AI Act: Mandatory Training Without Level, National Sanctions

Article 4 of the AI Act requires companies to enhance their teams' mastery of AI, without setting a specific competency threshold. The fines of €15 million or 3% do not target this obligation, and the stricter rules for high-risk systems will not come into effect until 2027. In the meantime, national authorities are in charge of oversight, and the certificates sold on the market are insufficient to prove the judgment capabilities in relation to the tools.

Limited European Fines and Delayed Deadlines

The scale of €15 million or 3% of global turnover is included in the AI Act, but it does not apply to Article 4, which does not appear in any list of violations. Instead, this scale applies to the obligations of Article 26, listed in Article 99. The implementation of the chapter containing these requirements has been postponed to December 2, 2027, for most affected systems. The common framework for AI mastery is yet to be developed by 2027. During this period, it is up to member states to introduce effective and proportionate sanctions into their national laws, while, since August 2, national market surveillance authorities have been responsible for controls, instead of the European Commission. According to the Commission, a sanction is mainly likely to be applied in the event of an incident highlighting inadequate training. This results in a period of about fifteen months without measurable obligations regarding teams' judgment in relation to AI.

Article 4 Requires Companies to Strengthen Their Teams' Mastery of AI

Since July 27, 2026, the revised Article 4 by the Omnibus requires providers and deployers to take measures to support their personnel's mastery of AI and those operating the systems on their behalf. It calls for these measures to be adjusted to the skills, experience, training, usage context, and target audiences. However, it is specified that there is no requirement to guarantee a particular level of mastery for an individual. The 2024 version mentioned a sufficient level, while the 2026 version refrains from setting one and leaves it to companies to define what is proportionate, without any provided instrument for measurement. At this stage, this is the only provision in effect covering all systems, and it applies without a staff threshold, from a twelve-person accounting firm to an industrial SME of eighty people using assistants.

The Certification Market Faces Commission Reservations

Many offerings address Article 4 in the form of two-hour or one-day awareness sessions, complete with quizzes and individual certificates, often displaying numerical warnings of €15 million or 3% of global turnover. However, this scale, while real in Article 99, does not pertain to Article 4. The European Commission, for its part, indicates in its FAQ that relying on a user manual or imposing its reading on staff is likely to be ineffective. It has published a directory of about forty training initiatives, clarifying that their reproduction does not create any presumption of compliance. A certificate confirms that a session took place, not that competence is maintained.

Real Business Uses: High Adoption and Gaps in Training

According to a YouGov study conducted for Microsoft France in January among 657 executives and leaders, 80% of decision-makers use generative AI weekly. Among these users, 61% use personal accounts, and more than seven out of ten have received no training. Cited studies also indicate that competence can erode with use: faced with a confident but incorrect assistant, participants' correct responses dropped from 46% to 31%.

High Risk: Human Oversight, Automation Bias, and Stop Powers

For high-risk systems, Article 14 stipulates that the person responsible for human oversight must remain vigilant regarding their automation bias and retain the ability at all times to refrain from using the system or to disregard its results. Deployers are required to assign this oversight to individuals who possess the necessary skills, training, and authority. These obligations, along with the scale of €15 million or 3%, will apply as planned, with the concept of automation bias now enshrined in law.

Proving Mastery: Three Proposed Metrics and National Oversight

To document mastery beyond certificates, three measures are proposed: residual autonomy, by redoing ten cases each quarter without the tool and observing discrepancies; the quality of discernment, by tracking the proportion of AI proposals modified, contested, or rejected, a metric available in usage logs but underutilized; and transmission, by measuring the time it takes for a newcomer to regain pre-AI autonomy. A team that accepts 97% of suggestions as they are does not understand the limits of the tool, according to this framework. No regulator requires these measures, but it is argued that they would be difficult to contest. Legally, member states must provide for national sanctions, and their authorities have been monitoring since August 2. In France, no authority is dedicated to Article 4, with competencies shared among CNIL, DGCCRF, Arcom, and others. The Commission believes that a sanction is especially likely after an incident revealing insufficient training. In this context, treating Article 4 as a cost of €200 excluding VAT per person for a certificate reflects one interpretation, while another sees it as a matter of know-how. The European AI Committee must also develop common mastery objectives by 2027.