Writing in IT Executive, Rad Parvin, Senior Value Partner at Valliance argues that businesses have bought AI but are not yet getting the return from it. Adoption has risen sharply while the share of organisations making a profit from AI has stood still. His explanation is that what works stays with a handful of individuals and never spreads across the organisation. The fix he sets out is to establish what really delivers value and make it available more widely, with the testing, management and agreements that let colleagues rely on it, and with leaders who have used AI themselves.
The article was published in Dutch. All quotes below are our translations.
Adoption has doubled, but the return has not grown with it
Rad starts from how ordinary AI has become at work. ChatGPT, Copilot and Claude are now as normal a part of the toolkit as Microsoft Office or Google Drive. Expectations were high, with less routine work, faster decisions and lower costs, but two years after the first big wave of rollouts, the reality is more stubborn.
He cites McKinsey's annual AI research, in which 88 per cent of organisations say they use AI structurally in at least one part of the business. The share actually making a profit from it is 37 per cent, almost exactly the same as a year earlier, and two thirds have not yet started scaling. He concludes that adoption has doubled while the return has not grown with it.
The Netherlands is both ahead and behind. Rad cites CBS figures showing that in 2025, 66 per cent of large companies used AI technology, against 30 per cent of SMEs. The sharpest rise is among companies with 50 to 250 employees, from 20 per cent in 2023 to 45 per cent in 2025.
'The licences are there. The question is what they deliver.' (translated from the Dutch)
Shadow AI grows when the official tools do not fit the work
While use lags, the bill keeps running. Rad describes employees turning to other tools, sometimes out of sight of IT, which is how shadow AI takes hold. He writes that almost half of AI use in the workplace now runs partly or entirely through private accounts. He is clear that this is not unwillingness. It happens because the official solution does not fit the work.
That is why he sees the problem as rarely technical. What matters is whether employees know what AI is useful for and whether it really helps them in their daily work.
The people getting results are rarely AI specialists
Rad points out that almost every organisation has successful pioneers. Someone uses a prompt to cut an afternoon of research down to half an hour. A colleague has an agent handle the weekly admin, something that was a demo last year and now simply works. These are often people who just tried things out.
They tend to keep their approach to themselves or within their team, whether they mean to or not. On the other side of the office, someone starts again from zero. The experiments pile up without the organisation itself working any differently.
'You only take the next step when you establish what really delivers value and make it more widely available.' (translated from the Dutch)
A good prompt becomes a way of working only when others can rely on it
A discovery only becomes valuable, Rad argues, when others can trust that it will work just as well for them. Colleagues need to be able to use someone else's prompt safely and consistently. The same goes for an agent that takes over manual work. It has to be tested, managed and widely used before the investment pays back, and that matters more now that agents act as well as advise.
This takes more than a prompting workshop. Organisations need to build applications, people need to learn to work with them, and there need to be agreements on data, control and responsibility.
'If you only train employees, you get end users of applications that don't exist. And only drawing up rules leads to reluctance.' (translated from the Dutch)
Leaders need to use AI themselves before they set the rules
Rad writes that AI policy is still too often shaped by presentations and demos, which produces rules that quickly turn out too strict or too vague. He argues Brussels did not help that reflex this summer. By his account, the Digital Omnibus, in force since the end of July, pushed the heaviest obligations for high-risk AI back to December 2027, and many companies read that as breathing room. He points out that the obligation to ensure employees are AI-literate has applied since February 2025 and has not been postponed, and that the transparency obligation came into force this month.
'Postponing the paperwork is not postponing competence.' (translated from the Dutch)
Leaders do not need to be able to code, he writes, but they do need to experience how convincing a wrong answer can sound and when human judgement remains essential. That understanding makes agreements more concrete, such as which data may be used, which outputs always need checking, and when an application is reliable enough for wider use. It also stops employees reaching for their private accounts out of convenience.
The return comes from sharing what works
As long as AI stays mostly with a small group of enthusiasts, Rad argues, the return stays limited. It produces isolated successes and no real change. The gain lies in sharing what works. When good applications find their way to other teams and become part of standard work processes, the investment starts to deliver real value.
'That 37 per cent isn't moving because the models fall short, but because organisations aren't spreading the gold their employees already hold across the shop floor. That doesn't take a new tool or training. It takes an organisation willing to go and collect it.' (translated from the Dutch)
First published in Dutch in IT Executive on 22 September 2026. Read Rad's full article, Bedrijven betalen voor AI, maar laten opbrengst liggen.
FAQs
Does the Digital Omnibus delay the EU AI Act's AI literacy requirement?
According to Rad, no. He writes that the Digital Omnibus pushed the heaviest high-risk obligations back to December 2027, but that the duty to ensure employees are AI-literate has applied since February 2025 and has not been postponed.
How should AI governance be set up so it does not hold adoption back?
Rad argues that rules on their own lead to reluctance, and that policy shaped by presentations and demos tends to be too strict or too vague. His answer is for leaders to experience AI themselves, so that agreements become concrete, covering which data may be used, which outputs always need checking and when an application is reliable enough for wider use.
Why are employees using private AI accounts instead of company tools?
Rad writes that almost half of workplace AI use now runs partly or fully through private accounts. He argues this is not unwillingness. It happens because the official solution does not fit the work. He sees the problem as rarely technical. It comes down to whether employees know what AI is useful for and whether it helps them in their daily work.
Why do AI experiments stall instead of changing how the organisation works?
In Rad's view, successful experiments stay within one person or team, and others start again from zero. The organisation only moves forward when it establishes what really delivers value and makes that more widely available, with applications that colleagues can use safely and consistently.
Why do so many AI programmes show activity but no profit?
Rad Parvin points to McKinsey research in which 88 per cent of organisations use AI structurally in at least one part of the business, while only 37 per cent make a profit from it, a figure that has barely moved in a year. His explanation is that what works stays with individual pioneers and is never tested, managed and spread across the organisation.
