This piece first appeared in German. The English version is a rewrite rather than a line-by-line translation, and the German original stays online in the archive: KI-Skills: erst Gehaltsprämie, jetzt EU-Gesetz.

On 2 August 2026, national market surveillance authorities across the European Union began supervising and enforcing Article 4 of the AI Act. The duty itself is older and has applied since February 2025. What changed this summer is that somebody now checks, and that the duty became lighter a week earlier. A company now has to show that it took measures to support its staff's AI literacy.

Authorities are the smaller half of the story. Two forces are pulling in the same direction at once, which they seldom do. Regulation has made AI competence a legal duty of the employer. The labor market has made it the most valuable single line an application can carry.

Article 4 now asks for measures

The text is shorter than its effect suggests. Since the Digital Omnibus on AI, providers and deployers of AI systems have to take measures to support the development of AI literacy among their staff and anyone acting for them. The amended article states that it does not require any specific level of AI literacy of any individual. No uniform certification is required. What holds up in practice is a documented training plan proportionate to the risk.

Deployer covers almost every company. Use ChatGPT or Copilot in your work and you are typically operating an AI system within the meaning of the regulation. There is no de minimis threshold, so a mid-sized firm with fifteen employees has to act as well, scaled to its size, risk and context.

Article 3(56) defines what competence means: the skills, knowledge and understanding that allow informed use of AI systems and an awareness of the opportunities, the risks and the possible harm. That is deliberately open. In its questions-and-answers document the Commission makes two things clear. Article 4 carries no obligation to measure employees' AI knowledge, and after the Omnibus no specific or sufficient level is mandated.

The flexibility has a floor, though. Relying on the system's instructions for use, or asking staff to read them, might be ineffective, the Commission warns. It points to what a sensible program considers instead: a general understanding of AI across the organization, the company's role as provider or deployer, and the risk of the systems it uses.

The penalty is indirect, which makes it easy to underestimate

Honesty requires the other half. The AI Act sets no specific fine for missing training, and Article 4 does not appear in the penalty catalogue of Article 99. National authorities can still act case by case, and the Commission notes that they could impose proportionate penalties. The 35 million euros that get quoted in this context belong to prohibited practices such as social scoring, and they have nothing to do with the training duty.

Risk runs along a quieter path. A breach of Article 4 is unlikely to be pursued on its own, and more likely to weigh as an aggravating factor in wider enforcement proceedings. Civil law adds the rest. Where missing AI competence produces damage, say a faulty automated decision in hiring or a data protection breach through uncontrolled input, the absent training can be read as a failure of due care.

Brussels has lowered the bar

In its Digital Omnibus proposal of 19 November 2025, the Commission wanted to move the obligation from organizations to the member states and to itself. Deployers of high-risk systems would have kept the duty to train for human oversight.

The law that came out of the negotiations keeps Article 4 as a company duty with a lower bar. Regulation (EU) 2026/1744, adopted on 8 July 2026 and in force since 27 July, has companies take measures to support AI literacy. Before, they had to ensure a sufficient level of it. The Commission and the member states now have to support that effort, with particular attention to smaller firms. For deployers of high-risk systems, the duty to train staff for human oversight remains.

Anyone reading this as the duty disappearing is confusing a fall in legal pressure with a fall in economic pressure, and the second one is stronger.

The market is paying 62 percent

PwC published its Global AI Jobs Barometer in mid-June 2026, an analysis of more than a billion job advertisements across 27 countries. The average wage premium for workers with AI skills has reached 62 percent. Roles that ask for specific AI capabilities, prompt engineering among them, grew by 69 percent while the job market as a whole grew by 9, roughly eight times the pace.

The reading that belongs beside it: PwC sells AI transformation consulting and has a commercial interest in impressive numbers. Its data measures what employers advertise, and not what they end up paying. PwC is also not quite consistent with itself, because the 2025 edition put the premium at 56 percent and the 2026 edition restates that same year as 57, probably a data revision. Its 2024 edition reported premiums of up to 25 percent in selected markets, a different measure. Across the editions, the direction is unambiguous all the same.

Where does a premium like that come from? Scarcity. The Microsoft AI Economy Institute reports that 17.8 percent of the world's working-age population, people between 15 and 64, had used a generative AI product in the first quarter of 2026. A quarter earlier the figure was 16.3 percent.

Such a measure captures breadth of use and not depth. The order of magnitude carries the point anyway. Between what companies report as AI adoption and what individuals can actually do there is a gap, and the premium lives in it.

Three skills, and a degree is not one of them

The mistake at this point would be to think of a computer science degree. Article 4 asks for no deep technical grasp of transformer architectures or token embeddings. Employers ask for it even less. What counts is demonstrable practical skill, and three kinds of it currently have the best ratio of effort to market value.

The first is designing AI-supported workflows, repeatable processes in which a model takes defined steps and a person holds defined control points. That is process thinking. It can be learned in weeks and it involves no programming.

The second is tool integration through the Model Context Protocol, or MCP, the open standard that lets AI models reach calendars, documents and company data. It has arrived in everyday office software. Microsoft has made MCP support generally available in Copilot Studio, and MCP-based agents have since reached Microsoft 365 Copilot itself, which lets developers wire their business processes in with much less work. An agent built there can use the same tools as Claude, ChatGPT and other MCP-compatible platforms.

