How many jobs will automation replace? What the studies show

Ask how many jobs automation will replace and you get answers that differ by a factor of five. None of the numbers is wrong. Each one answers a different question: how many occupations a machine could do, how many tasks inside jobs it could do, how many workers touch a tool, or how many jobs employers expect to open and close. This page lines the studies up, shows what each one measures, and ends with the jobs that sit at both ends of the scale. For the wider argument about talent and technology, see the technology section.

How many jobs will be lost due to automation?

Seven widely cited estimates, read side by side, point in no single direction. The table gives each study’s scope and its headline figure exactly as published.

The Oxford and OECD 2016 figures rest on the same list of occupations and the same 2013 scores. They differ only in what they do with them². The employer surveys, such as the Forum’s Future of Jobs Report 2025, count jobs created as well as jobs displaced, which no occupation-risk study does³. The ILO counts exposure, and it is explicit that exposure is not loss⁶. Setting these against one another as if they forecast the same thing is the most common error in automation headlines.

Why the studies disagree: occupations versus tasks

Frey and Osborne scored the probability of computerisation for 702 detailed occupations, and hand-labelled only 70 of them as the training set¹. They put occupations above a 70% probability in a high-risk band, below 30% in a low-risk band, and the rest in between². Their summary was that 47 percent of US employment sits in the high-risk band, meaning those occupations are potentially automatable over some unspecified number of years, perhaps a decade or two¹. The wording is “potentially automatable”, which is not a forecast of job loss, and that qualifier is the first thing headlines drop.

The OECD authors relaxed one assumption. Occupations are bundles of tasks, and not every task in a bundle is easy to automate². Working from the micro data of the OECD’s adult-skills survey, which hold job-task information, in its 2016 working paper, which records the tasks workers do, they found that even in occupations Frey and Osborne rated high risk, workers still perform tasks that are hard to automate, such as face-to-face interaction². The result was 9% for the United States, against 47% for the same country in the Oxford paper².

Worked example (our arithmetic). Apply each share to the same base, the 138.44 million US jobs across the 702 occupations¹. The result is illustrative, because the two studies do not use the same worker counts.

The gap is the whole debate in one line: it is the number of jobs that look doomed when you grade occupations, but ordinary once you look at the task mix inside them. Neither estimate says when, and neither counts jobs created.

The OECD paper also shows how much the occupation approach depends on assumptions. Other studies that copied the Oxford scores onto national job lists reached 35% for Finland and as high as 59% for Germany². The task-based method reached a share between 6% in Korea and Estonia and 12% in Germany and Austria².

Four kinds of number

The literature mixes four measures, and each answers one question and not the others.

Is automation a job killer or a job creator?

On the evidence of three Future of Jobs surveys the answer depends on the year, which says more about employer mood than about technology. The 2020 edition estimated that 85 million jobs may be displaced by 2025⁵. It also expected 97 million new roles to emerge⁵. The 2023 edition projected 83 million jobs lost and 69 million created over five years, a net reduction of 14 million jobs, or 2%⁴. The 2025 edition expects 170 million jobs created, equal to 14% of today’s employment, and displacement of 8% (92 million), a net gain of 78 million or 7%³.

Worked example (our arithmetic). Churn is displaced plus created, taken as a share of the dataset. Checking it against each report’s own figure shows how stable it is.

Employers agree on how much turbulence is coming, a fifth of all jobs in five years, and disagree on the net sign^3,4. The 2025 survey covers more than 1,000 employers representing more than 14 million workers in 22 industry clusters and 55 economies, and was run in late 2024³. Treat the sign as a mood captured at one moment, not as a measurement.

The OECD’s reading of the early evidence is more cautious than either. It found that the impact of AI on job levels has been limited so far⁷. Adoption is at a very early stage and concentrated in the largest firms⁷. Many early adopters appear reluctant to dismiss staff, and adjust through slowed hiring, voluntary quits and retirement⁷. In a separate study cited by the OECD, of the firms adopting AI, 26% said it increased employment, against fewer than 10% that saw employment decrease⁷. The same chapter cites a US census sample in which just 3.2% of firms used AI between 2016 and 2018⁷.

