AI literacy is mandatory. But a training course won't change how your teams work.

"We're already doing something with AI." It's a phrase we hear often, and it's usually true. You've rolled out licenses. Everyone has access to an AI assistant. There's an AI policy, neatly established and shared. But if you ask six months later what it's actually delivered, the room goes quiet.

Sound familiar? You're not alone. Since August 2, 2026, there's a second question to consider, because from that date the regulator can hold you accountable for your employees' AI literacy. Two questions that seem similar, but aren't. The first is about what AI actually gets you, the second is about the law. In this blog, we explain why you can answer the second question without doing anything about the first, and why that's exactly the pitfall.

Artificial Intelligence

Author: Allard Rothengatter

Published: 20 August, 2026

What Article 4 requires of you now

Article 4 of the European AI Regulation has applied since 2 February 2025. It requires both providers and deployers of AI systems to ensure AI literacy among their staff, proportionate to role, level of knowledge and context of use. No lower limit on company size, no exemption for low risk categories: if you use ChatGPT, Copilot or Claude in your work processes, you fall under it. Supervision started on 2 August 2026.

Why is this regulated at all? The thinking behind Article 4 is that AI systems make mistakes that don't look like mistakes. A candidate rejected because of a pattern in old data. A summary in which one figure is just slightly wrong. An answer that sounds fluent and convincing and is factually incorrect. If you don't know that can happen, you don't check for it. The legislator therefore places the responsibility with the organisation deploying the system: make sure the people working with it understand what they have in front of them.

What gets mentioned far less often: Article 4 was rewritten a week before that date. The Digital Omnibus, Regulation (EU) 2026/1744, has been in force since 27 July. Where the old text asked organisations to ensure a sufficient level, the new text asks for measures that support its development. The Commission sums it up this way: the obligation remains, but no specific level is prescribed, and you don't have to guarantee any particular level for any individual.

So how is this enforced? The honest answer: that isn't clear yet. Article 4 carries no fine of its own under the regulation. In its Q&A, the European Commission states that national supervisory authorities can enforce and impose penalties, but this has to be arranged in national legislation, and the Dutch implementing act for the AI Regulation is still in preparation. The regulation also creates no criminal offences and no right to compensation.

What remains is this. The Commission notes that a penalty becomes more likely where there is evidence of an incident caused by inadequate training and guidance. Not as a standalone infringement, then, but as a circumstance at the moment your organisation is being looked at for another reason. No certificate is required, and no AI officer either: an internal record of what you have done is enough.

In short: the obligation stands, the bar sits lower than it did six months ago, and anyone who bought a course to avoid a fine bought something against a risk that works differently than the brochure suggested.

"Complying with Article 4 takes an afternoon. Actually benefiting from it means changing the work itself."

– Allard Rothengatter

AI literacy is the minimum, not the goal

The question "are we compliant with Article 4?" and "is AI delivering value for us?" are often confused in practice, but the answer to the first tells you little about the second.

AI literacy is about understanding what an AI system is, what risks are involved, and what harm it can cause. Knowing that a model can sound convincing while giving a wrong answer. Knowing what data you should and shouldn't input. That’s the minimum to keep AI from being dangerous.

Beyond that is a second concept, rarely used in the Dutch conversation: AI fluency. The framework behind it was developed by Rick Dakan and Joseph Feller in collaboration with Anthropic, and it's not about knowing the risks but about mastering the collaboration. Not: what can go wrong. But: when do I use this, how do I phrase what I need, how do I assess the output, and where does my own responsibility remain.

Literacy prevents harm. Fluency creates value. And the numbers show that almost no one has reached that second stage yet.

Why it doesn't stick

The Global CEO Study 2026 by IBM shows that only a quarter of employees use AI regularly at work, while 86 percent of CEOs surveyed believe their people have the necessary skills. So skills aren't where things go wrong, at least not according to those making the decisions.

Deloitte surveyed over three thousand executives and came to the sharpest summary of this article: companies focus on building AI fluency instead of redesigning work. 84 percent have not restructured roles around what AI can do.

Closer to home, research by Sharp among Dutch SMEs shows that 97 percent have an AI policy, while nearly half of employees use AI tools without reporting it. Policy on paper, shadow use in practice.

You can't solve that pattern with a license. And not with a certificate, either.

The four competencies of AI fluency

The framework distinguishes four skills that together determine whether someone actually gets value from AI.

Delegation

Knowing what you delegate

And especially: what you don't

The first question isn't how to ask AI to do something, but whether you should. Which part of a task can you hand over, and which part requires judgment you shouldn't outsource? This is the competency most often skipped—and the one that costs the most if you do.

Description

Clearly stating what you need

Context matters more than wording

A good prompt includes the goal, the boundaries, and the background. In practice, we see employees ask questions as if they're using a search engine, only to be surprised by a generic answer. This is a skill you can learn, and the effect is immediately noticeable.

