AI mainly changes tasks, and hardly any whole jobs yet. The work is still software, film, healthcare or art. What shifts is how it gets made, who can make it and what people check. Those who use AI well and review it critically make the difference.
In 2025, 17 percent of Dutch companies used AI, compared with 8 percent in 2023. Among companies with 250 or more workers it was 66 percent, among companies with 2 to 10 workers 14 percent (provisional figures).
Danish research that links surveys to wage data finds no measurable effect on earnings or hours worked two years after ChatGPT. What does change is the content of the work: new tasks in creating content, overseeing AI and integrating AI.
In three Dutch case studies, simple tasks disappear and new tasks appear in checking and validation. At Deloitte Legal, the work of junior staff shifts from research to review, with a time saving in contract analysis of 40 to 50 percent by their own estimate.
Small businesses: better performance, same number of people
65 percent of small and medium-sized businesses that use generative AI see better performance from employees. 83 percent say it has no effect on the number of people they need.
In the United States, employment of 22- to 25-year-olds in occupations with high AI exposure is 19 percent below that in comparable occupations, mainly because fewer young people are being hired. Where AI complements people rather than replacing them, that effect does not appear. The researchers call these early signals, not a proven cause.
Expectations, not facts. They differ widely, and we say so.
Expectation
Most jobs change, few disappear
According to the ILO, one in four people worldwide works in an occupation with some exposure to generative AI. Because human input remains necessary, the ILO expects most jobs to change rather than disappear. Occupations in media and web score higher than before, due to better speech, image and video.
Between the mid-1990s and the late 2010s, almost one in five tasks in the Netherlands was automated, and roughly as many new tasks were added. The CPB expects AI, too, to bring both disappearing and new jobs; the distribution across groups is not equal.
The SER expects tasks and work processes to change, new work to emerge and some jobs to disappear. Whether AI contributes to broad prosperity depends on choices made by employers, workers and government, such as involving employees early and continuing to learn.
73 percent of AI experts expect a positive effect on how people do their jobs, compared with 23 percent of the public. That gap of 50 points is itself a reason to be careful with big promises.
AI agents that work on a computer by themselves went from 12 to about 66 percent of tasks completed on the OSWorld test in one year. They still fail about one time in three, so someone needs to keep an eye on them.
Organisations absorb AI by dividing tasks differently. New tasks appear, such as creating content with AI, overseeing it and building AI into existing processes.
AI wins a gold medal at the Mathematical Olympiad but reads an analogue clock correctly in only 50.1 percent of cases. Knowing where AI can and cannot be trusted is becoming a core skill.
Since 2 February 2025, organisations that develop or use AI must make sure their people have enough knowledge and skill to use AI responsibly (Article 4 of the AI Act). The law leaves open exactly which measures to take.
Since 2 August 2026 the transparency rules of the AI Act apply: chatbots must make clear that they are AI, and deepfakes must be labelled visibly or audibly. For the automatic marking of AI content by existing AI systems, a transition period runs until 2 December 2026.
Almost a third of small and medium-sized businesses use generative AI. Sole traders most often say they outsource less work as a result (16.5 percent). For makers that means: the value lies in what a client cannot do on their own with a chatbot.
Open and self-hosted models are becoming fully capable
The gap between open and closed models shrank in one year from 8 to 1.7 percent on some tests, and the cost of a model at GPT-3.5 level fell more than 280-fold. That makes models you run yourself or in Europe a realistic option.
GPT-NL, from TNO, SURF and the NFI, is trained only on data for which permission was given. Government projects are now trialling the model; commercial use will follow later.
Dutch language models of our own. GPT-NL is running with its first government users; teams are building it into their own applications. Digitale Overheid (Dutch government)
Reading and writing text in business software. The most widely used kind of AI at Dutch companies, according to Statistics Netherlands (CBS). CBS
filming Directing images you create yourself.
Script, generated shots and editing by one maker or a small team, with real footage where it matters.
Made example
Product film for a bicycle repair shop
Script, images and editing by one maker, music by a colleague.
Offered example
From storyboard to shot list
A tool a maker built for their own use and now offers to others.
The Night Watch complete. The Rijksmuseum reconstructed the cut-off parts using neural networks. Rijksmuseum
Painting after the algorithm. Pixel Pioneers shows work painted by hand from generated images. Boijmans
improving Making what already exists work better.
Makers use AI to look at an existing product, process or piece of work and make it faster, smarter or better looking, together with the people who use it.
Made example
Customer service that answers faster
Existing answers reviewed and an assistant built for the team.
Offered example
A week looking over your process
A maker looks at where AI really helps and where it does not.
The examples show what the work can look like; they are not real assignments or makers. The applications under "Where it's already happening" are real, with a source.