What does AI training actually deliver? The cost of doing nothing

Around 2000, Europe gained a standard for basic digital skills. Nobody made it mandatory, and almost thirty years on the result is measurable. When it came to AI, the law appeared to have addressed the problem. Since 27 July 2026, however, that obligation has been significantly weakened.

A mistake we have made before

In 1997, European computer societies established a foundation in Dublin with a single aim: to raise the level of digital skills. The result was the European Computer Driving Licence, ECDL, today known as ICDL. A vendor-neutral standard with course material, examinations and a certificate demonstrating competence in widely used office software.

It still exists. More than a hundred countries, twenty thousand test centres. Over almost thirty years, seventeen million people worldwide have participated in the programme.

Seventeen million. Worldwide. Over almost thirty years. The European Union alone has almost two hundred million people in employment.

A number of governments rolled it out on a large scale, mostly outside Europe. But it never became a general requirement for anyone who had to work with a computer, and in most organisations it never became an established part of staff development or training policy. Companies bought the computers, they bought the software, and the people who had to use them every day largely worked it out for themselves.

The consequences can still be seen in the figures. In 2025, 60% of EU residents aged 16 to 74 had at least basic digital skills. The target for 2030 is 80%. That figure covers the population as a whole; the proportion is somewhat higher among people in employment. Even so, almost thirty years after the computer driving licence was introduced, four in ten Europeans fall short of the basics.

Was that a disaster? It was survivable. Poor digital skills mainly cost time, and where things did go wrong the mistakes were usually visible: an email sent to the wrong person, a formula that did not add up. Annoying and expensive. But recognisable. And most competitors were no further ahead. Everyone muddled through, and muddling through became the norm.

This time the law had it covered

For AI, the lesson appeared to have been learned. From 2 February 2025, Article 4 of the EU AI Act required what had been entirely missing around 2000. Providers and deployers of AI systems had to take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff and among anyone operating those systems on their behalf. The measures had to reflect the person's knowledge and experience, as well as the context in which the system was used.

It was not a recommendation and it was not optional. An active duty to ensure, with a target built into it: a sufficient level. Doing nothing became hard to defend.

That was exactly the right move. Not because regulation is pleasant, but because it settled the argument inside organisations. If the law asks you to do your utmost to make sure your people know what they are doing, there is no longer any need to debate whether training is necessary.

Since 27 July the bar has been lowered

On 24 July 2026, Regulation (EU) 2026/1744, the Digital Omnibus, appeared in the Official Journal. Three days later it entered into force, with no transitional period for this element.

Article 4 now reads differently. The words "ensure", "to their best extent" and "sufficient level" have gone. What remains is the duty to take measures that support the development of AI literacy. And it now states explicitly that providers and deployers are no longer required to guarantee that any individual member of staff reaches a particular level.

From ensuring to supporting. From a level to be reached to a development to be encouraged.

The heavier version applied for almost eighteen months. Less than a week before national enforcement began, the bar was lowered.

Note what did not happen. The obligation has not disappeared. Organisations still have to act, and those measures must match the roles, the systems and the risks. A newsletter with no connection to any of that will not do the job.

What has disappeared is much of the pressure. A standard that says "make sure your people can do this" forces a decision. A standard that says "support the development" leaves considerable scope for a minimal response. And experience tells us which way organisations go once that scope exists.

Nor should the risk of penalties be either exaggerated or dismissed. Article 4 never carried its own specific maximum fine under the AI Act. But national supervisors can certainly act, and for high-risk systems there is the additional requirement that human oversight be carried out by people with the necessary competence and authority. That obligation does fall under the penalty provisions. And if a member of staff causes damage because they did not know what they were doing, whether the employer met its duty of care may become a matter of employer liability under national law.

So the legal bar has been lowered. The risk has shifted; it has not disappeared.

Why the consequences could be far greater this time

Software proficiency mainly enabled people to perform the same work more quickly. Anyone handy with a spreadsheet produced the same table in ten minutes instead of thirty. Useful. But it was still the same table.

AI skill also changes what a person is capable of. Someone who is unable to write confidently in French has a usable first draft of a quotation in French within a minute. Someone facing eighty pages of meeting minutes ends up with a workable summary.

That is not an assumption. Researchers from Harvard, Wharton, MIT and Warwick ran a pre-registered field experiment with 758 consultants together with Boston Consulting Group. On tasks the model was suited to, participants with AI completed 12.2% more tasks, worked 25.1% faster, and their work received quality scores that were approximately 30% higher.