Assessment

Assessing if it's correct

Convincing isn't the same as correct

Evaluating output requires subject matter expertise, not AI knowledge. If you don't master the topic, you can't see where things go wrong. That's the limit of what AI can deliver: without in-house expertise, using AI is just guesswork with extra steps.

Diligence

Staying accountable

For what you use and how you use it

What goes out under your name? What data do you share? Where do you record that AI was used? This is the competency where legal obligations and business interests meet.

Additionally, the framework distinguishes three ways people work with AI: automation (AI executes what you instruct), augmentation (AI as a thought partner), and agency (configuring AI to independently handle recurring tasks). That third form is where things get interesting—and also where standalone tools fall short.

What fluency solves—and what it doesn't

What AI fluency does deliver

Employees who know when to use AI and when not to, who formulate better prompts, critically assess output, and take responsibility for what leaves the organization. Less shadow use, fewer throwaway results. And as a bonus, it covers much of what Article 4 requires.

What AI fluency doesn't deliver

A working process. A skilled team working in isolated, unconnected tools still copies data from one screen to another. The gains stay in one person's head and leave when they do. Skill without workflow is an individual achievement, not organizational capacity.

From skill to workflow

This is how we work ourselves. Every two weeks, the entire team gathers for an AI sync-up: where do we stand with AI adoption, in the code and everything around it. Not because there's something to celebrate, but because the answer changes every two weeks. We record our findings in Confluence, so everyone is up to date and we build a track record of the choices we made and why. We have the same ongoing conversation with our clients: where do we use AI, where do we deliberately not, and what does that mean for how their software is set up.

That's the core of why training alone doesn't solve it. AI fluency isn't a box to check off, it's an ongoing conversation. And at some point, that conversation gets translated into code: the same checks on AI-generated code as on manual work, the same roles and permissions, the same documentation. A more powerful model doesn't take over that work; it just makes it more important.

Frequently asked questions about AI literacy

Yes. Article 4 of the AI Regulation applies from February 2, 2025, to both providers and users responsible for AI systems. Company size and risk category do not matter: if you use AI tools in your work processes, you are subject to the regulation. The required level is proportional and depends on your role, knowledge, and the context of use.
The Digital Omnibus, Regulation (EU) 2026/1744, came into effect on July 27, 2026, and revised Article 4. The obligation to ensure an adequate level has been replaced by a duty to take measures that support the development of AI literacy, without requiring you to guarantee a specific level for any individual.
Article 4 does not specify a separate fixed EU fine, but the European Commission states that national regulators can enforce and impose sanctions in case of violations. The specific powers in the Netherlands will depend on the upcoming Implementation Act for the AI Regulation.
No, not in that sense. The law requires measures that fit the context. This could be training, but also internal guidelines or on-the-job coaching. The Commission does note, however, that simply referring to the user manual of an AI system will often not be effective.
AI literacy means understanding what AI is and what can go wrong: the minimum required by law. AI fluency is the skill to work effectively with AI, built on delegation, description, evaluation, and diligence. Literacy prevents harm, fluency creates value.

The law asks less. Your organization doesn't.

If you've only been watching the law in recent months, you saw a strange shift: a requirement about to be enforced, and at the same time being watered down. It's tempting to conclude that it's not a big deal after all.

That would be a shame, for a reason that has nothing to do with the law. The organizations that get real value from AI don't differ from the rest in which model they use or how many licenses they've bought. They differ in how much they've actually redesigned their workflows.

Just like with everything we build: the technology is rarely the hard part. The choices around it are.

Want to know more about AI literacy?

Want to discuss where AI workflows are stalling in your organization, and which issues are training-related and which are software-related? Get in touch. We're happy to help, even if the outcome is that you don't need us.

More interesting blogs

The Power of ChatGPT

25 Jan 2023 by Allard Rothengatter

ChatGPT is a smart chatbot developed by OpenAI. The model is trained on millions of texts from the internet and is able to understand and generate language in a way that was previously impossible. In this article, we will discuss some of the main benefits of ChatGPT and how it can change the way we communicate and process information.

ChatGPT o1: The new standard in AI language models?

17 Sep 2024 by Allard Rothengatter

In the dynamic world of artificial intelligence, one name is synonymous with innovation: OpenAI. With the launch of ChatGPT o1, also known as the “Strawberry Model,” they are taking the next step forward in AI-driven communication. This advanced language model not only promises a leap in efficiency and flexibility but also introduces various additional capabilities that will inspire both clients and developers.

ChatGPT o1 is designed to seamlessly align with the needs of modern organizations. Whether it’s improving customer interactions, accelerating development processes, or discovering new forms of creativity, this model offers a range of possibilities that raise the bar for what AI can achieve.

ChatGPT-4: What We Know and What We Can Expect

28 Feb 2023 by Alan Linnenbank

The release of OpenAI's GPT-3 language model in 2020 marked a major milestone in the development of artificial intelligence. With its impressive 175 billion parameters, GPT-3 demonstrated unprecedented capabilities in natural language processing and text generation. But what comes next? Will OpenAI continue to push the boundaries of AI language models with ChatGPT-4, and if so, what can we expect?