There is a number you’ve already heard: 47% of jobs are at risk of automation. There is a second number, far less quoted, estimating the same phenomenon over the same years: 9%.
It isn’t that one of them got the arithmetic wrong. They measure two different things, and understanding the difference is the fastest route to actually answering “will my job disappear”. This guide reconstructs what the studies say, what field experiments show, and what you can concretely take from it.
47% and 9%: the difference is the unit of measurement
The 47% comes from Frey and Osborne, The Future of Employment (2013, later published in Technological Forecasting and Social Change, 2017). The authors classified occupations by their probability of being automated, concluding that around 47% of US employment fell into the high-risk category.
The 9% comes from Arntz, Gregory and Zierahn for the OECD (2016). Same period, same question, one decisive methodological change: instead of classifying occupations, they broke each job down into the tasks that compose it, using PIAAC survey data on what people actually do each day. Result: on average across OECD countries, only 9% of workers perform a task bundle that is almost entirely automatable.
The reason for the gap is simple and worth keeping in mind:
A job isn’t a single block. It is a collection of tasks, some automatable and some not. Count occupations and you get catastrophic numbers; count tasks and you get much lower numbers and a far more useful picture.
A later OECD study (Nedelkoska and Quintini, 2018) put the estimate at around 14% at high risk, adding that a further 32% or so will be substantially transformed without disappearing. Statistically, that last category is where you are.
What the field experiments show
Risk estimates are forecasts. In recent years direct measurements on real workers using generative AI have arrived, and they are more informative.
Customer support. Brynjolfsson, Li and Raymond (Quarterly Journal of Economics, 2025) followed 5,179 support agents equipped with an assistant built on language models. Average productivity increase: +14%. But the interesting result is the distribution: +34% for the least experienced workers, close to zero for the most experienced.
Professional writing. Noy and Zhang (Science, 2023) assigned writing tasks to 453 professionals: completion time down by about 40%, assessed quality up by about 18%, and a narrowing of the gap between strong and weak starters.
Consulting. Dell’Acqua and colleagues (Harvard Business School, 2023) worked with Boston Consulting Group consultants. On tasks inside the tool’s range: +12% tasks completed, +25% faster, markedly higher quality. On tasks outside that range, which looked similar but required information the model didn’t have, AI users were 19 percentage points less likely to produce a correct answer.
Three studies, one coherent conclusion:
| What AI does | Measured effect |
|---|---|
| Compresses the gap between novices and experts | Gains concentrated among those who know less |
| Accelerates standardisable tasks | Time −40%, quality +18% on writing |
| Degrades tasks outside its range | −19 percentage points of correct answers |
That third effect is the real professional risk, and it is rarely discussed: it isn’t that AI replaces you, it is that it makes you wrong with more confidence.
The macroeconomic piece that cools the hype
On aggregate effects, it is worth reading the most sceptical estimate around, because it comes from an economist who has studied automation for twenty years.
Daron Acemoglu (The Simple Macroeconomics of AI, 2024) computed the plausible effect of AI on total factor productivity, starting from the share of tasks actually exposed and from documented cost savings. Result: an increase on the order of less than 1% over ten years, far from predictions of radical transformation.
That is not a prediction of irrelevance: it is a reminder that between “this technology works” and “this technology changes the economy” sit adoption times, reorganisation costs and frictions measured in years.
On the other side, the World Economic Forum’s Future of Jobs Report 2025 estimates for 2030 roughly 170 million jobs created and 92 million displaced, with a positive net balance, and indicates that around 39% of key skills will change. Keep in mind it is based on employer surveys, so on expectations rather than measurements.
The two sources don’t really contradict each other: one measures aggregate economic effect, the other the recomposition of employment. The combined picture is “a lot of movement, little immediate net growth”, which is exactly the scenario in which what you can do matters.
Which tasks are exposed (and which aren’t)
The ILO analysis (Gmyrek, Berg and Bescond, 2023) is the most useful here, because it explicitly separates automation from augmentation.
The tasks most exposed in a substitutive sense share three properties: they are predominantly text-based, they have a quickly verifiable output, and they carry no legal responsibility. Clerical office work is the most affected category, and the study notes that in high-income countries this implies higher exposure for women, since that category is mostly staffed by women.
