Sources & Methodology

Methodology

The reskilling gap analysis synthesises 6 research prompts executed across Claude Deep Research, Gemini Deep Research, and Perplexity, plus 20+ primary source PDFs. The core methodology is task-based displacement estimation: AI does not replace jobs in their entirety but substitutes specific tasks within occupations. When >40% of constituent tasks are automated, the occupation requires formal reskilling.

Exposure coefficients are triangulated from three independent indices: the ILO Global Index of Occupational Exposure (2025), the Anthropic Economic Index (2025, observed Claude usage), and Microsoft’s “Working with AI” applicability framework. These are applied to Eurostat employment data (lfsa_egai2d) to produce displacement estimates.

Transition feasibility integrates skills distance (ESCO Skill-Occupation Matrix), certification barriers (national qualification frameworks), wage differentials (Eurostat SES + national statistics), and historical displacement evidence (Gathmann & Schönberg 2010, Jacobson/LaLonde/Sullivan 1993, Bertheau et al. 2022).

Key Limitations

Nine places where the headline numbers on this site rest on inference rather than direct measurement. Each is load-bearing; none is hidden.

1. Displacement estimates are modelled, not measured

   No one knows exactly how many workers AI will displace by 2035. The 38.72M figure is a triangulated estimate with ±8% confidence intervals. Country-level allocations use proportional distribution, not bottom-up measurement.

2. Retirement offset is derived, not observed

   The 8.67M 2035 figure starts from Eurostat lfsa_egai2d: ~22.4% of workers in the high-exposure ISCO groups are currently aged 55–64. Assuming a weighted average effective retirement age of ~63 across the EU-27 (with a 1-year rise by 2035 for statutory-age reforms already legislated in DE/FR/NL), the 55–64 cohort fully exits the labour market by 2035. A proportional share of the 45–54 cohort (~3% of high-exposure workers) crosses the retirement threshold within the 10-year window. The resulting 8.67M is roughly 22% of the gross exposure — consistent with but not directly measured. Legislation changes (further statutory-age increases, incentive schemes for working past pension age) could move this by ±1.5M.

3. The 7.55M / 15M / 7.5M split is a heuristic partition

   The split of the 30.05M net need applies a ~25 / 50 / 25 partition based on observed task-substitution depth in the Anthropic Economic Index and Microsoft “Working with AI” applicability: the top task-coverage quartile (>70% of tasks AI-capable) proxies “deep reskilling”, the middle quartiles proxy substantial upskilling, the lowest high-exposure quartile proxies partial task change. The split is an ordering heuristic, not a measured transition rate.

4. Channel throughput aggregates five data sources with different scopes

   The 3.34M/yr throughput figure combines five sources with different population scopes. University adult returners: Eurostat educ_uoe_enra02 (students 30+, completion-weighted). VET/apprenticeships: Cedefop Key Indicators on VET (CVET completers), cross-referenced to Cedefop Skills Forecast 2025 Employment_occupation_detail_qualification.xlsx for VET completion rates. Corporate L&D: Technavio Europe Corporate Training Market sized to training-hours / 137h meaningful-reskilling benchmark (Fosway 2025). ALMP: OECD SOCX (Social Expenditure Database) ALMP training-category spend as % GDP, combined with national LMP trackers (BA Arbeitsmarktberichte for DE; DARES Emploi for FR). The Eurostat empl_lmp_expsumm dissemination table has been withdrawn (returns 404 as of April 2026); OECD SOCX is the replacement primary. Bootcamps: Career Karma 2024 plus national tech-training registries (DE: Bildungsgutschein-aligned, FR: CPF-compatible). Each channel’s internal derivation is itself approximate; the total should be read as order-of-magnitude.

5. Net new capacity is an inference, not a measurement

   The ~450K/yr figure is 3.34M total throughput minus baseline churn, where baseline churn is estimated at ~2.9M/yr from Eurostat job-to-job transition flows plus green-transition reskilling obligations (EU Green Deal Social Climate Fund). The gap could plausibly be 250K–700K depending on how strictly “additional capacity for AI transition” is defined.

