What are you looking at?

On this interactive treemap you can see at a glance which of the 113 Dutch occupations are most exposed to AI — the larger the tile, the more jobs, and the redder the color, the higher the AI exposure.

The Dutch labor market is sharply divided: a third of all 8.2 million jobs and €101 billion in annual wages sit in occupations with low AI exposure, while nearly a quarter of jobs — particularly in administration and finance — face high exposure.

⚠️ High exposure does not mean your job will disappear. Anthropic's own research (March 2026) found no systematic increase in unemployment for highly exposed workers since ChatGPT launched — though there are early signs that hiring of workers aged 22–25 has slowed in exposed occupations. The full economic impact remains uncertain.

Color = level of AI exposure:

Green — low, AI has little impact
Orange — moderate
Red — high, AI could take over most tasks
Grey — no exposure score available

Click or tap a tile to learn more about that occupation.

Where does the data come from?

CBS — Statistics Netherlands
Employment figures per occupation (number of jobs, sector distribution).

ROA — Research Centre for Education and the Labour Market
Labour market forecasts, job outlook scores, and education requirements per occupation.

UWV — Dutch Employee Insurance Agency
Wage data and labour market tension (shortage vs. surplus per occupation).

AI exposure scores
Based on task-level analysis: how many tasks within each occupation can currently be performed or augmented by AI models.

Anthropic Economic Index — Massenkoff & McCrory (2026)
Introduces a distinction between theoretical AI exposure (what LLMs are capable of) and observed exposure (tasks actually being automated in professional use). Our scores are theoretical. Anthropic's research shows actual adoption is currently a fraction of theoretical capability — e.g. Computer & Math occupations: 33% observed vs. 94% theoretical coverage.

Microsoft Research — Tomlinson et al. (Dec 2025)
Analyzed 200k anonymized Microsoft Bing Copilot conversations to measure AI applicability across occupations. Independent US-based dataset reaching similar conclusions to the Anthropic research: AI is most applicable to information work (communicating, learning, writing). All aggregated data published openly at github.com/microsoft/working-with-ai.

Data Sources & Methodology

Total workers: 8,183,000

Source: CBS StatLine, table 85517NED — "Werknemers; uurloon en beroep", year 2024
What it measures: Werknemers — employees with a wage contract. Does NOT include self-employed (ZZP). CBS EBB puts total working population at ~9.4M, so this undercounts by ~1.2M.
Verify: cbsodata.get_data('85517NED'), filter Perioden == '2024', sum Werknemer_1 × 1000 for all 4-digit BRC codes.


Weighted average exposure: 5.2/10

Source: Our calculation — each occupation's exposure score × its worker count, divided by total workers.
Caveat: Scores are from Gemini Flash via OpenRouter, using ESCO descriptions. These are LLM judgments, not empirical measurements — treat as directional, not precise.
Verify: Open data.json, compute sum(exposure * jobs) / sum(jobs).


Tier breakdown (31% / 26% / 38%)

Source: Same as above. Direct count of workers in occupations scoring 0–3, 4–6, 7–10.
Caveat: Fully dependent on LLM scores being calibrated correctly.


€174B wages exposed

Source: CBS salary data (85517NED) × annual hours calculation.
Formula: median_hourly_wage × 1,872 hours (36 hrs/week × 52 weeks). 36 hours is the CBS average contracted hours.
What it measures: Median wage × headcount — not total wage bill. Underestimates high earners, overestimates low earners. Excludes ZZP.
Verify: CBS StatLine 85517NED, column k_50ePercentielMediaan_3.


Category exposure averages (ICT 8.6, Business & Finance 8.0, etc.)

Source: Same CBS worker counts + LLM scores, grouped by BRC 2014 sector code (2-digit prefix).
Caveat: ICT has only 4 occupation groups — small sample. Business & Finance has 13 groups covering a wide range of jobs.


Employment growth forecasts (the %/yr figures)

Source: ROA AIS tot 2030, DataverseNL doi:10.34894/DVQTOG, file AIStot2030_Arbeidsmarktinformatie_editie2025.csv
Variable: verwachte uitbreidingsvraag tot 2030, field gemjaarlijksperc
What it measures: Expected annual net employment growth 2025–2030. Does not include replacement demand (retirements).
Verify: Filter aggregatieniveau == 'beroepsgroep (BRC2014)' and onderwerp == 'verwachte uitbreidingsvraag tot 2030'.


