The patterns behind the data
I analysed 103 occupational groups on AI impact. 5.8 million jobs. What I found was not always what I expected. On this page I share the patterns that only become visible when you put the data side by side.
Three observations from the data: HBO and university graduates score higher on AI exposure than vocational (MBO) graduates. Middle incomes (€50–75K) are more vulnerable than the lowest and highest. And occupations with the highest AI score sometimes grow the fastest. Reality is more nuanced than the dominant narrative suggests.
The job balance: growth vs decline
The central question: are jobs being added or lost on balance? The answer is more nuanced than you might expect. The net shift is limited. But beneath that surface a substantial redistribution is taking place.
54 occupations are growing (together +178,700 jobs). 43 occupations are shrinking (together -252,350 jobs). 6 occupations remain stable. The net effect: a decline of approximately 73,650 jobs out of a total of 5.8 million. That is -1.3%. No mass job loss. But an enormous shift from office work towards healthcare, construction and technology.
+149K
-223K
-74K
Healthcare and social work grows fastest in absolute numbers (+68K jobs). The business and administrative sector shrinks the hardest (−153K). That is no coincidence: care work is physical and relational, administrative work is precisely what AI does well. Construction grows due to the energy transition (+27K), independent of AI. Creative and media shrinks the most in percentage terms (−16%), driven by AI content generation.
The 61% problem
AI impact is not a broadly distributed phenomenon. It is surgically precise. Five occupational groups account for 61% of all expected job losses. That is 155,080 out of 252,350 declining jobs. If you want to address the AI transition, start here.
This means AI transition policy does not need to be generic. These are specific roles, in specific sectors, with specific reskilling pathways. Administrative staff, customer service and secretaries share comparable skills and can be guided towards care coordination, installation technology or IT support.
Under the surface: 431K jobs in motion
The net figure (−73,650 jobs) suggests a modest shift. But it masks what is really happening. The gross movement is 431,050 jobs — 7.5% of the working population in this dataset.
+179K
54 occupations
-252K
43 occupations
431K
7.5% of total
The net figure is what economists report. The gross figure is what people experience.431K workers in a job that is changing or disappearing — that is more than the working population of Utrecht. Even when the balance is close to zero, the transition is enormous.
Three automation waves
Not all occupations are hit at the same time. The data reveals three waves, defined by the combination of AI exposure and NL vulnerability.
These are occupations where the technology exists and protection is minimal. Microsoft Copilot replaces administrative work, AI chatbots take over customer service, GitHub Copilot writes code. No policy is needed to make this happen. It is already happening.
Managers, marketers, policy officers, HR specialists. AI exposure is high but regulation, social preference or labour market tightness slows adoption. This wave accelerates once organisations treat AI as standard rather than experiment.
Healthcare, construction, installation, education, security. Physical work, relational work, regulated work. Here AI strengthens the professional rather than replacing them. These occupations are also growing due to ageing, energy transition and societal demand.
AI risk by sector
Individual occupations tell only part of the story. At sector level the patterns become visible. Creative work, media and the business and administrative sector score the highest. Construction, installation and healthcare score the lowest. That is no coincidence. AI excels at language, analysis and communication. Bricklaying, nursing or electrical installation fall outside what current AI models can handle. That line runs through the entire dataset: the more the work revolves around information processing, the higher the exposure.
Top 10 gainers and decliners
Forecasts are deliberately presented as ranges. Exact percentages suggest a certainty that does not exist. The direction is more reliable than the number itself. UWV signals, for example, that the ICT shortage is easing, which fits the conservative end of the range. What stands out: gainers are spread across healthcare, technology and ICT. Decliners concentrate in administrative and commercial work. Office work is shrinking. The shop floor and the screen (with the right skills) are growing.
Strongest gainers
Strongest decliners
The paradox: high AI score, yet growth
A group of occupations scores 7 or higher on AI exposure. Core tasks are technically automatable. And yet they are growing towards 2030. That seems contradictory, but the explanation differs per occupation.
Software developers become more productive through AI (GitHub Copilot, Claude Code, Cursor), but at the same time every organisation wants to build more software. Demand grows faster than productivity rises. For cybersecurity specialists it is even clearer: threats are increasing exponentially. AI helps with defence, but does not replace the human specialist.
The lesson is relevant for every occupation: if the underlying market demand is structurally growing, AI exposure can be more of a productivity boost than a threat. That is the Solow Paradox in miniature. The technology is there, but the effect on jobs depends on the context.
Pipeline-bouw en data-transformaties worden geautomatiseerd. Architectuurkeuzes en databeheer minder.
Who is hit hardest?
The common view is that AI mainly threatens low-skilled work. Robots in the factory, self-checkout tills in the supermarket. But the data shows something different. HBO-level occupations score an average of 6.0 on AI exposure, university-level occupations 6.1. Vocational level-1 scores 3.6. That difference is substantial. The explanation: AI is exceptionally good at precisely what knowledge workers do. Producing text, analysing data, writing reports, streamlining communication. A nurse, an electrician or a carpenter does work that largely falls outside the reach of current AI models.
With salary the same pattern emerges. The €50,000 to €75,000 group has the highest average AI exposure (6.6). Not the lowest incomes. Not the medical specialists or directors at the top. It is the accountant with ten years of experience, the marketing manager, the policy officer at the municipality. Occupations where the daily work largely overlaps with what ChatGPT, Copilot and specialised AI tools can already do.
