Methodology

How we measure
automation pressure

A transparent account of the data pipeline, scoring formulas, and sources behind the Automation Risk & Reskilling Intelligence platform — covering 185 professions across 127 countries.

185+
Professions mapped to ISCO-08
127
Countries with individual indices
4
Independent data sources
23,595
Profession × country data points

What this platform is — and isn't

The platform quantifies two things for every combination of profession and country: how likely a given profession is to be disrupted by automation and AI (the Automation Score), and how capable a country's workforce is of adapting to that disruption (the Reskilling Score).

Core question

For a given profession in a given country: how hard is automation hitting it, and how prepared is the workforce to respond?

This is not a prediction engine. The scores are relative risk indicators — a structured, comparable, evidence-based way to prioritise attention across a large matrix of professions and geographies. They do not claim to forecast exactly which jobs will disappear or when.

The automation probability figures come from peer-reviewed academic research (Frey & Osborne), the most widely cited quantitative study in this field. Country-level indices come from publicly available institutional datasets. Neither the underlying research nor this platform should be read as deterministic.

Four open, institutional datasets

All four underlying datasets are open, institutional, and freely available. No proprietary or paid data feeds are used — ensuring reproducibility and allowing the methodology to be independently verified.

1
Frey & Osborne — Automation Probability
Oxford economists analysed 702 US occupational categories and assigned each a probability of computerisation based on susceptibility to machine learning, mobile robotics, and computational intelligence across three bottleneck dimensions: perception & manipulation, creative intelligence, and social intelligence tasks. Supplemented with ISCO-08 and WEF Future of Jobs 2023 for additional emerging roles. Read paper →
2
Oxford Insights Government AI Readiness Index — AI Adoption
Annual index scoring countries on readiness to implement AI, drawing on 30+ sub-indicators covering government strategy, technology infrastructure, data availability, regulatory environment, and human capital. Used as a proxy for AI-driven automation adoption in each country's economy. Non-covered countries estimated from OECD Digital Economy Outlook, WEF Global Competitiveness Index, and ITU ICT Development Index. View data →
3
World Bank Open Data — Employment Growth
Employment-to-Population Ratio (SL.EMP.TOTL.SP.ZS), comparing the most recent available year against a five-year prior baseline. Enters the Automation Score as a moderating factor: shrinking labour markets amplify displacement risk; growing markets partially offset it. Growth rates clipped to ±30% to prevent extreme values distorting the composite score. View data →
4
UNDP Human Development Index — Education Index
Education Index component from the UNDP Human Development Report, measuring educational attainment via mean years of schooling for adults and expected years of schooling for children — normalised to [0, 1]. Serves as a structural capacity indicator in the Reskilling Score, representing the supply side of reskilling and the absorptive capacity of the workforce. View data →

Five deterministic steps

The pipeline is fully deterministic and reproducible. Given the same input datasets, the same scores will always be produced.

1
Occupation Normalisation & Taxonomy Mapping
Raw occupation titles are mapped to the ISCO-08 major group taxonomy using a two-stage process: a rule-based keyword matcher first (instant, high throughput), then an LLM via Ollama for ambiguous or non-standard titles. All LLM classifications are cached to disk — each title is classified only once.
2
Skill Change Rate Estimation (LLM-Augmented)
The most analytically nuanced input. Estimated per profession using a large language model with a calibrated prompt and anchored reference scale. Distinct from replacement risk — a truck driver and a software developer may both face disruption, but the pace and nature of skill change differ dramatically. Values cached indefinitely.
3
Country Index Construction
Three country-level indices (AI adoption, employment growth, education) are downloaded, normalised to [0, 1], and merged into a single country table. Live downloads attempted at each quarterly refresh; failures fall back to a curated embedded dataset covering all 127 countries.
4
Score Computation
Automation Score and Reskilling Score are computed for each profession × country pair using the multiplicative formulas below. Both outputs are clipped to [0, 1].
5
Output & Refresh Cadence
A single flat CSV is produced with one row per profession-country pair (23,595 rows). All source data cached with a 90-day TTL. Skill Change Rate estimates cached indefinitely and only re-queried when the underlying LLM model is updated.

