AI literacy training: compliance is not adoption

AI literacy training became a legal obligation in Europe on 2 February 2025, under article 4 of Regulation EU 2024/1689, the AI Act. National authorities begin supervision in August 2026, and businesses that have not adapted face penalties. In Italy, Law 132 of 23 September 2025 added transparency duties toward workers and required consultation with employee representatives before introducing systems that affect work organisation.

What article 4 actually requires

Article 4 sets no minimum number of hours and no specific format. It requires providers and deployers of AI systems to ensure an adequate level of AI literacy for all staff involved in operating, supervising or managing those tools. The word adequate carries weight: it means proportionate to the role, the risk level of the system and the operational context. A generic course on what artificial intelligence is, delivered to the whole workforce regardless of duties, formally satisfies a requirement without matching the logic of the rule. Supervisory authorities, active from August 2026, will assess the coherence between the AI systems in use, the roles that manage them and the documented training paths. Traceability is part of the obligation. Italian Law 132/2025 adds a layer: companies must inform workers about the use of AI that affects them directly, which requires the criteria to be explicit and consultable, and therefore defined in advance.

Adoption grows, transformation lags

ISTAT data for 2025 show acceleration: 16.4% of Italian companies with at least 10 employees use at least one AI technology, up from 8.2% the previous year and 5% in 2023. Effective adoption tells a different story. A report by Anitec-Assinform with Politecnico di Torino, presented in April 2026, finds adoption concentrated in large firms: more than 50% of companies with over 250 employees use AI, against 15.7% of those with fewer than 50. Even among adopters, the distance between using a tool and transforming processes remains wide. AI4Business reports that in 2025, 42% of companies abandoned most of the AI initiatives they had started, and fewer than half of the prototypes reached large-scale production. Il Sole 24 Ore, citing the Observatories of Politecnico di Milano, summarises the problem: the organisational leap, rather than the technological one, is the critical variable. Only 5% of companies bring their AI prototypes into stable production. 79% of business leaders consider AI strategic, while only 20% of employees use it daily in their workflows. A course does not close that gap.

Why AI training programmes fail

Three mechanisms recur systematically in small organisations. The first is training on the tool instead of the redesigned process. Employees learn to query a platform and interpret its output, then return to unchanged processes. The tool is used episodically, replacing isolated tasks rather than integrating into a revised flow, and the expected efficiency does not appear because the real flow is unchanged. The second is the separation between training and operations. The classroom, physical or virtual, and daily work remain separate environments, and skills acquired with generic examples transfer poorly to specific processes, often in sectors where cases do not match standardised study material. The Anitec-Assinform report documents this fragmentation, with a training ecosystem still largely disorganised. The third and most underestimated is the absence of top management involvement. The Politecnico observatory notes that training in Italian SMEs concentrates on operational levels, while the engagement of owners and management is often missing. People are asked to change how they work without anyone redesigning the processes they operate on or defining how to measure the expected change.

What separates adoptions that hold

Companies that obtain concrete results invert the usual order of priorities: process first, then training on the redesigned process. Training follows an organisational decision rather than preceding it. Ramp, an American corporate finance platform, reached near 100% internal AI adoption by first building a unified access infrastructure that removed the organisational friction preventing daily use, and only then investing in contextual training. The goal was to make redesigned processes understandable to the people who operate them. The same principle emerges from European data. The Fivetran AI Readiness report for 2026 finds that only 15% of organisations are effectively ready for AI adoption at scale, and the main barrier is not technological: it is data governance, the quality of documented processes and clarity on measurable objectives. Training employees on systems that operate on unstructured data and implicit processes produces frustration rather than skills.

Compliance and effective adoption

Two objectives are often confused. Regulatory compliance means documenting that employees who use AI systems have received training adequate to their role and to the risk level of the tools adopted. It is reachable in a relatively short time, with an inventory of AI tools, a map of the roles involved and a traceable training path. It is necessary, and it is not sufficient. Effective adoption means integrating AI tools into operational processes to produce a measurable return in recovered time, reduced errors and better decisions. It requires more time, an analysis of existing processes, the redesign of at least part of the flows and training built on those redesigned flows. It is the longer path, and the only one that justifies the investment. Treating the second as an automatic consequence of the first is the most common mistake. Compliance opens the door, and effective adoption requires walking through it with a method.

Training is a legal duty before it is a practice, and the EU AI Act obligations for SMEs put the burden on the user.