A new task-level atlas finds automation exposure ranges from about 3% of tasks in the poorest countries to over 60% in China, rising strongly with income; low-income countries face relatively more labour-substituting and low-tech automation, whereas AI is more prevalent and more often labour-augmenting in richer economies.
Global Automation Atlas
A global task-level atlas shows automation exposure varies widely across countries (3.3%–61.6% of tasks), rises with income, is skewed toward labor substitution especially in low-income countries, and shows that AI plays a larger augmenting role in high-income settings while contributing more to substitution in poorer countries.
Automation affects the labour content of work differently across different contexts. Yet, most existing exposure measures assign fixed scores to tasks or occupations, limiting comparisons of automation exposure across countries. We develop a task-based and country-specific approach to classify automation exposure across the world to disentangle labor-substituting from labor-augmenting automation, the relevant technology channel, and the material role of AI. Our measure spans 124 countries, generating an atlas of 2.33 million task-country labels for economies covering 99% of world population and GDP. We present five descriptive results. First, exposure is highly uneven, ranging from 3.3% of tasks in South Sudan to 61.6% in China, and rises strongly with income, although substantial variation remains within income groups. Second, across countries, exposed tasks are skewed towards substitution rather than augmentation, but low-income countries are disproportionately exposed to substitution, whereas middle-income countries are more heterogeneous. Third, less technologically advanced forms of automation account for more than half of exposed tasks in low-income countries but about one quarter in high-income countries; while other more complex channels generally rise with income levels. Fourth, AI tends to be less prevalent in simpler channels of automation, but also more prevalent in labour-substituting margins in lower income settings and to augment labour in higher income settings. Fifth, we find that females seem to be disproportionately more exposed to labour-substituting automation than males. Our methodology provides a basis for comparing automation exposure across development stages, linking it with cross-country data and allowing us to treat exposure levels, labour margins, technological channels and AI involvement as separate dimensions.
Summary
Main Finding
The authors construct a task-based, country-specific Global Automation Atlas covering 18,797 O*NET tasks across 124 countries (2.33 million task–country labels) that records, for each task-country pair: an economically meaningful exposure score (0–3), the dominant labour margin (substitution / augmentation / both), the dominant technology channel (5 archetypes), whether AI/ML is materially involved, and the dominant AI function when relevant. Using this architecture they show exposure is highly uneven and rises with income; substitution dominates globally (especially in low-income countries); automation channels and the role of AI vary strongly with development; occupational and industry aggregates reproduce plausible patterns (with women more exposed to substitution); and country covariates (GDP per capita, internet use, schooling, regulatory quality, capital intensity) predict exposure and margin composition.
Key Points
- Dataset and scope
- 18,797 standardized O*NET tasks × 124 countries → 2,330,776 task–country observations.
- Countries cover ≈99% of world population and GDP; aggregated outputs at country, occupation (ISCO), and industry (ISIC4) levels.
- Public atlas and data: automationatlas.org.
- Measurement schema (separate dimensions)
- Exposure level (0 none, 1 assistive, 2 partial economic exposure, 3 extensive economic exposure). “Exposed” = levels 2 or 3.
- Labour margin: substitution-only, augmentation-only, balanced-both, or unclear.
- Technology channel (dominant mechanism): physical execution; rule-based workflow; information transformation; planning/control; inference/scoring.
- AI materiality: whether contemporary learned models are central; dominant AI function when material: state inference, content transformation, recommendation/decision support, adaptive control.
- Country conditioning: each task is evaluated in the named country’s production/ institutional context.
- Five main empirical/descriptive results
- Exposure heterogeneity: exposed-task share ranges from 3.3% (South Sudan) to 61.6% (China); exposure rises with per-capita income but with large within-tier variation (especially middle-income).
- Labour-margin skew: substitution-only exposure exceeds augmentation-only in every country; low-income countries are particularly biased toward substitution; middle-income countries show more heterogeneity.
- Channel composition by income: simpler automation channels (e.g., rule-based workflow, physical execution) account for a majority of exposed tasks in low- and lower-middle-income countries but only about one quarter in high-income countries; more complex channels (planning/control, inference, information transformation) increase with income.
- AI’s role varies: the share of exposed tasks where AI is materially involved ranges ~35%–70% across income brackets; AI is less prevalent in simpler channels, ties more to substitution in lower-income settings and to augmentation/shared workflows in higher-income settings.
