On 9 November 2026 the Agenzia per l’Italia Digitale (AgID), Italy’s digital agency, adopted by Determination No 169 the “Linee Guida per l’adozione di IA nella Pubblica Amministrazione” (Guidelines for the adoption of AI in the public administration), provided for by the Three-Year Plan for IT in the Public Administration 2024-2026 (AgID, Amministrazione trasparente). The document is addressed to every body covered by the Digital Administration Code, from ministries to local authorities, from health authorities to public service operators, and sets out what each of them must do to introduce an artificial intelligence system: define a strategy, classify its systems according to the risk levels of Regulation (EU) 2024/1689, the AI Act, assign responsibility to the Digital Transition Officer and organise the governance of AI around the Plan-Do-Check-Act cycle and ISO/IEC 42001 (AgID). The Guidelines give the standard the role this blog has recognised from the outset: a methodological reference for the organisation, not a source of presumption of conformity with the AI Act. The document is extensive and largely aligned with the European framework; in some passages, on harmonised standards and on transparency obligations, it departs from it. We examine its addressees, its organisational model and its critical aspects.
The addressees: public administrations, public service operators and publicly controlled companies
The Guidelines are adopted under Article 71 of Legislative Decree No 82 of 7 March 2005, the Codice dell’amministrazione digitale (Digital Administration Code, CAD), and take effect after their publication in the dedicated area of AgID’s institutional website, notice of which is given in the Gazzetta Ufficiale (Normattiva). Publication on the website took place on 9 November 2026; the determination provides for the notice to be sent to the Gazzetta Ufficiale (AgID, Determination No 169/2026).
The document is addressed to the bodies referred to in Article 2(2) of the CAD, which it designates by the abbreviation PA. The audience is wider than the abbreviation suggests. Point (a) covers the public administrations listed in Article 1(2) of Legislative Decree No 165 of 30 March 2001, which include, among others, State administrations, schools of every level, the Regions, the Provinces, the municipalities, universities, chambers of commerce, non-economic public bodies and the agencies and bodies of the National Health Service, together with port system authorities and independent administrative authorities; point (b) extends the Code to public service operators, listed companies included, in relation to services of public interest; point (c) applies it to publicly controlled companies as defined by Legislative Decree No 175 of 19 August 2016, excluding listed companies that do not operate public services (Normattiva). A public service operator, a health authority or a school that introduces an AI system is therefore an addressee of the Guidelines in the same way as a ministry. The material scope covers AI systems and models as defined in Article 3 of the AI Act and referred to in Article 2 of Law No 132 of 23 September 2025.
The vocabulary is that of technical documents: the terms deve (must), dovrebbe (should) and può (may), which the document writes in capitals, are used in accordance with the ISO/IEC Directives, and the text states that deve and devono indicate a mandatory requirement. A legal reading of the document therefore requires distinguishing, paragraph by paragraph, what a public body must do from what it may do. The legislative references already include Regulation (EU) 2026/1744, the Digital Omnibus on AI, and Legislative Decree No 179 of 7 October 2026 (AgID, Guidelines).
The organisational adoption model: the Plan-Do-Check-Act cycle and ISO/IEC 42001:2023
Chapter 4 opens with two obligations: public bodies must adopt a structured set of processes, policies, resources and tools to govern, implement, monitor and improve the use of AI systems throughout their life cycle, and must implement an organisational model for the adoption of AI capable of responding promptly to changes in the regulatory and technological context. The proposed model is based on the Plan-Do-Check-Act (PDCA) continuous improvement cycle and on certain management practices defined by ISO/IEC 42001:2023, a standard which, in the words of the Guidelines, “constitutes a methodological reference for setting up an AI management system” (our translation). Public bodies may adapt the model and, for limited-risk or minimal-risk systems, merge its phases.
The path starts from an AI strategy, which each body must define consistently with its Piano integrato di attività e organizzazione (Integrated Plan of Activities and Organisation, PIAO), and continues with the analysis of context, the identification of use cases, performance indicators, risk management, impact assessment, the operational plan, monitoring and continuous improvement. For risk management the text allows reference to UNI ISO 31000 and ISO/IEC 23894; for impact assessment it recalls the fundamental rights impact assessment provided for in Article 27 of the AI Act and adds that, for minimal-risk systems as well, public bodies should define a process to assess, even in simplified form, the potential consequences for fundamental rights. Principle P.5 makes clear where responsibility remains: ultimate responsibility for the decisions adopted, whatever the level of automation, remains with the public body.