The scoring formulas

Automation Score — a composite measure of effective automation pressure, combining the inherent replaceability of an occupation with the country's actual capacity to deploy automation technology.

Automation Score Formula
automation_score = automation_probability × ai_adoption_index × (1 − employment_growth)
Three-factor multiplicative model. All inputs on [0, 1]. Output clipped to [0, 1].
VariableSourceRole
automation_probabilityFrey & OsborneOccupational baseline — does not vary by country
ai_adoption_indexOxford Insights 2022Country's structural propensity to deploy AI — same occupation in a low-adoption country scores lower
(1 − employment_growth)World Bank 2024Labour market correction — shrinking employment amplifies risk; growing employment moderates it

Reskilling Score — measures potential and urgency for workforce reskilling. A high score indicates both strong need (rapid skill change) and meaningful structural capacity to respond.

Reskilling Score Formula
reskilling_score = skill_change_rate × tech_adoption × education_index
Three-factor multiplicative model. All inputs on [0, 1]. Output clipped to [0, 1].
VariableSourceRole
skill_change_rateLLM-estimated (calibrated prompt)Rate of skill transformation per profession — distinct from replacement risk
tech_adoptionISCO group weightProfessionals (Group 2) carry higher weight than Elementary Occupations (Group 9)
education_indexUNDP HDR 2023Structural capacity to absorb reskilling — workforce's baseline ability to learn and pivot

Reading the two scores together:

↑ Automation · ↑ Reskilling

Strong pressure but real capacity to respond. Priority for structured reskilling programmes.

↑ Automation · ↓ Reskilling

Highest risk zone. Strong displacement pressure with limited structural capacity to adapt. Requires policy-level intervention.

↓ Automation · ↑ Reskilling

Stable profession in a country with high adaptive capacity. Proactive upskilling opportunity.

↓ Automation · ↓ Reskilling

Limited immediate risk. May warrant monitoring as automation expands to lower-probability occupations over time.

Skill Change Rate — Calibration Scale

0.00–0.15  Almost no change  ·  janitor, agricultural laborer
0.16–0.30  Slow change  ·  cook, security guard, construction worker
0.31–0.45  Moderate change  ·  accountant, cashier, paralegal
0.46–0.60  Fast change  ·  financial analyst, marketing specialist, nurse
0.61–0.80  Very fast change  ·  software developer, radiologist, journalist
0.81–1.00  Near-complete transformation  ·  AI/ML engineer, data scientist, prompt engineer

What we're transparent about

We believe in being direct about what the data can and cannot tell you. The following limitations are inherent to the methodology and should inform how you use the scores.

Skill Change Rate is model-estimated, not measured
Unlike other inputs, Skill Change Rate is not drawn from an empirical dataset — it is estimated by a large language model using a calibrated prompt. The estimates are broadly consistent with domain literature but are not independently validated. Users can adjust rates for specific occupations by editing the LLM cache file directly.
Country indices are structural aggregates, not sectoral
The AI adoption index, employment growth, and education index are national aggregates. They do not differentiate between, say, the tech sector in Germany and the agricultural sector in Germany. For high-level benchmarking this is an acceptable trade-off; for sector-specific analysis, treat country-level scores as broad context.
Scores are relative, not absolute
An Automation Score of 0.85 does not mean "85% of jobs in this profession will be lost." It means this profession-country pair sits near the high end of the relative scale. The scores are designed for comparison and prioritisation, not for forecasting headcount reductions.

Data provenance summary

Every input variable, its source, and its validation status at a glance.

automation_probability
Frey & Osborne · Oxford University
Empirical
ai_adoption_index
Oxford Insights 2022 · via Our World in Data
Empirical
employment_growth
World Bank Open Data 2024
Empirical
education_index
UNDP Human Development Report 2023
Empirical
skill_change_rate
LLM estimation · calibrated prompt
Model-estimated

The dataset refreshes quarterly. Each refresh attempts live downloads from all four source APIs; failures fall back to the most recently cached values. LLM-estimated skill change rates are versioned separately — when the underlying model is updated, the skill cache is cleared and all professions are re-estimated. Scores from different pipeline versions should not be directly compared without reviewing the changelog.