- Occupation/industry aggregates & distributional patterns: clerical, transactional, routine information-processing tasks show exposure at lower income levels; business admin, ICT, plant operation, certain manufacturing and information services rise with development. Crosswalks with ILOSTAT indicate women are more exposed to labour-substituting automation across occupations and industries.
- Predictors: GDP per capita, internet use, years of schooling, and regulatory quality are the most informative correlates for the exposed-task share. Within exposed tasks, substitution composition falls with income, regulatory quality, and capital intensity; augmentation composition correlates most with capital intensity.
- Validation: labels produced with Gemini 3.1 Flash-Lite under a fixed structured prompt protocol; validated via (i) construct comparisons with existing measures (LLM exposure, AI exposure, robot exposure, country AI preparedness, firm adoption), (ii) convergence tests with an independent model family, (iii) reasoning/prompt-paraphrase consistency checks, and (iv) face validity inspections.
Data & Methods
- Task universe: O*NET v29.1 (Nov 2024), 18,797 standardized task statements and occupation links.
- Classification pipeline:
- LLM-based structured classification (Gemini 3.1 Flash-Lite) applied to each task–country pair with a fixed protocol and prompts that elicit the five dimensions above.
- Country conditioning: tasks evaluated under named country conditions; also run context-free and income-group benchmarks.
- Exposure coding: 0–3 scale with level 2+3 counted as economically exposed.
- Channel taxonomy designed to be mechanism-first (the mechanism producing the economically relevant output determines the channel).
- AI materiality distinct from channel; AI function recorded only when AI is central to the route.
- Aggregation: task–country labels aggregated to produce country-level exposed shares, occupation- and industry-level exposure measures, and within-exposed margin shares (renormalized among exposed tasks when needed).
- External linkages: matched to ILOSTAT employment shares and country-level covariates (GDP per capita, internet use, years schooling, regulatory quality, capital intensity, etc.).
- Validation strategy:
- Construct validity: comparisons to Eloundou et al. (LLM/LLM-exposure), Felten et al. (AI exposure), Webb (robot/automation), Cazzaniga et al. (AI preparedness), Eurostat firm adoption.
- Convergent validity across model families and prompt variations; face validity via anchor occupations and label distributions.
- Limitations discussed by authors: LLM-based labeling requires careful validation; labels reflect current-technology plausibility and country context but are not direct measures of adoption or realized job loss.
Implications for AI Economics
- Measurement implications
- Task-level, country-conditioned measurement is crucial: occupation- or country-fixed task scores mask cross-country differences in feasible automation margins driven by capital, infrastructure, skills, and institutions.
- Separating exposure, labour-margin, channel, and AI materiality gives richer inputs for causal and policy analysis than single-index measures that conflate these dimensions.
- Research uses
- Comparative studies of automation impacts: atlas enables cross-country analyses of wage effects, employment compositional change, and sectoral reallocation that account for differing margins (substitution vs augmentation) and channels.
- Identification and structural work: country-task labels can be used as instruments, controls, or treatment definitions in studies of adoption, firm response, and labor-market adjustment.
- Distributional analysis: the strong substitution bias in low-income countries and gendered exposure patterns imply heterogeneous welfare and inequality consequences; the atlas can help quantify exposure-driven risks and guide targeted reskilling policy.
- Modeling diffusion and complementarities: channel and AI-function labels permit analyses of how complementary capital, skills, and institutions mediate the adoption of different automation technologies (e.g., robotics vs ML-driven information transformation).
- Policy implications
- Development-sensitive policy design: low-income countries face a higher share of substitutable tasks under simpler automation channels—policy priorities should emphasize social protection, job creation in non-substitutable tasks, and investment in complementary capital and skills.
- Education and connectivity matter: schooling and internet access are strong predictors of exposure shares; investments here can affect both exposure and the ability to benefit from augmentation.
- Regulation and governance: regulatory quality predicts exposure composition; policy frameworks shape whether automation leads to augmentation versus substitution and how firms deploy AI.
- Gender-responsive interventions: female-concentrated occupations are disproportionately exposed to substitution; active labor-market programs and retraining should consider gendered task composition.
- Cautions and future directions
- Atlas measures plausibility of automation today (exposure), not realized adoption or displacement. Linking labels to firm-level adoption and dynamic labor outcomes is necessary to trace realized effects.
- LLM-based classification is powerful and scalable but requires ongoing validation; the authors’ multi-pronged validation is reassuring but users should treat labels as informed assessments, not ground truth.