Technical standards: harmonised standards and the presumption of conformity under Article 40 of the AI Act
Chapter 4 distinguishes between two levels. In the cases provided for by the AI Act, public bodies must adopt the quality management system required by the Regulation and take account of the relevant harmonised standards, where available; non-harmonised standards, “including ISO/IEC 42001:2023, may help to demonstrate the adoption of appropriate organisational, management and control measures, but do not confer the presumption of conformity with the requirements of the AI Act” (our translation). Paragraph 4.4 specifies that this is the presumption under Article 40, which is recognised, within the limits of the requirements and obligations covered, for harmonised standards or parts thereof whose references are published in the Official Journal of the European Union (AgID, Guidelines).
The distinction coincides with the one this blog has followed in recent months. For a public body, ISO/IEC 42001 is therefore a model for organising its AI management system and documenting the adoption of appropriate measures. The legal presumption of conformity under Article 40, by contrast, flows only from harmonised standards cited in the Official Journal, and only for the requirements they cover. To date the only standard cited is EN 18286:2026, on the provider’s quality management system, whose reference was published by Commission Implementing Decision (EU) 2026/2274 of 12 October 2026, which entered into force on 14 October (EUR-Lex), as we set out in our analysis of EN 18286 and roles within the organisation. On a further, separate plane, which the Guidelines do not address, lies the certification of the management system against UNI CEI EN ISO/IEC 42001:2026, which grounds a presumption of conformity based on certification issued by an accredited certification body, as we explained in connection with the Italian adoption of the standard.
Governance: the Digital Transition Officer, the AI Officer and professional profiles
Responsibility for managing AI must be entrusted to the Responsabile per la Transizione Digitale (Digital Transition Officer, RTD) and the Ufficio per la Transizione Digitale (Digital Transition Unit, UTD); the RTD is given the responsibility and the authority to ensure that systems comply with the applicable legislation and with the Guidelines and to report to the governing bodies. The RTD works in close coordination, where appropriate, with the data protection officer and, in implementing the strategy, also involves the cybersecurity contact designated under Article 8(2) of Law No 90 of 28 June 2024. Only in contexts of intensive use of artificial intelligence may a public body establish a Responsabile dell’Intelligenza Artificiale (AI Officer), possibly with a dedicated unit, operating under the directions of the RTD. Public bodies using high-risk systems entrust human oversight to the same RTD and unit and must ensure that the staff in charge have the necessary competence, experience and authority.
Chapter 8 translates the model into competences. Public bodies must ensure that public managers, as well as Digital Transition Officers and their units, acquire and maintain a level of knowledge and competence in AI appropriate to their respective responsibilities. For specialist profiles the text states that UNI 11621-8:2026 “may constitute a useful reference for identifying and describing the professional profiles operating in the field of artificial intelligence” (our translation) and lists, by way of example, roles such as the AI Product Manager, the AI Data Engineer and the AI Machine Learning Engineer, alongside the IT lawyer and the AI Ethicist. The author took part, together with the other experts of the working group, in developing the standard, whose twelve profiles are described in a dedicated article. The relationship between the standard and the European framework is examined in our analysis of CEN Workshop Agreement (CWA) 18398. The chapter does not, however, refer to Article 40(2) of Legislative Decree No 179 of 7 October 2026, under which the Minister for Public Administration, after hearing AgID and ACN and with the agreement of the Conferenza Stato-Regioni, defines in ministerial guidelines the minimum framework of AI competences required of public employees (Gazzetta Ufficiale).
Critical aspects: harmonised standards, transparency, the standardisation request and references to texts not yet adopted
The first aspect concerns harmonised standards, on which the document is not consistent. Paragraph 4.4 links the presumption under Article 40 to the publication of the reference in the Official Journal; Chapter 11, on cybersecurity, although it refers back to paragraph 4.4 itself, provides instead that public bodies must take into account the standards produced by CEN/CENELEC JTC 21 because they “represent harmonised technical standards capable of guaranteeing the presumption of conformity with the cybersecurity requirements laid down in Article 15 of the AI Act” (our translation) (AgID, Guidelines). In the Official Journal of the European Union the only act citing harmonised standards for high-risk AI systems is Implementing Decision (EU) 2026/2274, which concerns EN 18286:2026 on the quality management system alone (EUR-Lex): no standard on the cybersecurity requirements of Article 15 confers that presumption today. EN 18286:2026, for its part, appears nowhere in the document, although its reference had been published twenty-six days before the Guidelines were adopted: a public body that takes on the role of provider of a high-risk system, and for that reason must adopt the quality management system, does not find in the document any indication of the standard that today confers the legal presumption of conformity for the requirements it covers.
