On 30 April 2026, FUNCAS published a working paper authored by Prof. Francisco Rodríguez-Fernández (UGR / FUNCAS) estimating the destruction of up to 2.3 million jobs by AI in Spain over the next ten years. This page explains how that work relates to the «AI and Employment in Spain» dataset (502 occupations, 22.73 million workers), deposited on Zenodo in March 2026.
The FUNCAS study adapts the AI Occupational Exposure index by Felten et al. (2023) to the Spanish National Classification of Occupations via a double SOC→ISCO→CNO crosswalk, and applies a parametric destruction-and-complementarity model calibrated on Eloundou (2024), Acemoglu (2024) and the INE-ETICCE Q1 2025 survey. The output is absolute figures of expected job destruction over a ten-year horizon, aggregated to the 9 broad CNO-11 groups.
The «AI and Employment in Spain» dataset (v15) scores 502 occupations at 4 digits with a theoretical vulnerability value between 0 and 10, adversarially validated across seven generative-AI models, and reallocates employment and salary to the 4-digit level via an LFS-Census-SEPE-WSS cascade. The output is a 4-digit occupational layer with vulnerable wage bill, with no assumed time horizon or adoption speed.
Both questions are legitimate, both answers are useful, and together they offer a fuller picture than either on its own. This page documents where they converge, where they complement each other, and how to read them together.
«Since disaggregation below the nine broad CNO-11 groups is not available with sufficient precision in INE press releases, the figures of employed persons per group have been obtained from the quarterly publication of the LFS itself.» Rodríguez-Fernández, FUNCAS DT-2026/04, §3.1
The comparison operates at the level of the 9 broad CNO-11 groups (1 digit), where both methodologies produce directly comparable values. Differences in granularity, time frame and output are intentional: each model answers a different question.
When the v15 dataset is aggregated to the 1-digit CNO-11 level and compared with the AIOE-CNO values published by FUNCAS, the two methodologies rank the broad groups in essentially equivalent ways.
The convergence is informative precisely because the two calculation paths share no dependencies: AIOE is built on the O*NET matrix and the AAAI index; v15 is built via per-occupation scoring with adversarial multi-model validation, without reference to AIOE. Two independent models converging at the macro level is a weak but real form of mutual validation.
0.936Pearson coefficient · Spearman ρ = 0.830
Each point represents one of the broad CNO-11 groups. Labels: 0 Armed Forces · 1 Managers · 2 Scientific technicians · 3 Support technicians · 4 Clerical support · 5 Services · 6 Agriculture · 7 Manufacturing · 8 Plant operators · 9 Elementary. The line is a least-squares linear regression. Raw data available in CSV format.
FUNCAS assigns expected destruction to groups 1, 2, 3 and 4. The 4-digit decomposition in the v15 dataset shows how vulnerability ≥7 is distributed within each of those groups. The findings most relevant to the public conversation are the following.
Group 4 produces the most informative result of the decomposition. Out of 2,132,500 workers in the v15 dataset, 1,039,883 people — 48.8% of the group — score vulnerability ≥7, spread across just 9 specific occupations. The concentration is not uniform: a single occupation absorbs nearly half of all the group's vulnerable mass.
Here the two models diverge, and the divergence is the interesting reading. FUNCAS assigns destruction to the group by applying a flat AIOE of 0.52 across the 870,000 employed. The v15 dataset covers 33 managerial occupations at 4 digits, and none reaches the vulnerability ≥7 threshold. The managerial occupations with the highest scores (commercial, financial, HR, R&D directors) score between 5.5 and 6.5.
The coherent interpretation: Felten's AIOE captures task exposure to generative-AI capabilities, and a manager's tasks — strategic analysis, drafting reports, executive communication — are highly automatable. The v15 model distinguishes between technical task exposure and net occupational vulnerability, which includes non-delegable tasks of judgement, legal-corporate accountability, and interpersonal interaction. AI will act in this group as augmentation, not substitution. That reading is consistent with both models when they are read together.
The figure of 2.3 million jobs destroyed over ten years headlining the FUNCAS coverage and the figure of 2.75 million workers with vulnerability ≥7 produced by the v15 dataset are different metrics that sit at comparable orders of magnitude. The first is expected destruction under explicit time scenarios; the second is a theoretical exposure ceiling with no assumption about adoption speed.
The most useful reading for the public debate is probably the conjunction of both: in the short term, pay attention to the adoption pace that materialises the FUNCAS estimates; in the medium term, pay attention to the specific occupational composition of the vulnerable mass that the v15 dataset documents at 4 digits. One metric answers «how much» over a concrete horizon; the other answers «where» at the granularity of a public-policy decision or a reskilling programme.
The full methodological note with the raw cross-validation data, the reproducible script and the technical appendices is deposited on Zenodo under the same series as the original dataset. Any researcher, journalist or institution can recompute the results from the v15 JSON in a single execution.