Key takeaways

  1. On this unadjusted occupation-mix measure, the gap between the most and least exposed economies looks small. The most exposed economy, Singapore at 0.34, scores under twice the least exposed, Burundi at 0.17, even though the share of their workforces in exposable office jobs differs by roughly 25 times and their income per head by more than 100 times. Exposure is held down because even the most exposed occupations, clerical and IT work, only score around 0.5 to 0.7 and never make up a whole workforce, and held up because manual and farm work still carry roughly 0.10 to 0.19. But the true gap is probably wider than it looks: this measure applies one global score per occupation, and the same job tends to be less digital in a poorer economy, so it likely overstates low-income exposure.
  2. The map is really a map of office work. One variable, the share of the workforce in clerical, professional, technical and managerial jobs, explains 96 percent of the variation in the index. Exposure does not fall unevenly across countries because of the technology. It does so because of how many people work at a desk.
  3. AI may pull up the ladder behind the rich. Office work concentrates in rich countries, so exposure does too. But the deeper risk is for poorer economies: the clerical and back office jobs AI is most likely to automate are the same rungs they have used to climb from farming toward higher income work. A low score today is not safety. It may mean the next steps up are being removed before they can be taken.

Context

Generative AI means the large language models, or LLMs, behind tools like ChatGPT. It is expected to reshape knowledge work, but almost all of the evidence so far comes from a single country, and most of it is about the United States. This map asks a wider question. Across the world, whose work is most exposed, and how is that exposure spread between countries? It combines the ILO's exposure scores for individual occupations with each country's actual mix of jobs, producing one comparable exposure figure for roughly 170 countries, so the global picture can be seen in one place.

Methodology

This method builds on the key outputs of the ILO's Working Paper 140, A Refined Global Index of Occupational Exposure, published in 2025. That paper rates 427 detailed occupations on a scale from 0 to 1 for how exposed each occupation's tasks are to generative AI. Clerical, data entry and IT work score highest, around 0.55 to 0.70. Cleaners, farm labourers and building trades score lowest, around 0.09 to 0.12. Most occupations, 231 of the 427, are rated not exposed at all.

Those scores are then applied across countries. The weights come from ILOSTAT, and from Eurostat for Europe: each country's share of employment by occupation group, classified under the ISCO occupational standard, at the latest available year for each country, mostly between 2018 and 2024.

For each country, the occupation scores are aggregated up to the level at which its employment data exists, then averaged across the workforce by how many people work in each group. The result is a single share-weighted exposure figure per country, the number that sets its colour on the map.

The result

Across the countries on the map, the index runs from 0.174, in Burundi and Madagascar, to 0.340, in Singapore, with most countries packed close to the middle, around 0.255. Rich economies cluster at the top, around 0.30 to 0.34. The lowest scores belong to low income, largely agricultural economies, around 0.17 to 0.18.

The pattern is that exposure mirrors wealth. The more of an economy's workforce that sits in offices, in clerical, professional, technical and managerial roles, the higher it scores, because those are the tasks today's AI is best placed to reshape. Where most people work in farming, construction or manual services, the score is low. The link is almost mechanical: the office share alone explains 96 percent of the variation.

But on this unadjusted measure the spread is narrow, and that is the more surprising result. Roughly 74 percent of Singapore's workforce is in office jobs, against about 3 percent in Burundi, a gap of some 25 times, yet their exposure scores differ by less than a factor of two. Even the most exposed economy is not mostly exposed, because the most exposed occupations top out below 0.7 and never dominate a whole workforce.

The outliers make the point. Djibouti scores 0.31, sitting among the European leaders, not because it is wealthy but because its small formal workforce is unusually administrative, built around its port and government. The index measures what people do, not what they earn.

For Europe, where finer data is available, the map resolves to a second level of detail. Luxembourg, the Netherlands and Switzerland lead, while Romania and Türkiye sit lowest. The finer view runs a little higher than the global one, because it picks up concentrated pockets of clerical and IT work that the coarse view averages away.

