How were these figures calculated?

The map covers 103 occupational groups based on the CBS Occupational Classification ROA-CBS 2014 (BRC 2014, 2025 edition). Together these occupations represent approximately 6.8 million jobs, roughly 69% of the Dutch working population (CBS: 9.8 million in 2025).

Labour market figures (jobs, growth, historical index, part-time share) come directly from CBS table 86320NED. 53 occupations match a CBS occupational group one-to-one, 37 share a CBS group with other occupations (trend from CBS, job count is an estimated share of the group total) and 9 are a narrow part of a larger group (trend from CBS, job count is our own estimate); for 4 the job count comes from a public register (e.g. judiciary, UWV, police), with the trend of the CBS group.

The remainder consists primarily of niche specialisations, undefined occupations and occupational groups that fall outside the CBS BRC 2014 top-level categories. The dataset does not claim to cover the full labour market.

The AI exposure score indicates what proportion of the core tasks of an occupation is technically automatable by current AI technology (LLMs, computer vision, robotics, agentic AI). It is a measure of technical exposure, not of actual displacement.

The vulnerability score indicates to what extent technical AI exposure also translates into real impact in the Dutch context. High exposure does not necessarily lead to high vulnerability when protective factors are present.

Factors taken into account:

• EU AI Act Art. 14: Requires human-in-the-loop for government decisions (lowers vulnerability of public sector roles)
• DBA legislation: Self-employment regulations influence the pace at which roles can be replaced
• Healthcare sector regulation: BIG registration and statutory requirements protect healthcare occupations
• Demand elasticity: When demand grows faster than AI displaces (e.g. ICT), vulnerability decreases
• Social preference: Societal preference for human contact (healthcare, education) slows adoption
• Labour market tightness: Acute shortages (healthcare, construction, installation) make displacement unlikely

Since v3 (September 2026) the vulnerability score no longer counts in the forecast. It correlates 0.96 with the JPE score and therefore added almost nothing. The score remains visible as context.

The model answers the question: how many people will work in this occupation in 2030, if expected labour demand materialises and AI and robotics do their work? The answer has three layers, also shown as separate scenarios on every occupation page.

Since September 2026 (v3) the model no longer starts from the trend of the last two years. A backtest on CBS data since 2013 showed that such a short trend does not persist: the following years tend to move the other way. The model now starts from an independent employment forecast per occupational group (ROA, Maastricht University), with AI and robotics added separately.

The AI Exposure Map uses two scoring systems side by side. The primary score is the Janssen Practical Exposure (JPE), developed for this map. Additionally, we show the academic Felten AIOE as a reference.

• AI exposure scores are model estimates, not empirical measurements. They have been cross-validated with academic indices but have not been independently audited.
• The dataset covers approximately 69% of the working population. Niche occupations are not included.
• Historical indices come from CBS, but CBS reports a series break between 2020 and 2021 (new survey design). At occupation level this can cause jumps, e.g. for general directors. Comparisons from 2021 onwards are the most reliable.
• Where occupations share a CBS group or are a narrow part of one, the trend follows the whole CBS group. A sub-occupation may in reality move differently.
• The Labour Force Survey is a sample. Small groups (below ~30,000 workers) fluctuate strongly from year to year.
• The 2030 forecast is a directional analysis, not a labour market forecast. The model does not account for economic shocks, policy changes or breakthrough technologies.
• Not all contraction is AI-related. Agriculture and fisheries are declining due to policy (nitrogen regulations, quotas), not automation.
• Salaries are medians; the spread within occupational groups can be substantial.
• The 2030 forecast is presented as a band, rounded to 5 percentage points, with a midpoint. The band combines the uncertainty of the ROA forecast with a favourable and unfavourable AI and robotics layer. Point estimates would imply false precision.
• The rule for structural decliners is based on the same backtest used to test it afterwards (ROA underestimated decline for 17 occupations). That is not an independent test.
• The robotics estimates were made relative to the historical CBS trend. Where automation has been under way for years (retail, warehousing), part of it may already be in the ROA base: a possible small double count.
• Some occupations overlap with a CBS group that is already split over other occupations (register counts, narrow estimates). Together about 78,000 jobs; coverage has been corrected for this.
• Demographic ageing is not included in the forecast model. Age distribution per occupational group is available in CBS 86320NED, but the model does not use it yet. Ageing data is shown as a contextual indicator in the career scan and the insights page. In heavily ageing sectors (e.g. administrative: 24.6% aged 55+), natural attrition may soften the AI transition. This is an important nuance that the model itself does not quantify.