Forecasts versus reality: how the task numbers aged

Employer expectations of automation have been revised more than once, which makes them a useful test. The 2020 survey respondents expected almost half (47%) of business tasks to be automated within five years⁴. The 2023 edition found that organizations estimated 34% of business tasks were performed by machines, with the remaining 66% performed by humans⁴. That was a 1% increase on the 2020 estimate⁴. Respondents revised their expectation down to 42% of tasks automated by 2027⁴.

The 2023 report offers a reason. The fraction of automated tasks may have risen less than expected because machines and algorithms augmented human performance rather than automating tasks in that period⁴. The same edition found the human-machine frontier moving at a slower pace than previously anticipated⁴. By 2027, task automation was expected to vary from 35% of reasoning and decision-making to 65% of information and data processing⁴.

The 2025 edition splits the question three ways. Respondents estimate that 47% of work tasks are mainly done by humans alone, 22% mainly by technology and 30% by a combination³. By 2030 the global estimate for tasks done mainly by technology is 34%³. The share augmented by technology is 33% globally³.

Worked example (our arithmetic). The report prints the 2030 shares for technology and for the combination, not for humans alone. Subtracting gives the human-only share and a check against the report’s own total³.

Of that nearly 15-point reduction in the human share, nearly 82% is attributable to advancing automation and 19% to expanded human-machine collaboration³. Our arithmetic: 82 + 19 = 101, a rounding artefact. The sectors differ sharply. In insurance and pensions management and in telecommunications, more than 95% of the fall in human-only tasks is expected to come from deeper automation³. In medical and healthcare services and in government, nearly half of the fall is expected from augmentation and collaboration instead³.

The Forum adds a caveat the headlines omit. The task-share analysis does not consider any change in the absolute amount of work getting done, so any concern about humans “running out of things to do” would be misplaced³.

Which jobs are most at risk?

The Oxford paper is the only source here that scores individual occupations, and its table makes the pattern clear. The scores below are from the 2013 paper’s ranked list of 702 occupations, where 1 would mean certain computerisation¹.

The bands are the OECD’s description of the authors’ cut-offs: low risk under 30%, medium 30 to 70% and high above 70%². Two points follow. First, the high band is full of routine administrative and sales work: the paper names cashiers, counter and rental clerks and telemarketers among the high-risk sales occupations¹. Second, wages and education run opposite to risk. The authors find a strong negative relationship between wages, education and the probability of computerisation¹, and the OECD’s later work reaches the same direction: it is mostly low-skilled and low-income workers who face a high risk of being automatable².

The two ends of the ranking show the same split at its sharpest.

The same table also shows the model’s age. The 2013 authors judged that perception and manipulation remained hard challenges for robots¹. That was written before generative AI, and the newer evidence from the ILO moves the exposure list toward clerical and professional work, discussed below. The Forum’s 2025 survey agrees in direction: clerical and secretarial workers, including cashiers and ticket clerks and administrative assistants and executive secretaries, are expected to see the largest decline in absolute numbers³. The fastest-declining roles include postal service clerks, bank tellers and data entry clerks³. Graphic designers and legal secretaries sit just outside the top 10 fastest-declining, a first-time prediction that the Forum reads as generative AI’s growing capacity for knowledge work³.

Customer service representatives

Customer service sits in the exposed middle in every source that mentions it. Frey and Osborne scored customer service representatives at 0.55, medium risk and well below telemarketers at 0.99¹. Client information and customer service workers ranked seventh among roles in decreasing demand in the Forum’s 2020 survey⁵, and are among the 15 largest net declines in absolute numbers in its 2025 survey³. Exposed roles such as clerical, financial and customer service jobs are concentrated in higher-income countries⁶. Yet the skill itself is not fading: service orientation and customer service are among the skills that have grown in relevance in the Forum’s survey³. The job that answers the phone is exposed, while the judgement behind good service is not.

Retail

Retail shows how one sector can sit on both lists. Frey and Osborne scored retail salespersons at 0.92 and cashiers at 0.97, both firmly high risk¹. The Forum’s 2023 and 2025 surveys expect cashiers and ticket clerks to fall fast^4,3. Yet salespersons are also named among the frontline roles expected to add the largest numbers of jobs in absolute terms by 2030, with farmworkers, delivery drivers, construction workers and food processing workers³. Editor’s reading: the two forecasts can sit together because checkout work and selling work are different tasks, but no source breaks the retail job down that way.

Roles that keep appearing on declining lists

Two Forum surveys five years apart give a rare comparison. The 2020 edition ranked roles in decreasing demand across industries, and the 2025 edition ranks the fastest net declines. The two rankings use different criteria and samples, so ranks are not comparable one for one. The names are^5,3.

