Methodology
AI Job Index: methodology (v0.2 draft, 31 Oct 2026)
What the index shows
For 242 occupations, grouped in 16 function families, the index shows how much of the work is technically exposed to generative AI, which tasks AI tends to take over or support, and how much of the work depends on empathy and on creativity or strategy. Exposure means technical potential. It is not a forecast of job loss, and it says nothing about when or whether a company adopts AI.
Inspiration: Bloomberg's 2017 graphic "Find Out If Your Job Will Be Automated" (linked as inspiration only; no Bloomberg data or graphics are used). The quadrant idea comes from Kai-Fu Lee, AI Superpowers (2018); the axes and scores here are our own, built from O*NET.
The two views
- Bubble view: x = technical exposure (ILO by default; GPTs are GPTs as alternatives), y = observed AI usage (Anthropic Economic Index), size = Belgian employment of the ISCO 2-digit group. The distance between the two axes shows where AI is possible but not (yet) used. 13 occupations without observed usage data are shown apart, not at zero. The wage band is shown per occupation in the detail view, not on an axis.
- Quadrant view: x = creativity and strategy, y = empathy (EQ), split at the medians.
Units of analysis
- Occupation list: 242 O*NET-SOC 31.0 occupations chosen for business relevance (src/selection.py), each assigned to one function family.
- Each occupation is linked to one ESCO occupation for NL/FR/EN labels and to its ISCO-08 4-digit group (taken from the ESCO code prefix, e.g. ESCO 1221.3.2 gives ISCO 1221). Links come from the official ESCO-O*NET crosswalk (best match type, then title similarity) and, for 40 occupations where the crosswalk pick was poor, a manual pick from ESCO search (src/esco_overrides.json, marked "manual (ESCO search)").
Fields and how they are computed
| Field | Source | Computation |
|---|---|---|
| exposure (primary, 0 to 100) | ILO Working Paper 140 (Gmyrek et al. 2025) | ILO mean task exposure score for the ISCO-08 4-digit group x 100. All O*NET occupations that map to the same ISCO group share this value. |
| ilo_gradient | same | ILO exposure gradient (Not Exposed, Minimal Exposure, Exposed: Gradient 1 to 4). |
| exposure_gpts_human, exposure_gpts_gpt4 | Eloundou et al., GPTs are GPTs (Science, 2024) | Mean over the occupation's O*NET tasks of beta = E1 + 0.5 x E2 (human annotators and GPT-4 respectively), x 100. |
| exposure_aei_observed | Anthropic Economic Index, labor_market_impacts job_exposure.csv | Observed Claude usage share for the 6-digit SOC, x 100. This is observed use, not potential. |
| exposure_low / exposure_high | derived | Min and max of ILO, GPTs human and GPTs GPT-4 for the occupation. Used for ranges in the P&L view. |
| task score (0, 25, 50, 75, 100) | GPTs are GPTs | Mean of the human and GPT-4 beta values per task (E1 = 1, E2 = 0.5, E0 = 0), x 100. |
| task tag | GPTs are GPTs + AEI 2026-06-26 | 1) score 0 gives "human". 2) If AEI task data exists (latest of the April and May 2026 periods, global Claude.ai): "automate" when the automation share exceeds the augmentation share, otherwise "augment". 3) Without AEI data: "automate" when GPT-4 automation rating is T3 or T4, otherwise "augment". 4) No source: "unrated". |
| top_tasks | O*NET 31.0 Task Ratings | The 6 Core tasks with the highest O*NET importance. |
| task_mix | derived | Share of automate / augment / human / unrated across all the occupation's tasks, weighted by O*NET task importance. |
| eq_score (0 to 100) | O*NET 31.0 | Mean importance of Social Perceptiveness, Service Orientation (Skills), Establishing and Maintaining Interpersonal Relationships, Assisting and Caring for Others (Work Activities); importance 1 to 5 rescaled with (v - 1) / 4 x 100. |
| creativity_score (0 to 100) | O*NET 31.0 | Same method for Fluency of Ideas, Originality (Abilities), Thinking Creatively, Developing Objectives and Strategies (Work Activities). |
| zone | derived | Split at the median of each score over all 911 rated O*NET occupations (both medians are 52.19; this was checked and is a coincidence of the discrete scale). Low creativity + low EQ: AI only. Low creativity + high EQ: AI inside Human. High creativity + low EQ: mix AI / Human+AI. High on both: Human+AI. |
| wage | Eurostat earn_ses22_28 | Mean annual gross earnings in EUR, Belgium, 2022, enterprises with 10+ employees, for the ISCO-08 major group (1-digit). Finer levels are not published for Belgium in this table. |
| employment | Eurostat lfsa_egai2d | Employed persons aged 15 to 74, Belgium, 2024, for the ISCO-08 sub-major group (2-digit). Shared by all occupations in that group, so bubble size shows the size of the occupational group, not of the single role. |
ONET Work Styles are not used: in ONET 31.0 they are rated on a different scale and partly by AI, so they are not comparable with the other descriptors.