Market tension (Tight / Average / Loose)

Source: ROA AIS tot 2030, same file
Variable: indicator huidige arbeidsmarktsituatie beroep (2024Q4), field typering
What it measures: Labor market tightness Q4 2024, ROA's composite of vacancies vs. job seekers. Dutch original: krap / gemiddeld / ruim.
Verify: Filter onderwerp == 'indicator huidige arbeidsmarktsituatie beroep (2024Q4)'.


Bottleneck (Almost none / Some / Large / Very large)

Source: ROA AIS tot 2030, same file
Variable: ITKB toekomstige knelpunten beroepsgroep in 2030, field typering
What it measures: ROA's forecast of recruitment difficulty by 2030. Dutch original: vrijwel geen / enige / groot / zeer groot.
Verify: Filter onderwerp == 'ITKB toekomstige knelpunten beroepsgroep in 2030'.


Education levels

Source: ROA AIS tot 2030, same file — opleidingsachtergrond * variables
Method: Took the education category with the highest share of workers per occupation group.
Caveat: This is the dominant education level among current workers, not a formal entry requirement.


Outlook description

Source: ROA AIS tot 2030, same file
Variable: verwachte arbeidsmarktontwikkeling beroep 2025-2030, field typering
What it measures: How labor market tightness is expected to evolve 2025–2030.


AI exposure scores (0–10 per occupation)

Source: Gemini Flash (via OpenRouter), scored using:
1. ESCO API — English occupation description + essential skills
2. CBS metadata — Dutch occupation scope description
Prompt: Each occupation's page sent with a fixed rubric anchored at known examples (roofer = 1, software developer = 9, data entry clerk = 10).
Caveat: Model judgments. Relative ordering between occupations is more reliable than absolute scores.


Theoretical vs. observed exposure

Our scores are theoretical — an LLM assessed each occupation's tasks and judged how many could be automated (following Eloundou et al. 2023). Anthropic's March 2026 paper introduces a complementary measure: observed exposure, based on actual Claude usage in professional contexts. Key finding: Computer & Math occupations are 94% theoretically feasible, but Claude currently covers only ~33% of tasks in practice. The gap is even larger for Office & Admin.
What this means for our scores: treat them as directional upper bounds on potential impact, not predictions of current automation. Relative rankings are more reliable than absolute scores.


Confidence summary

ClaimConfidenceMain risk
Worker counts (8.18M)HighZZP excluded, CBS 2024
Salary figuresMediumMedian hourly × fixed hours
Growth forecastsMediumROA macro model, Dec 2025
Market tensionHighROA/CBS composite, Q4 2024
Bottleneck 2030MediumModel forecast, not observed
AI exposure scoresLow–MediumLLM judgment, not empirical
€174B wages exposedMediumMedian × headcount approx.

Automation vs. augmentation

Microsoft Research (Tomlinson et al. 2025) identifies two distinct patterns in how AI affects occupations, based on 200k real Copilot conversations:

  • Delegation — workers hand off tasks to AI, which performs them autonomously. Most common in Business & Finance and Media occupations.
  • Augmentation — workers use AI as a tool while retaining control of the task. Most common in Computer & Math and physical occupations.

How this connects to the Dutch data: Our Dutch Business & Finance category averages 8.0/10 exposure and ICT 8.6/10 — the highest in the dataset. MS research suggests these two sectors likely follow different paths: Finance workers may increasingly delegate tasks like drafting, summarising, and data entry to AI; ICT workers are more likely to use AI as a productivity aid while their core judgment work stays human. Both show high theoretical exposure, but the real-world impact looks different.

A second relevant finding: MS found only a weak correlation between AI applicability and wages (r=0.13). In the Netherlands too, the most AI-exposed occupations include well-paid sectors like ICT, finance, and legal services — not just low-wage work. The economic impact of AI will be felt across the income spectrum.

About

🙋🏽‍♂️ This project was created by Daniel Siahaya, Founder of HeadFWD, an AI transformation partner for enterprise.

✅ The goal is to make labour market data on AI exposure accessible and understandable, not to alarm, but to inform.

📥 Have feedback, data corrections, or ideas? Reach out on LinkedIn.

👊🏽 Credits of the idea of this app goes to: Andrej Karpathy.