The least exposed tasks, conversely:
- those requiring accountability for a decision, because someone has to sign;
- those requiring unwritten context, i.e. information that exists only in people’s heads or in relationships;
- those that are physical and non-standardised, from installation work to care;
- those consisting of verifying an automated system’s output, a category that is, predictably, growing.
Four steps to gauge how exposed your job is
1. Break your job down into tasks. For a week, log what you do in half-hour blocks. Then mark each block: automatable, augmentable, irreplaceable. It is Arntz and Gregory’s exercise applied to your specific case, and it gives a more accurate result than any occupational ranking.
2. Shift time towards the third category. You don’t need to change professions: in most roles, changing the distribution is enough. Moving from 10% to 30% of your time on decisions and relationships reduces exposure more than taking a course.
3. Learn the tools before you have to. The measured gains are real, and they are largest precisely for those starting from a weaker position. The operational basics are in how to use ChatGPT at work and how to write effective prompts.
4. Build your own verification routine. The BCG study says damage arrives when you trust the tool outside its range. The countermeasure is trivial and nobody applies it: decide in advance which claims you will always check at the source, regardless of how convincing they sound.
5. Watch where demand moves, not where it disappears. The skills employers say they are looking for change faster than occupations do: we’ve collected them in most in-demand skills and most in-demand jobs.
My view: AI won’t replace you, someone using it might
I think “will AI replace my job” is badly posed not because the answer is reassuring, but because replacement rarely comes from the technology: it comes from a person using it.
Brynjolfsson’s finding hints at this indirectly. If the advantage is largest for the least experienced, the distance between a mediocre professional and a good one narrows, and the most uncomfortable position becomes the one held by people living off an average competence advantage.
I haven’t measured it, so I label it for what it is: my inference from the cited data, not a result of the studies.
Frequently asked questions
So the 47% was wrong? Not wrong: it answers a different question. It estimates the share of occupations technically automatable as wholes, not the share of workers who will lose their jobs. The OECD’s 9%, based on actual tasks, is closer to the question you care about.
Which jobs are most exposed? According to the ILO analysis, clerical roles with a high text component and low decision accountability. Least exposed are non-standardised physical jobs and those where someone must be legally answerable for a decision.
If AI helps beginners more, is specialising still worth it? Yes, but the value shifts: from producing the output to being able to judge it. In every study cited, the part that stays human is recognising when the result is wrong.
Do I need to learn to code to keep up? No, and the data is clear here: the measured gains concern writing, support and analysis tasks, not software development. If you’re interested anyway, we cover it in how to learn to code.
How much time do I have? The most cautious macroeconomic estimates point to contained aggregate effects over the coming decade; employer surveys point to much faster skill change. The practical reading is that the immediate risk isn’t redundancy, it is the gradual obsolescence of your profile.
Sources
- C. B. Frey, M. A. Osborne, The Future of Employment: How Susceptible Are Jobs to Computerisation?, Technological Forecasting and Social Change, 2017 (2013 working version). Estimate of 47% of US employment at high risk, on an occupational basis.
- M. Arntz, T. Gregory, U. Zierahn, The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis, OECD Social, Employment and Migration Working Papers, 2016. Estimate of 9% using a task-based approach.
- L. Nedelkoska, G. Quintini, Automation, Skills Use and Training, OECD, 2018. Around 14% at high risk and around 32% substantially transformed.
- E. Brynjolfsson, D. Li, L. R. Raymond, Generative AI at Work, Quarterly Journal of Economics, 2025. 5,179 support agents; +14% on average, +34% for the least experienced.
- S. Noy, W. Zhang, Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence, Science, 2023. 453 professionals; time −40%, quality +18%.
- F. Dell’Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School Working Paper, 2023. Gains inside the frontier and −19 percentage points of correct answers outside it.
- D. Acemoglu, The Simple Macroeconomics of AI, 2024. Estimate of an aggregate productivity effect below 1% over a decade.
- P. Gmyrek, J. Berg, D. Bescond, Generative AI and Jobs: A Global Analysis of Potential Effects on Job Quantity and Quality, ILO Working Paper, 2023. Distinction between automation and augmentation; exposure of clerical work.
- World Economic Forum, Future of Jobs Report 2025. 2030 projections on jobs created and displaced and on changes in required skills.
- D. Autor, Why Are There Still So Many Jobs? The History and Future of Workplace Automation, Journal of Economic Perspectives, 2015. Complementarity between automation and human labour.