6. Transition feasibility is scenario-based

   The flow distribution (15–25% to A+, 5–10% to C, 40–60% not transitioning) synthesises evidence from multiple displacement episodes but is not a prediction. Actual outcomes depend on policy intervention, economic conditions, and individual choices.

7. System comparison uses mixed indicators

   Adult learning participation rates (Eurostat AES) measure breadth, not depth. ALMP spending measures input, not outcome. Cross-zone transition rates are estimates derived from occupational mobility data, not direct measurements.

8. Currency conversion preserves US-specific institutional detail

   Wage and earnings data from US-sourced research (Massenkoff & McCrory 2026, Jacobson/LaLonde/Sullivan 1993) has been converted to euros at the prevailing EUR/USD rate at time of publication. US-specific institutional details (programme names, legislation) are preserved for traceability to the original sources.

9. US research applied to European context

   The entry-level hiring slowdown (Massenkoff & McCrory 2026) and occupation-level observed exposure data are derived from US Current Population Survey and Claude usage data. European adoption patterns may differ due to regulatory environment (EU AI Act), institutional frameworks (works councils, co-determination), and different occupational classification (ISCO-08 vs SOC). The data is included as the best available leading indicator, not as a direct measurement of European conditions.

Source Index

Tier 1 Academic & Institutional

Tier 2 Policy & Practitioner

Tier 3 Market & Journalistic

Derivation Appendix

The eight derived numbers below are upgraded from “described” to “grounded”: each is traced to a concrete dataset and reproduced by a script in scripts/. Every script prints its computation, writes a CSV to scripts/output/, and falls back to a cached pull if the Eurostat API is unavailable. Sections A–D were grounded in session S1 (retirement offset, task-coverage split, channel throughput, net new capacity). Sections E–H were grounded in session S2 (skills distance, speed gap, system radar scores, A→C rates).

A. Retirement offset, 2035 — 8.67M

Pool: 38.72M workers in high-exposure ISCO groups (weighted mix of ISCO-08 major groups 2, 3, 4, and parts of 5). Share aged 55–64 today: 20–26% by country, 21.0% EU-27 weighted average (Eurostat lfsa_egais, 2024). The 2035 offset equals this cohort, minus a ~5% residual who work past statutory age (Eurostat lfsi_esgan), plus a ~10% fraction of the 45–54 cohort that crosses the 2035 statutory threshold (DE, UK, ES, IE, CZ, BE all have legislated increases). The two second-order effects approximately cancel.

EU-27 central 8.15M is 6% below the Layer 5 headline of 8.67M; the headline sits inside the [7.74M, 9.45M] sensitivity bracket. Script: scripts/01_retirement_offset.py. Inputs cached at scripts/data/lfsa_egai2d_2024.json and scripts/data/statutory_retirement_2025_2035.csv. Cross-check against the Cedefop Skills Forecast 2025 Replacement_demand_occupation series (available locally at /Users/philippmaul/Documents/projects/european-ai-exposure-map/data/cedefop/skills-forecast-2025/Replacement_demand_occupation.xlsx) recommended for a future confirmation pass; the current 8.15M derivation is within the stated ±8% CI without it.

B. Task-coverage split — 7.55M / 15M / 7.5M

The 30.05M net-need pool is divided into deep reskilling, upskilling, and partial change. Two independent task-coverage indexes produce visibly different splits.

Anthropic Economic Index (Handa et al. 2025, Figure 4): a cumulative distribution of O*NET occupations by the fraction of tasks where Claude usage is non-trivial. 4% of occupations have AI usage on ≥75% of tasks, 11% on ≥50%, 36% on ≥25%. Within the ≥25% pool (operationally the “high-exposure” subset), the distribution is 11 / 19 / 69 top / middle / bottom.

Microsoft Working with AI (Tomlinson et al. 2025, Table 1): employment-weighted applicability scores for 97 SOC minor groups. Score range 0.03–0.38, roughly uniform. Employment-weighted top quartile captures ~23% by employment, middle two quartiles ~54%, bottom ~23% — close to the 25 / 50 / 25 Layer 5 partition.

Script: scripts/02_task_coverage_split.py.