A second study, published in the Quarterly Journal of Economics, followed 5,172 customer support agents. They handled on average 15% more queries per hour. But the average hides the most interesting part: the least experienced staff improved both their speed and their quality, while the most experienced colleagues saw little improvement in speed and a slight dip in quality.

The largest gains are therefore seen among employees who do not regard themselves as technically confident. They are precisely the employees least likely to start experimenting with AI on their own.

But there is another side to the evidence. That same experiment also included a task deliberately designed to fall outside the model's capabilities. The group without AI reached the right conclusion in 84.5% of cases. For the groups with AI, that fell to around 65%. They had help and performed worse than those who worked without it.

The most uncomfortable part comes last. Graders were shown the same answers without knowing which were correct, and asked how persuasive and well argued they were. The answers produced by the AI-assisted groups were rated as significantly more persuasive and better argued, including where the conclusion was wrong.

That is the difference from thirty years ago. Back then, incompetence was slow. Now it is invisible.

Knowing how to use AI is not the same as knowing when to trust it

The experiment included a third group. Alongside the AI, they received guidance on how to prompt the model effectively. On the tasks the model was suited to, they delivered work of significantly better quality than the group given only the tool.

But on the task beyond the model's reach, that same group performed worst of all: 60% correct conclusions, against 70.6% for the group with only the tool.

That looks contradictory and it is not. The guidance covered how to use the model, not how to judge the output. Anyone who learns to ask better questions gets better answers and, as a result, trusts the system more. That trust is exactly what proves damaging the moment the model moves beyond what it can do.

That is the key lesson. Learning to use a system without learning to judge it makes the risk bigger, not smaller. Training that only covers prompts does not solve the problem. What staff need is an understanding of where such a model is reliable, where it is not, what information may safely be entered into it, and when an answer has to be checked before it is used, shared or acted upon.

The decision is back with organisations

Around 2000 there was a standard, an examination and a certificate. Almost no organisation turned it into policy. Not because it was poor. But because nobody required it and it could always wait until next year. That next year never arrived.

With AI, the amendment has moved us significantly back in that direction. The law no longer prescribes what level your people must reach. It is now largely up to each organisation to decide what measures to take. And that is exactly the kind of decision that, around 2000, so often ended up being postponed.

Except that the consequences are heavier this time. People in your organisation are already working with AI, often without anything having been agreed. The question is whether they have enough basic knowledge to do so responsibly. Knowing where a model is reliable, what information may safely be entered into it, and when an answer needs checking. Since 27 July, that level is no longer a legal requirement. The responsibility, however, remains firmly with your organisation.

At AIAdopt we provide short online training modules for exactly this, with an exam and a certificate for each member of staff. No unnecessary technical detail and no legal jargon, but usable on the job straight away. And if you would rather read further first: our Insights cover regulation, certification and the use of AI across different sectors. Freely available, no registration.

Frequently asked questions

What changed on 27 July 2026 regarding the AI literacy obligation?

Regulation (EU) 2026/1744 amended Article 4 of the EU AI Act. Where providers and deployers of AI systems previously had to take measures to ensure, to their best extent, a sufficient level of AI literacy, they must now take measures that support its development. A sufficient level is no longer the legal benchmark, and no particular level has to be guaranteed for any individual member of staff. The obligation itself remains.

Do I still need to train my staff?

Yes. The duty to take measures remains, and those measures must match the roles, the systems and the risks in your organisation. The law simply no longer prescribes a level to be reached. More importantly, the law was never the reason to train. Research shows that AI used by unprepared staff can lower the quality of work without anyone noticing.

What does AI training deliver for my staff?

In a pre-registered experiment with 758 consultants, participants supported by AI completed 12.2% more tasks and worked 25.1% faster, producing work that received quality scores approximately 30% higher. That gain applied to tasks the model was suited to. Participants who also received guidance material delivered work of significantly better quality than those given only the tool.

Which staff benefit most from AI training?

Less experienced employees and those who initially perform less strongly. In a study of 5,172 customer support agents, they improved both their speed and their quality, while the most experienced colleagues saw little improvement in speed and a slight dip in quality. They are also usually the employees least likely to start using AI on their own.

Is a course on writing prompts enough?

No. The group given guidance on how to prompt the model effectively did better where the model was suited to the task, and worst of all on the task beyond its reach: 60% correct conclusions, against 70.6% for the group with only the tool. Learning to use the system more effectively can also increase trust in its output, and that trust is precisely what becomes risky where the model falls short.

Sources

Rob Ummels, AIAdopt. AI adoption, AI literacy and practical guidance for organisations that want to deploy AI sensibly.

Written by Rob Ummels in collaboration with Claude (Anthropic). Editorial responsibility: AIAdopt.