Back to Blog

AgID's AI Guidelines for Public Administration: where the AI Act meets ISO/IEC 42001 (and where it doesn't)

·5 min read
Isometric Tomato Blue illustration of a public administration adopting artificial intelligence: an institutional building, a Plan-Do-Check-Act cycle at the centre and a certified document, connected to one another, with security and data icons.

On 9 November 2026, AgID (the Italian Agency for Digital Italy) published, through Director General's Determination No. 169, the Guidelines for the Adoption of Artificial Intelligence in Public Administration (v1.0, 141 pages): not a generic handbook, but an operating model that requires public bodies to classify the risk of their AI systems under the AI Act and to manage adoption through a Plan-Do-Check-Act cycle built on the ISO/IEC 42001 standard. For public administrations this means a maturity self-assessment and AI-specific cybersecurity measures; for private vendors, a clear signal of where to direct compliance investment — with a couple of caveats worth stating upfront.

What the Guidelines are and where they come from

Their legal basis is Article 71 of Italy's Digital Administration Code (CAD), which grants AgID the power to issue binding guidelines for public administrations: the text states explicitly that the Guidelines are "issued following the procedure set out in Article 71 of the CAD". The document sits within the 2024-2026 Three-Year Plan for IT in Public Administration (Prime Ministerial Decree of 12 January 2024), of which it is a direct implementation.

The adoption process is longer than a simple announcement: it started with AgID Determination No. 225/2024, went through public consultation, an opinion from the State-Regions Conference (Conferenza Unificata), consultation with the Italian Data Protection Authority, and an EU notification procedure. The scope covers all the bodies listed in Art. 2(2) of the CAD — from central government to local authorities and non-economic public bodies — and was drafted within a working group that includes, among others, the National Cybersecurity Agency (ACN), ANAC, Consip and the Department for Digital Transformation.

The core of the model: PDCA + ISO/IEC 42001, with a caveat that matters

Chapter 4, "Organisational model for AI adoption", is the backbone of the document: it proposes a model based on the Plan-Do-Check-Act continuous-improvement cycle, built on "some of the management practices defined by the ISO/IEC 42001:2023 standard", which the Guidelines describe as a "methodological reference" for the AI management system.

One clarification matters a great deal here: the Guidelines themselves state that ISO/IEC 42001:2023 is a non-harmonised standard. It "can help demonstrate the adoption of adequate organisational measures", but it does not grant a presumption of conformity with the AI Act's requirements. For vendors selling AI systems to public administrations, pursuing ISO 42001 certification remains a sound organisational investment and a signal of maturity — not a legal shortcut.

Alongside this, according to AgID's official announcement, the model is "accompanied by practical tools for self-assessing organisational and technological maturity". Worth being precise here: this exact wording does not appear in the full text of the Guidelines; par. 2.2 refers more generally to "practical support tools" published separately on the agency's website. The maturity self-assessment is real and consistent with the PDCA framework, but it is described in AgID's press release more than in the body of the document itself.

AI Act risk classification, applied to public administration

The Guidelines weave the PDCA/ISO 42001 management model together with the risk classification set out by the AI Act, also referenced by Art. 2 of Law 132/2025 on AI systems and models. The cornerstone principle, set out in par. 3.4 as Principle "P.5 Responsibility", is unambiguous: "public administrations adopt AI as a tool to support human activity, aware that ultimate responsibility for decisions made — whatever the level of automation involved — by AI systems remains with the public administration." This is a more nuanced formulation than the generic "the final decision always rests with the administration" that has circulated in some press summaries: it does not ban automation, but it firmly anchors responsibility to the public body. The principle echoes the "non-exclusivity of automated decision-making" referenced in Chapter 10 on personal data protection.

AI-specific cybersecurity measures

Chapter 11, "Cybersecurity" (pp. 124-139), is the most technically dense part of the document and goes beyond the "traditional" IT security already required of public administrations under ACN's cyber perimeter framework. The Guidelines identify an attack taxonomy specific to AI systems, organised into four families: evasion attacks, poisoning attacks (availability, targeted, backdoor and model poisoning), privacy attacks (data reconstruction, membership inference, model extraction, property inference) and abuse attacks, the latter specific to generative AI. The picture is completed by a risk-management model based on the NIST AI RMF, the ENISA threat taxonomy, and nine security objectives drawn from the NCSC UK Guidelines. Data quality and security are treated as a distinct requirement, with a dedicated chapter (Ch. 9) and a specific chapter on personal data protection (Ch. 10), which also references EDPB Opinion No. 28/2024 on AI models.

What changes for private vendors

This is the point where caution is warranted: the adoption Guidelines do not directly address the effects on tender specifications and public contracts. The text repeatedly refers — in par. 3.4 (P.20), 4.1, 4.3 and 9.3 — to a separate document, the "Guidelines for AI Procurement in Public Administration", which does not appear to be attached to Determination No. 169/2026: the published package contains only the Determination and the adoption Guidelines.

The impact on tender specifications is therefore, at this stage, a plausible interpretation — consistent with a document that cites the Public Contracts Code among its regulatory references — rather than a fact confirmed by the primary text. According to secondary sources such as DIAC and ICT Security Magazine, risk-classification and ISO 42001-compliant management requirements are expected to flow into contractual clauses for public AI procurement. For system integrators and software vendors working with public administrations, moving now on audits, documentation and an ISO/IEC 42001-aligned management system remains a prudent choice regardless of when — or whether — the procurement guidelines are eventually published. One piece of context reinforces this picture: Law 132/2025 designated AgID as Italy's National Authority for Artificial Intelligence, consolidating its role as the reference point for anyone dealing with public administration on these matters.

Operational take-away

For public administrations: map AI systems in use or under adoption, classify their risk under the AI Act, launch the organisational and technological maturity self-assessment (tools published separately on AgID's website), and document the management system along the Chapter 4 PDCA cycle. For vendors: ISO/IEC 42001 does not grant a presumption of AI Act conformity, but it remains the strongest organisational foundation to build an offering for public administration — especially ahead of contractual requirements that, while not yet codified in a dedicated public text, are clearly the direction this document points to.

Taking an AI system into a public administration (or selling one)?

AI Act risk classification, PDCA cycle and an ISO/IEC 42001 management system: we can help you map your systems, set up the maturity self-assessment and document the measures before they become a tender requirement.

Let's talk about your AI adoption →