---
title: "AI literacy at work: Article 4 of the AI Act after the Omnibus and Italy's Decree 179/2026"
url: https://zerofive.ai/en/blog/compliance/ai-act-article-4-ai-literacy-training-omnibus-decree-179
canonical: https://zerofive.ai/en/blog/compliance/ai-act-article-4-ai-literacy-training-omnibus-decree-179
language: en
published: 2026-11-12
updated: 2026-11-10
author: "ZeroFive.AI"
tags: AI literacy, AI Act, Article 4, Digital Omnibus, Decree 179/2026, AI training, AI inventory
abstract: "What Article 4 of the AI Act now requires on AI training, who must be covered, how to document it and what Italy's Decree 179/2026 adds for employers."
---

# AI literacy at work: Article 4 of the AI Act after the Omnibus and Italy's Decree 179/2026

**In short.** Since 27 July 2026, Article 4 of the AI Act requires providers and deployers to take measures that support the AI literacy of the people who use AI systems on their behalf, with no minimum level to guarantee, and since 23 October Article 42 of Italy's Legislative Decree 179/2026 adds, for employers, training on the specific risks of those systems. No certificate is required: an internal record is enough, provided it links the measures to roles and to the systems in use, and that link cannot be shown without an inventory of the systems.

Between 27 July and 23 October 2026, AI training in Italian companies changed its legal basis twice, first with Regulation (EU) 2026/1744, the so-called Digital Omnibus on AI, which rewrote Article 4 of the AI Act, then with Legislative Decree no. 179 of 7 October 2026, which brought training on AI risks into the obligations of the employer. Two texts in three months, written by different legislators with different logics.

In the classrooms where we work with management and operational teams the questions are almost always the same, and they almost always come from phrases heard elsewhere: you need a certificate, one online course for everyone is enough, the obligation disappeared with the Omnibus. All of them are wrong, each in its own way, and a company that decides on the basis of any of them risks spending on training without gaining useful skills or a defensible position when an inspection comes. Here we have set out what the texts say, with sources, and the points that still lack clarification.

## The new text of Article 4

The 2024 version required providers and deployers to take measures "to ensure, to their best extent, a sufficient level of AI literacy of their staff". The text in force since 27 July 2026 states instead that providers and deployers shall "take measures to support the development of AI literacy" of their staff and of the other persons dealing with the operation and use of AI systems on their behalf, and adds a sentence that was not there before: the obligation does not require them "to guarantee any specific level of AI literacy of any individual".

Paragraph 2 shifts part of the work onto the institutions, because the Commission and the Member States must support and facilitate the efforts of providers and deployers, SMEs in particular, including through practical examples published on the single information platform provided for in Article 62.

The proposal the Commission presented at the end of 2025 went much further and would have turned the companies' obligation into mere encouragement, entrusting the promotion of skills to the Member States and the Union. Parliament and Council chose a different path. The obligation stayed with organisations, it has applied since 2 February 2025 and the Omnibus did not postpone it by a single day, but it did change its nature, because the old text looked at the result (a "sufficient" level, hard to define and even harder to prove) while the new one looks at the means, that is, the measures taken, how well they fit the systems in use and whether they can be demonstrated.

For whoever organises training the practical consequence is immediate. The company does not have to prove that every person has passed a competence threshold, and in fact the questions and answers published by the Commission make clear that Article 4 does not require measuring employees' knowledge; what it must be able to show is that it did something proportionate to the systems it uses and to the risks they carry.

## Who the obligation covers

A deployer, in the language of the AI Act, is anyone who uses an AI system under its own authority in a professional context. A company that has switched on a generative assistant in its office suite, answers customers with a chatbot or has its catalogues translated by a language model is a deployer, and the Commission gives an explicit example, because even those who use ChatGPT to write advertising or translate texts fall under Article 4, and the people involved must be informed of the specific risks, starting with hallucinations.

The perimeter of people is wider than that of employees. The text speaks of staff and of the other persons who operate the systems on the organisation's behalf, and the Commission includes contractors, service providers and, where the risk justifies it, even clients. For an agency or a consultant working with AI tools on one of your processes the same principle applies, people working for a supplier must have the skills appropriate to the task, while the choice of asking for proof of it in the contract depends on the type of system and on the risk. Procurement teams ask us about this often, and the answer changes with the case: a proportionate clause where AI touches sensitive processes, no formal requirement where the use is marginal.

Not even a degree or years of experience with AI settle the matter. For the Commission these people are "normally" considered AI literate, yet the organisation still has to check that they know the specific systems it uses, their risks and the internal rules that govern them, because the data scientist arriving from another company knows perfectly well how a language model works and has no idea which data your sales department may upload to an external service.