Sensitivity — observed usage vs capability ceiling

The headline split uses Microsoft’s capability-rubric framework — what tasks AI could cover if deployment were uniform across occupations. The Anthropic Economic Index measures what Claude is observed to be used for today, which is more bottom-heavy (shallow-task-change dominates over full-role substitution). The two indices measure different constructs, not competing estimates of the same thing. For a 2035-horizon reskilling plan, capability is the right anchor; for 2026 urgency, observed usage is.

If deployment stalls and the capability gap does not close, the Anthropic reading becomes the more accurate long-run picture: deep-reskilling need drops by ~56% and partial-task-change nearly triples.

C. Annual channel throughput — 3.34M

Five channels, each with its own primary source. Low / central / high sensitivity brackets reflect the differences between Eurostat administrative counts and survey-based estimates for the corporate and bootcamp channels.

Central total matches the Layer 5 headline. University and ALMP channels are tightly bounded by administrative counts. Corporate L&D has the widest uncertainty; the 137h Fosway benchmark versus 40h average is the dominant parameter. Script: scripts/03_channel_throughput.py.

D. Net new capacity — 450K

Total channel throughput (3.34M) minus baseline absorption. Baseline = training-requiring subset of annual job-to-job flow (Eurostat lfsa_etpgan, 2024) plus green-transition and demographic replacement commitments.

EU-27 annual job-to-job flow ~ 20.1M (10.2% of employment). IAB occupational-mobility evidence (Rhein et al. 2013) puts ~16% of these transitions in the “training-required” category: ~3.2M. Add EU Green Deal Social Climate Fund commitments (~250K/yr) and Cedefop replacement demand in shortage occupations (~180K/yr). The critical free parameter is the share of that training-required flow that runs through the five tracked channels (rather than employer-internal routes) — the headline 450K corresponds to ~78% absorption. Uncertainty bracket: [~250K, ~700K]. Script: scripts/04_net_new_capacity.py.

E. Skills distance 0–10 — 16 transition pairs

Each transition label on transitions.html is mapped to a pool of ISCO 4-digit codes. Every ESCO occupation whose iscoGroup matches any code in the pool is pulled into a cluster (stub rows with fewer than five bucketed skills are dropped). For each cluster, a weighted skill vector is built from the ESCO occupation-skill relations file (essential = 1.0, optional = 0.5), with each leaf skill aggregated to its Level-2 parent in the ESCO skill pillar hierarchy (156 buckets, down from ~14K raw skills — necessary because raw-skill resolution produces cosine 0 for almost every A→C pair). The final distance is 10 × (1 − median cosine) over the src × dst cross-product.

Script: scripts/05_skills_distance.py. Spearman rank correlation between ESCO-derived distances and transitions.html headlines: 0.821. The ESCO-based absolute distances run systematically higher than the human-anchored 0–10 scale — this is an artefact of cosine over sparse Level-2 skill vectors, where even same-zone occupations top out around 0.35. The rank-order is grounded; the absolute calibration shift is retained as a delta rather than overwriting transitions.html.

F. Speed gap — 5 occupations

Two independent clocks per occupation. Disruption years derive from a doubling-time model over Anthropic Economic Index observed exposure + Microsoft Working with AI capability ceiling: low bracket = years to cross the 40% material-displacement threshold (T_double = 1.5 yr); high bracket = low + institutional absorption lag (1 yr) + years to cross the 60% full-automation threshold capped by capability ceiling (T_double = 2.5 yr). Response years are a weighted blend of pathway durations: bootcamp (0.5–1.5 yr), CPF/Bildungsgutschein (0.5–1.0 yr), corporate L&D meaningful reskilling (3–5 yr, Fosway 2025 137h benchmark), CVET ordinance update (3–5 yr, BIBB Neuordnung cycle), Umschulung (5–7 yr, SGB III §180 + integration lag).

Script: scripts/06_speed_gap.py. Because Anthropic Economic Index observed exposure is already in the 0.33–0.70 range across all five occupations, the derived disruption-low bracket compresses toward 1 year: AI capability to displace is already present; deployment speed is the binding constraint. The writers/translators response-years figure (1–2 yr) diverges materially from the 4–6 yr headline — the pathway mix (CPF-heavy, self-directed) gives a short duration, while the headline reflects the reality that freelance writers lack institutional reskilling scaffolding. Pathway-definition-dependent disagreement; retained for review.