The second concerns transparency. Paragraph 7.1 states that “Article 50 of the AI Act provides that high-risk systems and certain AI systems are subject to stringent transparency requirements” (our translation) and brings under it the technical documentation of high-risk systems and the instructions for their use as well (AgID, Guidelines). Article 50, however, governs the transparency obligations of providers and deployers “of certain AI systems” and specifies, in paragraph 6, that paragraphs 1 to 4 “shall not affect the requirements and obligations set out in Chapter III” (AI Act Service Desk): the technical documentation of high-risk systems and transparency towards their deployers are the subject of Articles 11 and 13. A reference to those provisions would help public bodies to identify the legal basis of those obligations.
The third concerns the standardisation request. Footnote 19 indicates Decision C(2023) 3215 as the source of the mandate to CEN and CENELEC; Implementing Decision (EU) 2026/2274 instead indicates as the basis of the work Implementing Decision C(2025) 3871 of 23 June 2025, which repealed the earlier one (EUR-Lex). A reference to the decision in force would link the document to the list of standards to be expected, which depends on it.
The fourth concerns the organisational set-up. Strategy, data governance, regulatory compliance and human oversight of high-risk systems all converge on the Digital Transition Officer; the dedicated AI role is optional, operates under the RTD’s directions, and the text gives no criteria for identifying the contexts of intensive use in which it may be established. Paragraph 4.10 requires public bodies to determine and allocate the financial, technological and human resources needed to develop, maintain and continuously improve their AI systems, without specifying which of them should go to the RTD’s unit: in our view the actual functioning of the model depends on that choice, which is left to each public body.
The fifth concerns references to texts not yet adopted. The document refers to specific paragraphs of the guidelines on development (paragraph 3.2 on the levels of autonomy of systems, paragraph 3.5 on classification as a basic operator, Chapter 3 on the reference architecture) and requires the operational plan to include procurement and development activities governed by the dedicated guidelines. Specific obligations also refer to those texts: the information notice to data subjects must state, as regards the type of AI used, the architectural elements described in the guidelines on development, and for systems delivered through cloud services public bodies must take account of the operational indications contained in the guidelines on procurement. Those two documents have not yet been adopted: AgID put them out to public consultation from 12 March to 11 April 2026 (AgID) and states that their adoption procedure is under way (AgID).
As to timing, the Guidelines set no deadline for public bodies to comply; nor does Article 14 of Law No 132/2025, which requires them to adopt technical, organisational and training measures (Gazzetta Ufficiale). The AI Act deadlines still apply: the Digital Omnibus has deferred the application of Sections 1 to 3 of Chapter III, on high-risk systems, to 2 December 2027 for Annex III systems and to 2 August 2028 for Annex I systems (EUR-Lex). For systems used by the police forces, Article 21 of Legislative Decree No 160 of 9 September 2026 further provides that they must be made compatible with Chapter II of Title I of that decree within one year of its entry into force (Gazzetta Ufficiale).
Conclusions
Since 9 November 2026 public bodies have had a single framework for the adoption of AI, built with the vocabulary of obligations and options and anchored to the AI Act and to Law No 132/2025. It falls to each body to define its strategy consistently with the PIAO, classifying its systems, assigning responsibilities to the Digital Transition Officer and documenting competences. The guidelines on development and procurement, to which the text already refers, have yet to be adopted, and some passages have yet to be aligned with the framework in force: the presumption attributed to the JTC 21 cybersecurity standards, the absence of any reference to EN 18286:2026, the citation of the repealed standardisation request and the reference to Article 50 for obligations that the AI Act places in Articles 11 and 13. We consider it useful for public bodies to begin mapping the systems in use and the roles they hold in relation to each, keeping distinct in their own documents the management system set up on ISO/IEC 42001 and the quality management system that the AI Act requires of the provider. Implementation now rests with the organisational measures of each public body.
Author: Valentina Grazia Sapuppo










