Augmentation means AI takes over some tasks in a job while a person keeps the rest. Automation means most of a job's tasks could be done by AI. The ILO, the United Nations agency for work and employment standards, sorted occupations along that line in a May 2025 update of its global exposure index: about one in four workers worldwide holds an occupation with some generative AI exposure, and 3.3 percent hold one in the highest category, where automation is the closest fit. The index measures potential, not layoffs.
Key takeaways
- The ILO's 2025 index scores tasks, then groups occupations into four exposure gradients based on the average task score and how much scores vary within the job.
- Gradients 1 to 3 together cover 20.5 percent of world employment, and gradient 4, the closest match to automation, covers 3.3 percent.
- In high-income countries, which include Canada and the United States, 34 percent of employment is in an exposed occupation, against 11 percent in low-income countries.
- The ILO calls its figures an upper threshold and states that they are exposure, not actual impact on employment.
How task scores become an augmentation or automation label
The ILO's researchers did not rate whole occupations. They rated tasks. A survey of 1,640 employed people, drawn from each one-digit group of the international occupation classification ISCO-08, produced 52,558 ratings of how far generative AI could automate 2,861 tasks. Those tasks were a sample from the 29,753 tasks in Poland's detailed occupational classification. Delphi-style rounds with international experts were then used to check the scores, and an AI assistant was built to predict scores for every task in the ISCO-08 documentation.
Each occupation ends up with two numbers: the mean of its task scores and the spread of those scores. A job whose tasks all score high looks very different from a job where a few tasks score high and the rest score low, even if the means match. That spread is what turns a single exposure score into a distinction between automation and augmentation.
The four gradients, defined
| Gradient | Rule (mean task score, spread) | What it describes |
|---|---|---|
| Gradient 1 | Low mean, high task variability | Some tasks highly automatable, many still needing a person. The paper ties this to the older idea of augmentation. |
| Gradient 2 | Mean from 0.4 to below 0.5, high variability | A mix of exposed and unaffected tasks, so impact is uneven. |
| Gradient 3 | Mean from 0.5 to below 0.6, high variability | A significant share of tasks consistently exposed, with adaptation strategies needed. |
| Gradient 4 | Mean of at least 0.6, low variability | Most current tasks have high automation potential. The paper ties this to the older idea of automation risk. |
Occupations outside the four gradients are labeled either minimal exposure, where some interaction with generative AI may occur without changing the role much, or not exposed, where the technology currently has no observable potential to automate tasks.
How much employment sits in each group
Worldwide, about a quarter of employment falls in one of the four gradients. Gradients 1 to 3 together hold 20.5 percent, and gradient 4 holds 3.3 percent. The ILO's 2023 version of the index used different labels: 2.3 percent of employment in an automation category, 13 percent in augmentation, and 16.2 percent in a "big unknown" group. The 2025 paper says the new numbers are consistent with the earlier ones but more refined, and that the share left in the uncertain middle shrank.
Income level changes the picture. In high-income countries, 34 percent of total employment falls in an exposed gradient, and 17.3 percent sits in gradients 3 and 4. In low-income countries the exposed share is 11 percent. The paper links this to occupational structure, since clerical, financial, and customer service jobs are concentrated in wealthier economies. Clerical occupations have the highest exposure overall. The paper also reports that women are more exposed than men: 4.7 percent of female employment worldwide is in gradient 4, against 2.4 percent of male employment.
Why a high score is not a layoff notice
The ILO states plainly that its estimates reflect exposure and not the actual effect on employment, and that they mark the upper limit of what could be affected if the technology were fully adopted. It lists barriers to that adoption: electricity and broadband infrastructure, limited digital skills, the cost of the technology, and operational difficulty. It adds that even where automation is technically possible, institutions and customers may resist it. Few consumers might want the evening news read by an avatar, and a bank may keep a person responsible for a mortgage decision.
The paper also declines to treat augmentation as safe. A job made of many tasks acts as a buffer against full automation, but efficiency gains could still mean fewer workers are needed, and a job stripped down to its leftover tasks may become unattractive to do. Its estimates also leave out new jobs that the technology might create.
Questions to put to any claim about augmentation or automation
- Does the claim describe tasks, or whole jobs?
- Is the figure a measure of exposure (what AI could do) or of observed change in hiring, hours, or pay?
- Does the source report how much scores vary inside an occupation, or only an average?
- Which year and which version of the technology does the score reflect?
- Does the claim count new tasks and new jobs, or only the tasks that disappear?
FAQ
Is augmentation always better for workers than automation?
No such rule appears in the ILO paper. It notes that even partial automation of tasks can reduce demand for an occupation if the job does not take on new tasks, so the outcome depends on how work is reorganized.
Does the 3.3 percent figure mean that many jobs will disappear?
No. It is the share of world employment in occupations whose tasks score highest and most consistently on automation potential. The ILO describes all of its figures as exposure and an upper threshold.
Why did the ILO change its categories between 2023 and 2025?
The 2025 update added survey data, expert review, and a revised framework with four gradients, plus a distinction between minimal exposure and not exposed. The authors point to newer models that handle images, audio, and video as one reason for revisiting the scores.
Do these figures describe Canada and the United States specifically?
Not directly. The index reports world, regional, and income-group totals. Both countries belong to the high-income group, whose 34 percent exposed share is the closest published reference point, and national micro data would be needed for country-level projections.
Related reading
- Challenges and opportunities for employment
- How AI and automation differ
- Effects of automation and AI on society
- Job displacement concerns and adaptation
More in AI and Society.
Last verified: October 2026
Sources
- Generative AI and Jobs: A Refined Global Index of Occupational Exposure, Gmyrek P, Berg J, et al., ILO Working Paper 140, May 20, 2025
- Generative AI and Jobs: A 2025 Update, research brief, ILO, 2025