## Training by role and by risk

The single course, the same from the chief executive to the warehouse operator, is the cheapest solution and the one that protects least, for the reasons we set out when discussing [standard AI courses](/en/blog/compliance/mandatory-ai-course-not-enough). The Commission's FAQ describe a path that starts from a general understanding (what AI is, which systems the organisation uses, with which opportunities and which risks), moves through the company's role, which may develop systems or only use them, and through the risk level of the systems, and ends with actions calibrated on the knowledge people already have, on the context and on who is affected by the system's decisions. There is no mandatory format and the classroom is not the only possible measure, although the Commission warns that relying only on the systems' instructions for use may prove ineffective.

In the projects we follow, the groups that emerge are almost always these, and we described them in detail in the [roles and competence matrix](/en/blog/compliance/roles-competence-matrix-ai):

- executives and board, on the risks, responsibilities and deadlines of the AI Act and on the criteria for deciding where AI makes sense for the business;
- people who use generative tools every day, on approved tools, data that must not be uploaded, checking outputs and reporting problems;
- function managers, on human review of outputs, overreliance on the system and the use cases allowed in their team;
- HR and recruiting, on the use of AI in decisions about people and on the principle that the final decision rests with a human being;
- procurement, legal and compliance, on questions for suppliers, clauses, documentation to request and the link with the GDPR;
- those who oversee high-risk systems, on the system's capabilities and limits, automation bias and the moment to intervene or stop it.

![Training by role and by risk: a common foundation for everyone, modules by role and a layer dedicated to human oversight of high-risk systems](/ai-act-art4-formazione-ruoli-en.png)

For those who oversee high-risk systems the room for discretion is narrower. For these systems Article 26 requires deployers to assign human oversight to people with the necessary competence and authority, and the FAQ confirm that the obligation to train those who oversee them survives the Omnibus. The obligations for Annex III systems have slipped to 2 December 2027, true, but a company that today introduces a CV screening or credit scoring tool is already deciding who will supervise it, and with what preparation.

## Records and certificates

"You need a certificate" is the phrase we hear most often, and it is the one the Commission rejects most bluntly: no certificate is needed. Organisations can keep an internal record of training and other guidance initiatives, with no prescribed format, and they do not have to appoint dedicated figures such as an AI officer or a specific committee. A certificate can be useful for other reasons, for example for the continuing education credits of a professional body, but the AI Act does not require it.

The collection of practices the Commission has put together, with more than forty initiatives from companies and public bodies, should also be taken for what it is: a repertoire of ideas, which according to the Commission itself does not automatically grant a presumption of compliance with Article 4.

A useful record contains few things and keeps them linked: a short plan with audience, content, frequency and owner, the link between each trained group and the systems that group actually uses, attendance, dates and version of the materials, acknowledgements of the internal AI policy, revisions of the programme when a new tool arrives or an incident occurs. What is missing most often is the link with the systems, with formally impeccable records that do not say which tools they refer to, because nobody in the company keeps an up-to-date inventory of the AI systems in use.

## Supervision and penalties in Italy

Since 2 August 2026 national market surveillance authorities can check compliance with Article 4, while the European AI Office has no direct competence over this article. In Italy, Law 132/2025, in Article 20, designated AgID as notifying authority and ACN as market surveillance authority and single point of contact, leaving supervision in their respective sectors to the Bank of Italy, CONSOB and IVASS, and Decree 179/2026 completed the rules on the powers of the two agencies, with the regime we analysed in our piece on the [penalties in Decree 179/2026](/en/blog/compliance/penalties-sandbox-italy-decree-179-2026).

On penalties the picture is less linear than many posts suggest. Article 4 does not appear in the list in Article 99(4), the one for fines of up to 15 million euro or 3% of worldwide turnover, neither in the 2024 version nor after the Omnibus. Member States must still provide penalties for infringements of the regulation, and the Commission recalls that national authorities can also act on Article 4, with measures proportionate to the seriousness of the case and to any negligence.

The most concrete weight is indirect. Among the criteria used to set the penalty for any other infringement, Article 99(7)(g) considers the technical and organisational measures implemented, and the Commission notes that a lack of appropriate training may count when an incident occurs. Selling training by waving the Article 4 fine means telling an inaccurate story, ignoring it because "there is no penalty" means underestimating the day something goes wrong and the company has to show it prepared the people who were using the system.