Four roles sit on both lists: data entry clerks, administrative and executive secretaries, accounting, bookkeeping and payroll clerks, and material-recording and stock-keeping clerks^5,3. Data entry clerks were first in 2020 and are third in 2025^5,3. Administrative and executive secretaries moved from second to fifth^5,3. Record-keeping and administrative work is the stable core of every declining list. In 2020 the same Forum survey expected increasingly redundant roles to fall from 15.4% of the workforce to 9% by 2025, and emerging roles to grow from 7.8% to 13.5%⁵.

Where the new jobs are

The Forum’s 2025 survey expects growth in three families of work. Frontline roles are predicted to see the largest growth in absolute terms of volume³. Farmworkers top the list, with 35 million more jobs expected by 2030 in a role that currently employs more than 200 million workers worldwide³. Care economy jobs, such as nursing professionals, social work and counselling professionals and personal care aides, are also expected to grow significantly³. Technology roles grow fastest in percentage terms, including big data specialists, fintech engineers, AI and machine learning specialists and software and application developers³.

The same two surveys give the growth side. Again the criteria differ, and again the names recur.

Big data specialists, AI and machine learning specialists, software and applications developers and internet of things specialists appear on both lists^5,3. The growth side has been consistent for five years, and it is the technology-building side^5,3.

In skills terms, AI and big data top the list of fastest-growing skills, followed closely by networks and cybersecurity as well as technology literacy³. Employers also expect creative thinking, resilience and curiosity to rise alongside technical skills³. Manual dexterity, endurance and precision, and reading, writing and mathematics are expected to lose relevance as digital access, AI and robotics automate these tasks³. The 2017 edition of the GTCI had made the same pairing its third message, that technical skills plus social and project competence form the new talent profile⁸.

Jobs AI will not replace by 2030

No source provides a list of safe jobs, and the question assumes a certainty that none of the forecasts claims. What the sources do support is a description of the low-risk end. In the Oxford ranking of 2013 the lowest probabilities belong to recreational therapists at 0.0028 and mental health and substance abuse social workers at 0.0031¹. Elementary school teachers score 0.0044 and registered nurses 0.009¹. The OECD’s 2023 work agrees that high-skill occupations have the lowest risk of automation⁷. It also finds high-skill jobs the most exposed to advances in artificial intelligence, a difference worth holding onto: exposure means the tool can do part of the work, and risk means it may replace the worker⁷. AI can already answer 80% of the literacy questions of the OECD adult-skills survey, and two-thirds of its numeracy questions⁷.

The common thread is the bottleneck tasks. Frey and Osborne judged that perception and manipulation remained hard, and the paper’s other limiting factors, creative and social intelligence, sit behind the same low scores¹. Care and education roles are the ones the Forum expects to grow³. A low score is not a promise. The 2025 ILO paper shows that some strongly digitised occupations have increased exposure as the tools improve⁶.

Who is exposed to generative AI?

The ILO’s 2025 working paper is the newest measure and takes the third view, exposure, as opposed to prediction. Globally, one in four workers are in an occupation with some generative AI exposure, and 3.3% of global employment falls into the highest exposure category⁶. The revised estimates are lower than the ILO’s 2023 paper, though still one in four jobs globally (24%) has some degree of exposure⁶.

The paper’s finding for the job-killer question is its most useful line. As most occupations consist of tasks that require human input, transformation of jobs is the most likely impact of generative AI⁶. Exposure does not imply the immediate automation of an entire occupation, but the potential for a large share of its tasks to be performed with the technology⁶. Clerical occupations continue to have the highest exposure⁶.

The pattern by country income runs opposite to what many readers expect. Overall exposure is 11% of total employment in low-income countries against 34% in high-income countries⁶. Women are more exposed at the top: in high-income countries 9.6% of female employment sits in the highest gradient, against 3.5% of male⁶. The ILO sorts exposed occupations into four gradients, summarized below⁶.

Countries and workers: where risk concentrates

The OECD’s 2016 task-based analysis found that the share of jobs at high risk is highest in Germany and Austria (12%) and lowest in Korea and Estonia (6%)². Countries with a strong focus on high-qualified workers typically have lower shares at high risk, since these workers perform fewer automatable tasks². One caution from the paper: a country can look more automatable than another simply because the other has already automated and replaced labour with capital².

