AI literacy: the Article 4 training obligation is already in force
While everyone looks at 2 August 2026, one AI Act obligation has been in force since 2 February 2025 and almost no company can demonstrate compliance with it: training the staff who use artificial intelligence.
Content reviewed on 13 July 2026
What Article 4 says
Providers and deployers of AI systems shall take measures to ensure, to their best extent, a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf — taking into account their technical knowledge, experience, education and training, and the context the systems are to be used in.
Note the subject: not only those who build AI, but also those who use it. A company whose customer-service staff work alongside an AI assistant is squarely in scope. So is a company whose marketing team drafts copy with a language model.
Why it is the most neglected obligation
Because it has no headline deadline: it came into force quietly, alongside the ban on prohibited practices, and made no news. And because it is worded elastically — "to their best extent", "sufficient level" — which convinced many that it was a recommendation.
It is not. It is an obligation, and one whose result must be demonstrable: the question you will be asked is not "did you do training?" but "show me what you did".
An elastic obligation is not a weak one: it is an obligation you have to fill with content yourself, and document.
What "sufficient level" means
The provision sets no hour count and no syllabus. It sets a criterion: the level must be proportionate to the person's role, to the system's risk and to the context of use. Whoever configures a chatbot's instructions needs to know things that whoever uses it to answer customers does not, and vice versa.
- People who use AI daily must know what the system can and cannot do, where it typically fails, and when not to trust the output.
- People who configure it must know which obligations bind the company, and how their choices satisfy or breach them.
- People who decide must be able to classify risk and know which uses Article 5 prohibits outright.
- Everyone must know what never goes into an AI system: third-party personal data, trade secrets, confidential customer information.
How you prove it
With three things, none of which is the training itself: a written programme (who trains whom, on what, how often), a register of the sessions delivered (date, attendees, contents, duration), and individual certificates.
That is the only part visible from the outside. Training done well but never recorded is worth, in front of an authority, exactly as much as training never done.
Training is eligible for the Chamber of Commerce voucher
Staff training costs are an eligible line in the Italian Chambers of Commerce "Voucher Doppia Transizione" schemes, where the local Chamber has joined. It is one of the rare occasions when a legal obligation coincides with a cost somebody else co-finances. Contribution rates and ceilings vary from Chamber to Chamber: the local call is the only authoritative source.
One point makes all the difference: for consultancy and training, nearly every call admits only a closed list of qualified supplier categories — explicitly including innovative start-ups. That is not a box a generic software vendor ticks.
Frequently asked questions
How many hours of training are needed?
- The provision does not say, and being wary of anyone who gives you a hard number is a good rule. The criterion is proportionality to role and risk. For an SME with a website chatbot and a few generative tools in-house, a serious programme is a handful of hours differentiated by role, repeated when the system changes.
Can I train staff in-house?
- Yes. Article 4 does not require an external body: it requires a result and the ability to demonstrate it. What matters is not who teaches, but that a programme, a register and certificates exist. If, however, you want to fund the training with a Chamber voucher, then nearly every call does require a qualified supplier to deliver it.
Does it apply if we only use ChatGPT to write emails?
- Yes, and that is the most typical case. Staff using a generative model for work are persons dealing with the use of AI systems on the company's behalf. There, the training is mostly about what not to put in the prompt and how not to trust the output blindly.
The AI literacy course
A programme differentiated by role, a register of sessions and certificates — the three things that make Article 4 demonstrable. Delivered by ShopBrain SRL, a registered innovative start-up.
See the courseSources
This page is informational and does not constitute legal advice. The authoritative texts are Regulation (EU) 2024/1689 and Italian Law no. 132 of 23 September 2025. We are engineers: we know how transparency is implemented in an AI system — we do not replace your lawyer. For hard cases, consult a qualified professional.