G. System radar scores 1–10 — 5 dimensions

One observable indicator per dimension: Speed ← non-formal adult-learning participation rate (Eurostat trng_aes_100, 2022); Scale ← total FE+NFE participation rate 25–64; Quality ← share of enterprises providing CVT (Eurostat trng_cvt_01s, 2020 CVTS round); Equity ← ratio of low-education (ISCED 0–2) to high-education (ISCED 5–8) AES participation (Eurostat trng_aes_102, 2022); Funding ← ALMP training spend % GDP (OECD SOCX 2022 + national LMP trackers; Eurostat's open LMP table has been withdrawn from the dissemination API). Per-country percentile ranks are averaged within each system and mapped linearly to 1–10 (1 + 9 × avg_percentile).

Script: scripts/07_system_radar.py. Headline scores in parentheses. The Speed dimension is the weakest proxy — the Skills2Capabilities 2024 "VET responsiveness" index would anchor it directly but is not locally accessible. Non-formal participation is a reasonable stand-in but it undervalues Liberal Market speed (UK/IE rely on bootcamps outside the AES definition) and overvalues Germanic Dual speed (high NFE participation co-exists with the Beruf-system hysteresis). Two notable deltas: Germanic Quality 7.6 vs 10 (headline reflects VET depth that CVTS enterprise-share doesn't capture); Liberal Speed 5.5 vs 9 (market-agility isn't in the NFE metric). Retained for review.

Germanic-Dual sensitivity panel. Three Germanic-Dual cells shift materially depending on whether the system is scored by depth-of-system (expert judgment anchored on Ausbildungsordnung depth and Umschulung integration) or by breadth-of-participation (CVTS enterprise-share + AES non-formal participation + AES equity ratio). Quality moves 10.0 → 7.6, Equity 7.0 → 4.8, Speed 5.0 → 7.6. The pattern — depth undercounted when the indicator captures breadth — is the same divergence shape as the Microsoft-vs-Anthropic split flagged in the Sensitivity panel above for B. Task-coverage split (capability ceiling vs observed usage). Cross-referenced; no separate sensitivity table added in this session.

H. System A→C transition rates — 6 models

Reconstructed from two components: (i) 5-year re-employment rate post-displacement per country, sourced from Bertheau et al. 2022 (IZA DP 15033) for AT/DK/FR/IT/PT/ES/SE, with system-peer medians for the remainder; (ii) Zone-C destination share modelled linearly on ALMP training spend % GDP with slope calibrated so Nordic systems (ALMP ~0.37%) produce ~10% share and CEE/Southern (~0.09%) produce ~4% — consistent with Bertheau's finding that "a 10pp increase in the share of ALMP spending is associated with a 5% decrease in earnings losses" (Section 5). Per-country rate = re-employment × zone-C share; system rate = simple mean over members. Low/high brackets span the ALMP slope uncertainty envelope.

Script: scripts/08_a_to_c_rates.py. Four of six systems land within or adjacent to the headline bracket. Two material disagreements: Germanic Dual +3.4 (AT's 0.51% GDP ALMP spend alone predicts higher mobility, but headline reflects Beruf/Ausbildungsordnung hysteresis that locks occupational mobility — the ALMP-linear model can't see this); Liberal Market −3.3 (UK ALMP at 0.01% GDP predicts near-zero, but bootcamp-driven market response produces higher de-facto mobility than the model captures). Both are flagged for review rather than overwriting systems.html. Underlying PDF: scripts/data/bertheau/iza_dp15033.pdf.

Flagged deltas (all eight derivations)

Consolidated delta summary. None of the site numbers are overwritten in this session; deltas are flagged for human review.

Connection to Other Products

This analysis is Layer 5 of the 7-layer Nexalps European AI Labour Market suite. Each layer builds on the previous:

Layer	Product	Question Answered
1	AI Exposure Map	Which jobs does AI hit?
2	European Careers Map	Where is the job market going?
3	Historic Disruptions	What happened every other time?
4	Demographics & AI	Is Europe running out of workers?
5	Reskilling & AI (this site)	Can Europe reskill fast enough?
6	Synthesis (coming)	What does this mean for Europe?