- Future empirical work should leverage the atlas to study causal determinants of automation adoption, the dynamics of task reallocation, interaction between AI function and labor margins, and welfare impacts across development stages.
Short takeaway: the Global Automation Atlas provides a scalable, validated, task-level, country-conditioned framework that disentangles exposure magnitude, labour-margin, technology channel, and AI role—an essential measurement advance for cross-country research on automation, AI impacts, and development-sensitive policy design.
Assessment
Paper Typedescriptive
Evidence Strengthmedium — The paper assembles a large, systematic task-by-country dataset covering 124 countries and 2.33 million task-country labels, enabling credible descriptive comparisons across contexts and careful disaggregation by automation channel and AI involvement. However, it does not estimate causal effects; results depend on mapping rules, exposure scoring thresholds, and task-occupation crosswalks (often derived from high-income sources), which can introduce measurement error and classification bias, especially in low-income settings.
Methods Rigormedium — Methodological strengths include a task-level approach, country-specific adjustments, explicit separation of substitution vs. augmentation and technology channels, and broad geographic coverage. Rigor is limited by reliance on external task datasets (likely produced for advanced economies), subjective choices in labeling and thresholds for 'exposure', potential crosswalk/misalignment between tasks and local job content, and limited validation of labels against on-the-ground automation adoption.
SampleTask-level dataset covering 124 countries (economies representing ~99% of world population and GDP) producing 2.33 million task-country labels; exposure measured at the task-country level and aggregated by country income groups, sectors and worker demographics (e.g., gender). Likely construction uses standard task banks crosswalked to country employment structures and technology-channel taxonomies (including AI vs non-AI channels).
Themeslabor_markets adoption
GeneralizabilityTask descriptors and exposure scores originate from task datasets developed in high-income countries and may not fully capture actual job content in low- and middle-income countries., Crosswalks between tasks and national occupation/industry classifications can introduce misclassification, especially where informal work is large., Static exposure measures may not reflect rapid, recent adoption or local institutional constraints on automation., Definitions of 'AI' and classification into technology channels involve judgment calls that affect cross-country comparability., Aggregating task-level exposure to country-level summaries can mask within-country regional and firm-level heterogeneity.
Claims (7)
| Claim | Direction | Confidence | Outcome | Details |
|---|---|---|---|---|
| Our measure spans 124 countries, generating an atlas of 2.33 million task-country labels for economies covering 99% of world population and GDP. Other | positive | high | coverage of task-country labels / dataset scope |
n=2330000
124 countries; 2.33 million task-country labels; 99% of world population and GDP
0.3
|
| Exposure to automation is highly uneven across countries, ranging from 3.3% of tasks in South Sudan to 61.6% in China, and exposure rises strongly with income (with substantial within-group variation). Adoption Rate | positive | high | share/percentage of tasks exposed to automation |
n=124
3.3% (South Sudan) to 61.6% (China); exposure rises with income
0.18
|
| Across countries, exposed tasks are skewed towards labour-substituting automation rather than labour-augmenting automation; low-income countries are disproportionately exposed to substitution, whereas middle-income countries are more heterogeneous. Job Displacement | mixed | high | proportion of exposed tasks classified as labour-substituting vs labour-augmenting |
n=124
0.18
|
| Less technologically advanced forms of automation account for more than half of exposed tasks in low-income countries but about one quarter in high-income countries; more complex technological channels generally rise with income levels. Adoption Rate | positive | high | share of exposed tasks attributed to 'less technologically advanced' channels vs 'more complex' channels |
n=124
more than half (>50%) in low-income countries; about one quarter (~25%) in high-income countries
0.18
|
| AI is less prevalent in simpler channels of automation overall, but AI is more prevalent on labour-substituting margins in lower-income settings and tends to augment labour in higher-income settings. Adoption Rate | mixed | high | prevalence of AI involvement in automation channels and by labour margin (substitution vs augmentation) |
n=124
0.18
|
| Females seem to be disproportionately more exposed to labour-substituting automation than males. Inequality | positive | high | gender gap in exposure to labour-substituting automation (female vs male exposure) |
n=124
0.18
|
| The paper's methodology enables classification of automation exposure that disentangles labour-substituting from labour-augmenting automation, identifies the relevant technology channel, and records the material role of AI — allowing exposure levels, labour margins, technological channels and AI involvement to be treated as separate dimensions across development stages. Other | positive | high | granularity/capacity of the measurement methodology (ability to separate multiple dimensions of exposure) |
n=124
0.18
|