There is a more important observation here. Countries have generally grown rich by moving workers up a ladder: out of farming, into factories and basic services, and then into the clerical, administrative and professional work that pays well. Outsourced back office and call centre work, the kind that built large parts of the Indian and Philippine middle classes, sits on the lower rungs of that office economy, and those are precisely the jobs the exposure scores rate most highly. If AI automates that work before developing countries can grow into it, the ladder is pulled up behind the economies that already climbed it. This is the older idea of kicking away the ladder in a new form: the route that today's rich countries used to get rich may be closing just as poorer ones reach for it. Seen this way, a low exposure score is not protection but a warning. The danger is less that AI displaces poorer countries' current workers, and more that it removes the path they have always used to become richer, turning AI from an engine of catch up into a brake on it.

Explore it in the map above. Click any country to see which parts of its workforce drive its score.

Limitations

This is an experiment, not an authoritative forecast, and being open about its limits is part of the point.

  1. Exposure is not job loss. The score measures how much of an occupation's work AI could plausibly reshape if it were fully adopted, not how many jobs disappear. It says nothing about how fast AI is adopted, what it costs, how it is regulated, or whether employers prefer to keep people. A high score is a measure of potential, not a prediction.
  2. The resolution is capped by the employment data. The scores are detailed, at the level of individual occupations, but harmonised employment data is not, sitting at ten groups worldwide and around forty in Europe. Fine scores are averaged up to a coarse level, so each country's figure is a faithful approximation of the ILO's, not an identical reproduction. Where finer two-digit data exists, in Europe, the index tracks the coarse version closely (correlation about 0.96, running roughly 0.01 higher), so the coarseness does not distort the broad picture, though it is too rough for fine rank claims.
  3. The same occupation is treated as the same job everywhere. The scores are built from Polish occupational task data and then applied worldwide, which assumes a given occupation involves much the same work in every country. In practice the same job is often more manual and less digital in poorer economies, a point the ILO's later work addresses. Because this index applies one score per occupation everywhere, it probably overstates exposure in lower income countries, so the true gap between rich and poor may be wider than the map shows. For comparison, the ILO's own task-content work finds a far wider income gap on a related measure: roughly 11 percent of workers in low income countries have some exposure to generative AI, against 34 percent in high income countries.
  4. National data is not perfectly comparable. Each country's occupation categories are treated as consistent, but national surveys differ in how they define, count and classify jobs.
  5. The snapshot is not synchronised. Each country uses its latest available year, and those years vary, so the map is not a single moment in time.
  6. The scores reflect AI as it stood in 2025. Working Paper 140 was published in May 2025, and its exposure scores capture what AI could do then. The technology has moved quickly since: 2026 alone has brought far more capable agentic tools, from Claude Code and its Cowork mode to OpenAI's Codex, that already automate work the 2025 scores did not fully count. So real exposure today is almost certainly higher than this map shows, and these figures are best read as a 2025 baseline that is already dating.

How AI was used

This piece was made with AI, and it is worth being precise about where and how. AI was used throughout, both in the methodology and in producing the article itself.

The methodology was designed by a human: its structure, the steps it follows, and the data sources it draws on were all defined and then reviewed by hand. AI was used to execute that method, carrying out the specific calculations once the formula and the sources had been set. The calculations themselves were run using Claude Opus 4.8, then passed to OpenAI's Codex (GPT-5.5) to scrutinise the working and flag any errors of method or arithmetic, before a final human review.

The interactive map and its visual elements were built with Claude Opus 4.8. The written sections, including these key takeaways, were produced jointly: drafted with AI, shaped by human input, and signed off in a final human review of the whole piece.

The exposure scores themselves come from the ILO's published research, Working Paper 140. Every figure on this page traces back to a published source: the ILO for the occupation scores, and ILOSTAT and Eurostat for the employment data.