## Decree 179/2026 and workers' training

The other change is Italian and has received far less attention. Legislative Decree no. 179 of 7 October 2026, published in the Official Gazette no. 234 of 8 October and in force since 23 October, devotes the whole of Chapter III (Articles 36 to 51) to AI training, from schools to the professions. For companies the article to read is Article 42, which brings AI into health and safety at work: systems that affect the organisation and pace of work enter the risk assessment required by Article 28 of Legislative Decree 81/2008, Italy's workplace health and safety code, and employers must ensure adequate information and training for workers on the specific risks connected with the use of AI systems and on the prevention and protection measures adopted. How this rule relates to the delegation level of the systems is the subject of our piece on [AI delegation and risk assessment](/en/blog/compliance/ai-delegation-risk-assessment-article-42).

Next to it sits [Article 41](/en/blog/compliance/ai-assisted-decisions-effective-human-oversight), which reserves to a person the decisions on the employment relationship, from hiring to dismissal, and for those using AI in recruiting or staff evaluation this makes the training of HR and managers the condition for using those tools without exposing the company. Professionals registered with professional bodies are covered by Article 47, which asks the national councils to include AI in initial and continuing training. Agenda Digitale has noted that the decree organises AI literacy for schools, public administration, the professions and healthcare, while asking private employers only for training on the risks of the systems.

For a private company the result is a dual track. Article 4 of the AI Act asks for literacy measures calibrated on roles and risks, Article 42 of the decree asks for training on health and safety risks, which in practice has to be integrated with the obligations in Articles 36 and 37 of Legislative Decree 81/2008, and therefore with the work of the health and safety officer and with the risk assessment document. The concrete danger is that the two tracks run in parallel without ever meeting, with an AI Act course bought from a supplier, a module added to the safety plan and neither of them knowing which systems are actually running in the company.

## An inventory of systems before the course

All the choices described so far, who to train, on what and to what depth, how to document it and how to update the risk assessment, need the same starting information, that is, the list of AI systems the company uses, with the people who use them and the data that flows through them. Without an inventory, training by role is reduced to training by org chart, and the record becomes an attendance list that does not show the fit between measures and risks, which after the Omnibus is exactly what Article 4 asks companies to be able to show. In another piece we explained [what an inventory should contain and how to keep it up to date](/en/blog/compliance/ai-system-inventory-iso-42001), and the diagram below shows a typical structure, highlighting the fields that training depends on.

![Typical structure of an AI system inventory, with the fields that feed the training plan, the risk assessment and the record of measures](/ai-act-art4-inventario-en.png)

This is why in our programmes training does not start from a catalogue of modules, it starts from a map of systems and roles. In the ZeroFive AI Rating model that map feeds the Readiness dimension, where skills and culture sit, and the Confidence dimension, where trust and adoption sit, and from there we build the paths for leadership, for operational teams and for those who oversee systems. Measuring skills, which Article 4 no longer requires, serves to understand where training has changed the way people use the tools and where it has produced only one more line in the record.

The Commission has not yet published guidelines dedicated to Article 4, and in its FAQ it acknowledges that guidance has so far come from webinars and questions and answers. For many companies the first concrete request, from an authority or from a client during due diligence, will arrive before that document: those with a record linked to the systems in use will be able to answer system by system, the others will have to reconstruct backwards what they did, and why.

If you want to start from the map of systems and roles in your organisation, you can book a call at [calendly.com/fabiolalli/zerofive](https://calendly.com/fabiolalli/zerofive) or write to us at [hello@zerofive.ai](mailto:hello@zerofive.ai).

*ZeroFive.AI is not a law firm: this article reconstructs the regulatory framework for operational purposes, updated to 9 November 2026, and is not a substitute for legal advice.*

## Sources

- European Commission, [AI Literacy: Questions & Answers](https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers)
- Regulation (EU) 2024/1689, [AI Act on EUR-Lex](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng)
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), OJ EU of 24 July 2026
- Italian Law no. 132 of 23 September 2025, Article 20
- [Legislative Decree no. 179 of 7 October 2026 on Normattiva](https://www.normattiva.it/eli/id/2026/10/08/26G00200/CONSOLIDATED)
- Agenda Digitale, [Digital Omnibus AI, cosa cambia davvero per le imprese](https://www.agendadigitale.eu/sicurezza/digital-omnibus-ai-cosa-cambia-davvero-per-le-imprese/)
- Agenda Digitale, [Alfabetizzazione IA, il decreto italiano lascia scoperte le imprese](https://www.agendadigitale.eu/cultura-digitale/competenze-digitali/alfabetizzazione-ia-il-decreto-italiano-lascia-scoperte-le-imprese/)
