AI system inventories in the public sector: the FRIA slips to 2027, the Italian duties do not

A municipality, a local health authority, a regional agency almost always already hold some artificial intelligence tool — arrived inside a management suite, switched on by one office, trialled by one manager — and almost never a list saying how many there are, who decided on them, which data they run on. Since 2 August, for some of those systems, that list is no longer an internal document: it is a public registration.

The postponement you have heard about

Article 27 of the AI Act is the fundamental rights impact assessment, the FRIA: it applies to deployers that are bodies governed by public law or private bodies providing public services, and it must be done before use. Regulation (EU) 2026/1744 of 8 July 2026 — the Digital Omnibus on AI, in OJ L 2026/1744 of 24 July, in force since the 27th — moved it: Article 1, point 40(b) rewrites the third paragraph, point (c) of Article 113, so that Chapter III, Sections 1, 2 and 3 apply from “2 December 2027 as regards AI systems classified as high-risk pursuant to Article 6(2) and Annex III”. Article 27 sits in Section 3.

What did start instead

The postponement covers only Sections 1, 2 and 3 of Chapter III. Article 49 — registration — sits in Section 5; the database of Article 71 is Chapter VIII. Neither is among the exceptions in Article 113: both apply from 2 August 2026, the general date.

Article 49(3) provides that “Before putting into service or using a high-risk AI system listed in Annex III, with the exception of […] systems listed in point 2 of Annex III, deployers that are public authorities […] shall register themselves, select the system and register its use in the EU database referred to in Article 71”. Article 71(4) adds that such information “shall be accessible and publicly available in a user-friendly manner”.

Annex VIII, Section C sets what goes in it, and its point 4 asks for “A summary of the findings of the fundamental rights impact assessment conducted in accordance with Article 27”. That is the whole knot: the duty to register has applied since August, while the assessment whose summary is required starts in December 2027.

One detail changes the arithmetic for anyone working in the public sector. Article 111(2), rewritten by the Omnibus, spares high-risk systems already in use before Chapter III becomes applicable, as long as their design does not change significantly. Then it adds: “In any case, the providers and deployers of high-risk AI systems intended to be used by public authorities shall take the necessary steps to comply […] by 2 August 2030”. A private company with an old, untouched system walks away; a public body does not.

The Italian floor has no postponements

Law 132 of 23 September 2025 waits for no European date. Article 14(1) requires public administrations to ensure the «conoscibilità» of how the system works and the «tracciabilità» of its use — that it can be known, and that its use can be traced. No transitional period, no risk threshold: it holds for every system, since 10 October 2025.

A warning: the text in force is no longer the one of that day. A corrigendum of 17 October 2025 fixed Article 26, decree-law 159/2025 (converted by law 198/2025) amended Article 19(6), and decree-law 107 of 26 June 2026 added paragraph 5-bis to Article 24. Article 14 is unchanged, and Article 14 is the one that concerns you.

On 15 September 2026 the Official Gazette carried legislative decree 160 of 9 September 2026, the first act implementing the delegation: in force from 30 September, it inserts Article 437-bis into the criminal code, whose fourth paragraph punishes the «utilizzatore professionale» — the professional user — of high-risk AI systems who intentionally fails to adopt human oversight measures, where danger to life or safety follows. The second decree, on the powers of the authorities and on public-sector use of AI, does not appear to have been published.

The AgID guidelines already ask the right questions

The AgID guidelines on adopting AI in the public sector went to consultation with determination no. 17 of 17 February 2025. Eighteen months on, the AgID document card still carries the tag «(iter in corso)» — procedure under way — and the PDF opens with the heading «Bozza di linee guida», draft guidelines. They are not binding. But Annex C, «VALUTAZIONE D’IMPATTO DELL’INTELLIGENZA ARTIFICIALE», version 1.0 of 14.02.2025, is already the list of questions: six sheets — from the introductory ones to fundamental rights, data governance and accountability — plus an appendix on the risk level.

That is where the real work sits: which units will use the system, for which processes, what role the administration holds (provider, deployer, user), which people it will affect, what legal basis legitimises the decisions taken on its output. None of these questions is answered with an internet search: they are answered with council resolutions, managerial decrees, tender specifications, contracts, minutes. Paper that lives inside the organisation and cannot leave it.

What to do now

Count before you classify. System name, the office using it, the act that introduced it, the supplier, the data processed — including whatever arrived inside a management suite.
Record who decided. For each entry, the administrative act and the person: that is the traceability of Article 14, and the FRIA needs exactly the same.
Check Annex III line by line: if a system falls under it, registration under Article 49 is already due.
Keep the list beside the ICT census questionnaire: its section on AI in internal processes asks for part of the same data, with a signature underneath.

Why a closed system

Those documents describe how the organisation decides about citizens. Handing them to an outside service to be summarised means exporting, in one go, the map of public decision-making for that territory. You need a model running inside the organisation’s perimeter, on open weights you can hold still for years: the FRIA is updated whenever something changes, and an analysis produced with a model that no longer exists six months later cannot be compared with the previous one. You must be able to tell an inspector what is inside the system you used to classify your own. And it has to be cheap per document, because the archives run to thousands of pages. The map that comes out of that work stays with the organisation.

Comply and decide

On complying: the inventory and the FRIA stop being an annex filled in against a deadline and become a control running permanently over decrees, contracts and systems in service, with the record of who checked what and when, ready for registration under Article 49 or for an inspection. On deciding: the same setup holds acts, case files, management systems and contracts together in a single operating model, and on that model AI agents carry out decisions with an operator in command — not “the list is ready”, but “these three systems fall under Annex III and none of them has an impact assessment”. For large enterprises, defence, the public sector and healthcare, always in both delivery modes: on-premise on standalone machines that need no deep integration into the organisation’s network, or a dedicated CSIDIA cloud with data centres in Italy and premises we staff ourselves. Management included.

Could you say today how many AI systems are running in your organisation, and who decided on them? Half an hour with one of our experts for a first check.

What we do not know

We could not verify from a primary source whether the Article 71 database is operational and open to public deployers, nor does any Commission clarification appear on how a registration due from 2 August 2026 sits with a FRIA summary whose underlying duty starts fifteen months later. For systems under point 2 of Annex III, Article 49(5) provides for registration “at national level”: we found no Italian act saying who keeps it. And the AgID guidelines on development and procurement (determination no. 43/2026) do not appear to have been adopted in final form.