Coverage (242 occupations)
- ESCO link, NL/FR labels, ISCO, wage, employment: 242 of 242.
- ILO exposure: 241 (Team Assemblers, ISCO 3139, has no ILO score).
- GPTs exposure: 242. AEI observed exposure: 229.
- EQ, creativity and zone: 241 (Financial and Investment Analysts, 13-2051.00, has no O*NET 31.0 ratings).
Known limits
- US task descriptions (O*NET) are applied to Belgian jobs through ESCO/ISCO. Job content differs between countries.
- Several O*NET occupations share one ISCO group, so they share ILO exposure, wage band and employment.
- Wage is a broad band (9 values in total). It places bubbles in tiers rather than giving a role salary.
- AEI reflects use of one AI product (Claude) and the people who use it; it is a signal of current use, not a market measure.
- Task texts in NL (je-form) and FR (vous) are machine translations (Argos Translate, offline OPUS-MT models) of the second-person version of each O*NET task, followed by a rule-based and manual review. Two independent passes were run and the cleaner one kept per task (they agree on about 95% of tasks); 88 tasks were corrected by hand. A random check of 30 unflagged tasks before hand-correction found noticeable errors in about a quarter of the NL and a fifth of the FR texts, mostly clumsy wording and some wrong terms. They are not professional translations; a native-speaker read is advised before publication.
- One label is overridden by hand: FR title for HR Specialists (13-1071.00) is "chargé de recrutement / spécialiste RH" instead of ESCO's "directeur des ressources humaines".
- Sector (NACE) is not available per occupation; the filter uses function family and ISCO major group.
- ESCO labels are taken as published; a few read more senior or more specific than the O*NET role (see spot_checks.md).
Reproduce
Run, from src/: selection.py (list), fetch_esco.py after pick_esco.py (ESCO links and labels, cached in raw/esco_cache.json), then build.py, translate_tasks.py and mt_second_pass.py (two independent machine-translation passes), review_translations.py (rules, flags, manual fixes from task_translation_fixes.json) and enrich.py. Output: out/ai_job_index.json and out/ai_job_index.csv.
Licence status per source
| Source | Version used | Licence | Required credit / conditions | Status |
|---|---|---|---|---|
| O*NET Database | 31.0 | CC BY 4.0 | "This site includes information from the ONET 31.0 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA). Used under the CC BY 4.0 license. ONET® is a trademark of USDOL/ETA. NUVAI has modified all or some of this information. USDOL/ETA has not approved, endorsed, or tested these modifications." | OK with this credit |
| GPTs are GPTs (Eloundou, Manning, Mishkin, Rock) | GitHub repo openai/GPTs-are-GPTs, full_labelset.tsv | MIT (Copyright 2024 OpenAI) | Keep copyright notice with redistributed data; cite the Science 2024 paper | OK |
| Anthropic Economic Index | Release 2026-06-26 (task data) and labor_market_impacts job_exposure.csv (Hugging Face Anthropic/EconomicIndex) | Data CC BY; code MIT | Cite "Anthropic Economic Index report: Cadences" (Massenkoff et al., 2026-06-26) and the dataset | OK; licence wording on the HF card says "CC-BY" without version, so we write "CC BY" |
| ILO Working Paper 140 (Gmyrek et al. 2025) | Final_Scores_ISCO08_Gmyrek_et_al_2025.xlsx | CC BY 4.0 | Cite: Gmyrek, P., Berg, J., et al. 2025. Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper 140. © ILO. https://doi.org/10.54394/HETP0387. Because we adapt it, add: "This is an adaptation of a copyrighted work of the International Labour Organization (ILO). This adaptation has not been prepared, reviewed or endorsed by the ILO and should not be considered an official ILO adaptation. The ILO disclaims all responsibility for its content and accuracy. Responsibility rests solely with the author(s) of the adaptation." Do not use the ILO logo. | OK with citation and disclaimer |
| ESCO (European Commission) | ESCO API, default release on 31 Oct 2026 (one record via v1.1.1) | Free reuse under Commission Decision 2011/833/EU | Credit: "This service uses the ESCO classification of the European Commission." | OK |
| Eurostat | earn_ses22_28, lfsa_egai2d (downloaded 31 Oct 2026) | Free reuse with acknowledgement (Eurostat copyright notice, Decision 2011/833/EU) | "Source: Eurostat" plus dataset codes | OK |
| ESCO-O*NET crosswalk | ESCO publication | as ESCO | as ESCO | OK |
| Bloomberg 2017 graphic | not used | n/a | Link only, as inspiration | No data or design reused |
| Kai-Fu Lee quadrant (AI Superpowers, 2018) | idea only | n/a | "Quadrant framing inspired by Kai-Fu Lee, AI Superpowers (2018)." | Confirmed by Renout |
| Statbel 4-digit employment/wages | not used | n/a | Blocked by captcha; manual download needed if wanted | |
| Argos Translate + OPUS-MT models (task translations) | argostranslate, en-nl 1.8, en-fr 1.9 | Software MIT; models from OPUS-MT (CC BY 4.0) | Mention "machine-translated" in the UI; no credit required for the output | OK |