected by market, topic and operating context.

Continue with Atlas

Move from this development into the relevant research, comparison and workflow.