The OECD’s Employment Outlook 2023 adds a country ranking for the highest-risk occupations. Luxembourg, the United Kingdom and Sweden have the lowest shares of employment in those occupations, while Hungary, the Slovak Republic and the Czech Republic have the highest shares⁷. It finds that low- and middle-skilled jobs are the most at risk, including construction and extraction, farming, fishing and forestry, and to a lesser extent production and transportation⁷.

The information technology sector is leading the way in AI adoption, while industries such as construction are lagging behind³. Austria expects 42% of tasks augmented by technology by 2030, against a global average of 33%³. The Forum’s survey adds the employers’ side. Process and task automation is the second most common workforce strategy, at 73% of employers, and is more pronounced in high-income economies (77%) than in upper-middle-income (74%)³. Employers expecting to be affected by ageing populations are more likely to accelerate process automation, 79% against 73% globally³. Global robot density reached 162 units per 10,000 employees in 2023, double the number measured seven years earlier³. The GTCI itself added robot density and investment in emerging technologies in its 2020 edition, as a Technology Adoption component⁹. For how that fits the model, see what the index measures and its methodology.

What employers and workers expect

Employers’ stated plans, set side by side, show where the two surveys asked about the same thing^5,3.

The two surveys word these questions differently and sample different employers, so the table lines up topics, not trends. Within the 2025 survey, half of employers plan to reorient their business in response to AI, two-thirds plan to hire talent with specific AI skills, and 40% anticipate reducing their workforce where AI can automate tasks³. Skill gaps are the biggest barrier to business transformation, cited by 63% of employers³. The OECD found the same obstacle from the other side: two in five companies say a lack of relevant skills is a barrier to using AI at work⁷.

Workers are less uniformly gloomy than the headlines. The OECD surveyed over 2,000 employers and 5,300 workers in the manufacturing and finance sectors of seven OECD countries in 2022⁷. Nearly two-thirds (63%) of workers reported that AI had improved their enjoyment at work, because it automated dangerous or tedious tasks⁷. Yet 60% of workers are also worried about losing their jobs to AI in the next ten years⁷. Employers in the same survey said AI had raised the importance of specialised AI skills, and raised the importance of human skills even more⁷.

The GTCI’s 2017 foreword went further and said that, with today’s rapid technological change, the Luddite fallacy is now perhaps less of a fallacy⁸. The same edition added an Employability sub-pillar to its model, measuring the extent of skill gaps and labour market mismatches in each country⁸. The GTCI’s 2017 edition framed the stakes in the same terms. Its first message was “Think beyond automation”⁸. It said many unskilled jobs would continue to be automated, through the replacement of humans by robots and algorithms empowered by big data⁸, and that low-skilled workers continue to be replaced by robots while knowledge workers are displaced by algorithms⁸. The 2020 edition, on artificial intelligence, concluded that AI in the workplace requires a massive re-skilling of the workforce, and that human-AI teams could be more productive than either AI or workers alone⁹. The longer story of that theme is on the talent and technology page. The skills side of the same problem is in skills mismatch, and the cross-border side in international talent mobility.

How to read an automation headline

Five questions sort most claims.

• Which kind of number is it? Occupation risk, task automatability, exposure or an employer expectation, as in the four-way table above.
• What is the base? Jobs in a survey dataset, the whole workforce, or the formal sector^3,4.
• What is the horizon? The Oxford paper gives “a decade or two”¹. The Forum’s surveys give five years³.
• Does it count creation? Only the employer surveys do³.
• How old is it? Employer forecasts from 2020 and 2023 were revised within three years⁴.

An editor’s rule of thumb: when a headline says “X million jobs lost” and the source is an exposure index or an occupation score, the headline has quietly turned a risk into a forecast.

Conclusion

The evidence supports three statements and no more. Estimates of the share of jobs at risk fall between 9% and 47% because they measure different things, and the task-based figure is the more cautious one^1,2. Employer surveys expect large churn, near a fifth of jobs over five years, with a net sign that has flipped between editions^3,4. And the most recent exposure data point to clerical and routine administrative work, with transformation, not disappearance, as the likelier outcome for most occupations^6,7.

What the sources do not support is a single number for jobs lost, a date, or a list of survivors. The honest answer to “how many jobs will automation replace?” is a range, a method and a caveat, in that order.