Education vs AI exposure
HBO and university graduates score significantly higher than vocational graduates. Knowledge work is more vulnerable than physical work.
Salary vs AI exposure
The €50–75K group is the most vulnerable. Not the lowest wages, but the knowledge worker on a median-plus salary.
Largest absolute impact
Percentages are useful for comparison. But they abstract from what is actually happening. A decline of 34% in an occupational group of 245,000 people means that an estimated 83,000 people need to find different work. That is more than the entire working population of a city like Delft. With families, mortgages and careers that need to be rebuilt.
Below are the occupations where absolute job loss is greatest. These are not necessarily the occupations with the highest percentage decline, but the occupations where the societal impact weighs heaviest. Administrative employees are at the top by a wide margin. If there is anywhere where the conversation about guidance and retraining is urgent, it is there.
Ageing: soft or hard landing?
Not every job that disappears has to cost someone their position. If an occupational group is heavily aged, employees will retire in the coming years. Their role is then not replaced rather than someone being made redundant. That makes the transition "soft."
CBS data shows that the business and administrative sector, where the decline is greatest (−153,000 jobs), is also heavily aged: 24.6% are aged 55 or over. That is 457,000 people who will retire in the next 10 years. The AI transition therefore largely overlaps with natural attrition. In ICT it is the opposite: only 17.4% are aged 55 or over. A young profession. But demand is growing there, so ageing is not a factor.
Source: CBS Statline table 85276NED, employed workforce 2024. "Soft landing" = high ageing + high AI score (transition via retirement). "Hard landing" = low ageing + high AI score (retraining needed).
The administrative sector is estimated to lose 153,000 jobs towards 2030. But 457,000 employees in that sector are aged 55 or over. Natural attrition is greater than the AI-driven decline. That makes the transition manageable, provided we start now by not filling departing roles rather than waiting until automation is forced. In creative and language work (23.6% aged 55+, AI 7.2) the same applies. In ICT (17.4% aged 55+) it does not, but demand is growing there.
Exposure × Vulnerability
This chart places the two dimensions of this analysis side by side. Horizontal: the technical AI exposure. What can AI theoretically take over? Vertical: the NL vulnerability. What will actually happen in the Dutch context, given our regulation, labour market tightness and societal preferences?
The most interesting quadrant is bottom right: high exposure, low vulnerability. That is where occupations like software developers and IT project managers sit. AI can technically take over much of their work, but demand is so great that replacement does not materialise. Top left is the reverse: low exposure but rising vulnerability. That is where care coordinators and teaching assistants sit: AI is affecting their administrative tasks while their core task (human contact) remains safe for now.
The size of each bubble represents the number of jobs. The larger, the more people affected. Click a bubble for the full analysis of that occupation.
High exposure + high vulnerability. Admin workers, copywriters, customer service.
High exposure but NL context protects. ICT: demand grows faster than AI replaces.
Low exposure but vulnerability is rising. AI in admin, not in care work.
Low on both axes. Construction, installation, physical care.
What the data reveals (and what it doesn’t)
Correlations tell you something different from what you might expect. AI exposure predicts almost nothing about historical growth (r = −0.06). Occupations most exposed to AI have not systematically declined. Four of the ten fastest growers score AI 7 or above. The paradox is not the exception. It is the pattern.
The r = 0.87 between JPE and Felten AIOE confirms that our LLM-estimated scores align with the academic reference framework. The 13% difference sits exactly where you’d expect: teachers, nurses, judges — occupations where Felten says “AI can do it” but JPE says “it isn’t happening.” That is the core distinction of this analysis.
Part-time and AI vulnerability
The Netherlands is Europe’s part-time champion. In this dataset 43% work part-time. That raises the question: is there a correlation between part-time work and AI vulnerability?
That correlation exists, but is not linear. Occupations with the highest part-time percentage (nurses 72%, primary school teachers 72%) have low AI scores. The work is physical and relational. Occupations with the lowest part-time percentage (fire brigade 15%, welders 8%, lorry drivers 12%) also score low. That work is likewise physical.
The vulnerable middle group has a part-time percentage of 30 to 45%. Administrative workers (58% part-time, AI 8/10), secretaries (65%, AI 7/10), marketers (45%, AI 8/10). These are occupations where the work largely takes place behind a screen. And that is precisely the domain where AI is advancing fastest.
17 occupations. Avg. AI: 4.2
17 occupations. Avg. AI: 2.9
43% part-time. Highest in Europe.
Occupations with a high part-time percentage (healthcare, education) have an average lower AI exposure. Occupations with low part-time (construction, transport) are physical and likewise less vulnerable. The most AI-vulnerable occupations sit in the middle: administrative, commercial and ICT.
Most part-time
Least part-time
Embed and share
Free to embed under CC BY 4.0.
<iframe src="https://simondjanssen.nl/en/ai-map" width="100%" height="800" frameborder="0" loading="lazy" title="AI Exposure Map"></iframe>
<iframe src="https://simondjanssen.nl/en/career-scan" width="100%" height="900" frameborder="0" loading="lazy" title="AI Career Scan"></iframe>
https://simondjanssen.nl/en/ai-map
What now? These are not predictions. They are directions. The question is not whether the percentages are exactly right. The question is: do you recognise the pattern? And what will you do with it? The career scan provides a personal report with forecast, reskilling advice and concrete tools. The full methodology and dataset are openly available under CC BY 4.0.