The AI exposure scores and forecasts have been compared with five independent sources:

• Felten, Raj & Seamans (2023) — Occupational Heterogeneity in Exposure to Generative AI. Our top scores (translators, writers, developers) align with their highest AIOE values. Our lowest scores (construction, frontline healthcare) match their lowest.
• Georgieff & Hyee (2021) — Artificial intelligence and employment: New cross-country evidence. OECD Social, Employment and Migration Working Papers, No. 265. Confirms the pattern: knowledge work high, physical work low.
• Frey & Osborne (2017) — The Future of Employment. Nurses scored 0.009 automation probability, consistent with our 2/10.
• WEF Future of Jobs Report (2025) — Confirms that software developers are among the fastest-growing specialist roles. +6 million developers projected globally by 2030. Growth direction consistent with our data.
• UWV Labour Market Forecast (2025) — Signals that labour market tightness for ICT occupations is easing. Less tightness is not the same as fewer jobs, but the growth rate is moderating. Used as a check on the direction of the forecast.

License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Cite as: Janssen, S.D. (2026). Dutch AI Exposure Map: 103 occupational groups analysed for AI impact. simondjanssen.nl/en/ai-map. Retrieved [date].

Methodological inspiration: Karpathy/jobs (MIT License)

This section describes what this research can and cannot tell you. Transparency about methodology is not a weakness — it is what separates research from marketing.

The Janssen Practical Exposure scores are estimated using large language models (Claude, Anthropic). The prompt asks the model to assess what percentage of an occupation's tasks can be assisted or replaced by current AI capabilities, on a scale of 1-10. Scores were generated in March-April 2026.

They are cross-validated against the Felten AIOE academic index for US occupations (r = 0.87) but no equivalent academic benchmark exists for European occupations. Cross-country proxy validation against Felten AIOE for equivalent ISCO codes shows JPE scores are systematically 1.7 points lower than Felten across all three non-US countries (UK: −1.72, DE: −1.67, FR: −1.72, based on 56, 66, and 48 ISCO matches respectively). This gap is consistent with the known methodological difference: Felten measures theoretical capability while JPE measures practical disruption. The uniformity of the gap across countries suggests consistent scoring, not country-specific calibration error. The scores reflect a single point-in-time assessment and may not be reproducible with different models or prompts.

Since September 2026 the Dutch forecast uses JPE forecast v3 (ROA base + AI layer + robotics layer, see above).The US, the UK, Germany and France use the same JPE forecast v3 since 27 September 2026, each with its national projection as the base (see above). The earlier method described here was used for the other countries and has been withdrawn after the audit (see above); every country now uses JPE forecast v3. It used country-specific dampening factors and AI pressure tables. These parameters are informed by labour market research but are ultimately judgment calls, not empirically measured values. A formal sensitivity analysis is available (see below).

The US forecast uses a fundamentally different projection method (direct BLS growth percentages, factor 1.0) compared to other countries (index-based, factor 2.5). Results are presented on the same scale but are not directly mathematically comparable. The direction of findings is consistent; the magnitude may differ.

A sensitivity analysis (available in full at /docs/sensitivity-analysis.md) shows:

• For the Netherlands, see the sensitivity analysis in the forecast model section above (JPE forecast v3). For the UK, the number of shrinking occupations (below −3%) is 13, 14 and 31 at an AI slope of −0.10, −0.25 and −0.50 percentage points per year.
• The earlier sensitivity analysis of the other countries (dampening, AI pressure and the resulting country ranking) described the withdrawn model. Because that model rested on synthetic or estimated history, we no longer present its country ranking as a finding. The sensitivity of the rebuilt forecasts is reported per country on the methodology page of each map.

The Netherlands, the US, Germany and the UK report gross annual salary. French figures are DARES net medians for employees (net monthly salary × 12, not converted to gross), so they are lower than a gross figure and exclude the self-employed. Cross-country salary comparison should account for purchasing power parity, tax regimes, and social benefits, which this dataset does not adjust for.

The datasets cover major occupational groups but not all workers in each country. Estimated coverage: NL ~69%, US ~54%, DE ~64%, UK ~53%, FR ~69%. Gig economy workers, military personnel, and some creative/freelance occupations are underrepresented or absent.

Five different national classification systems are referenced (CBS BRC 2014, BLS SOC 2018, KldB 2010, UK SOC 2020, PCS 2020), cross-walked via ISCO-08. The audit found that not all codes are valid: a few US codes are still SOC 2010, the UK list mixed SOC 2010 and SOC 2020 codes, many German KldB codes are invalid or point to a different occupation, and the French codes are not valid PCS 2020 codes. These are being corrected in the rebuild; for the UK this is done, and each occupation is now linked to one or more SOC 2020 unit groups; for the US, where each occupation is linked to SOC 2018 codes; for Germany, where each occupation is linked to valid KldB 2010 codes; for France too, where each occupation is now linked to one or more FAP 2021 families (DARES). Occupation definitions do not map perfectly across systems. A French “ingénieur informatique” may cover a broader scope than a UK “programmer.”