Connecting an agent in Word, Outlook or Excel to a data source is a skill that did not exist eighteen months ago. Competition for it is thin in proportion.

The third is programming with AI assistance, as an extension of a job and not a change of career. Small automations, analysis scripts and prototypes used to go to the IT department. Specialists increasingly build them themselves.

The duty sits with the employer, the competence sits in the employee

Here the two forces meet. An employee who arrives with documented evidence of practical AI skills solves two problems in one move: their own pay problem, and their employer's compliance problem. In a dispute, the written record of who was trained on which AI topics and when is what proves that the Article 4 duty was met. Training has rarely been this directly usable from both sides.

The window is real and it is finite

Premiums of this size attract supply. The history of digital skills, from web design in the nineties to data analysis in the 2010s, suggests that the head start of the early movers normalizes in something like a decade. Build the skills in 2026 and you buy cheap. Wait until market surveillance rings the bell, or until the job advertisement takes them for granted, and you pay the full price.

ThinkBeyondAI runs training and certification on AI fundamentals and on the three skills above, together with Prof. Inhwa Kim and Shay Luo, for individual employees, apprentices and students as well as for companies and teams.

Sources

Grouped by the section they support.

Opening

  • European Commission (2026), AI Literacy – Questions & Answers, updated 27 July 2026, digital-strategy.ec.europa.eu. Supports the application of Article 4 since 2 February 2025 and the start of supervision and enforcement by national market surveillance authorities in August 2026.
  • Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), Official Journal of 24 July 2026, eur-lex.europa.eu. Supports the lighter duty and its entry into force on 27 July 2026.
  • The two forces pulling in the same direction are the author's own reading.

Article 4 now asks for measures

  • Regulation (EU) 2026/1744, amending Article 4 of Regulation (EU) 2024/1689. Supports the duty to take measures to support AI literacy and the statement that no specific level is required of any individual.
  • Regulation (EU) 2024/1689 (AI Act), Article 3(56), eur-lex.europa.eu. Supports the definition of AI literacy.
  • European Commission (2026), AI Literacy – Questions & Answers. Supports the absence of any duty to measure employees' knowledge, the absence of a mandated level, the warning about relying on instructions for use, and the elements a program should consider.
  • The documented training plan and the reading that almost every company is a deployer are the author's own assessment.

The penalty is indirect, which makes it easy to underestimate

  • Regulation (EU) 2024/1689, Article 99. Supports the penalty catalogue and the fines of up to 35 million euros for prohibited practices under Article 5.
  • European Commission (2026), AI Literacy – Questions & Answers. Supports the point that national authorities could impose proportionate penalties.
  • Treatment as an aggravating factor and the civil-law risk are the author's own assessment.

Brussels has lowered the bar

  • European Parliament, Digital Omnibus on AI, Legislative Train Schedule, europarl.europa.eu. Supports the date of the Commission proposal, 19 November 2025, and the course of the procedure.
  • Morrison Foerster (2025), EU Digital Omnibus on AI: What Is in It and What Is Not?, mofo.com. Supports the proposal to turn the literacy duty into an obligation of the Commission and the member states.
  • Regulation (EU) 2026/1744. Supports the adopted text, its dates and the new support role of the Commission and the member states.
  • European Commission (2026), AI Literacy – Questions & Answers. Supports the remaining duty to train staff for human oversight of high-risk systems.

The market is paying 62 percent

  • PwC (2026), AI reshapes global labour market into two distinct paths, rewarding human skills: PwC 2026 Global AI Jobs Barometer, press release, 15 June 2026, pwc.com. Supports the more than one billion job advertisements in 27 countries and territories, the 62 percent premium, the restated 57 percent for the previous year, and the 69 percent growth against 9 percent.
  • PwC (2025), AI linked to a fourfold increase in productivity growth and 56% wage premium, press release, pwc.com. Supports the 56 percent in the 2025 edition.
  • PwC (2024), AI-exposed sectors experience productivity surge as AI jobs climb and see up to 25% wage premium, press release, pwc.com. Supports the premiums of up to 25 percent in the 2024 edition.
  • Microsoft AI Economy Institute (2026), Global AI Diffusion Q1 2026: Trends and Insights, May 2026, microsoft.com. Supports the rise from 16.3 to 17.8 percent of people aged 15 to 64 who used a generative AI product.
  • The caveats on PwC's interest and on breadth versus depth are the author's own.

Three skills, and a degree is not one of them

  • Model Context Protocol, What is the Model Context Protocol (MCP)?, modelcontextprotocol.io. Supports MCP as an open standard for connecting AI applications to external systems.
  • Desai (2025), Model Context Protocol (MCP) is now generally available in Microsoft Copilot Studio, Microsoft Copilot Blog, 29 May 2025, microsoft.com. Supports general availability in Copilot Studio.
  • Microsoft (2026), Build a plugin for a declarative agent from an MCP server, Microsoft Learn, learn.microsoft.com. Supports MCP-based agents in Microsoft 365 Copilot.
  • The choice of the three skills is the author's own assessment.

The duty sits with the employer, the competence sits in the employee

  • No external source. The argument is the author's own.

The window is real and it is finite

  • No external source. The comparison with earlier digital skills is the author's own observation.
The link has been copied!