Is AI taking jobs?

It depends on the job. Desk tasks of every length reach AI, yet in Europe the jobs where AI is used grew as its use spread.

The debate about AI and work swings between two claims. One says AI makes people far more productive. The other says AI takes their jobs. Both claims depend on what AI is used for and on how much of that use the data can see. This paper measures both.

About this paper

The paper starts from the tasks people hand to AI. breinz built a catalogue of 608 such tasks from a public record of AI conversations and its own keyword dataset.¹ breinz then matched the catalogue to O*NET, the US government’s catalogue of the tasks in every occupation.² It also matched it to the Anthropic Economic Index, which assigns Claude conversations to those tasks.³ Employment, job ads and output data from the US and Europe show what has happened to the jobs behind the tasks.^{4 5 6 7}

The analysis covers AI use worldwide and compares the US and the EU, and within the EU Denmark, the Netherlands, France, Germany, Spain and Italy. The usage data covers one vendor, and the model itself estimates how long each task takes. So the figures show direction and size, never exact hours. The methodology, limits and full data are at the end.

What the data shows

1. AI goes to desk work.
   Desk tasks take 94% of AI use, and physical work almost none.
2. Countries differ in how much they use AI, far less in what they use it for.
   The mix of tasks is almost the same in all seven markets, while use by people and use in firms rank the countries differently.
3. AI use spreads across tasks of every length.
   Almost half goes to tasks under four hours, and success shows no pattern by task size.
4. Office work draws less AI use than its share of desk work suggests.
   Office and administrative support is two thirds desk work, and its tasks rarely reach AI.
5. Job ads fell from late 2022, and in the job types with the most AI use they were already falling behind before AI use spread.
   In Europe, employment grew in the occupations that use AI.
6. A large part of the economy is out of view.
   The self-employed and the smallest firms fall outside the statistics this paper combines. Physical work barely appears in conversation data. The share of each differs from market to market.

Desk work takes 94% of AI use

Across jobs, AI use lands on the central tasks more than on side tasks.

O*NET breaks every US occupation into the tasks it involves, about 18,800 in all.² breinz sorted the work behind these tasks into three kinds. Desk work means writing, analysing, planning, calculating and communicating. Physical work, such as building or treating a patient by hand, is done with the body or in person and involves no desk work. The third kind is physical or in-person work with a real share of desk work. Examples are the plan behind a repair or the report after a visit. The opening chart shows how AI use divides between the three. breinz’s catalogue of 608 tasks comes from other records, and it covers work that takes 98% of the AI use the Economic Index records.

O*NET also marks each task as core or supplemental for its occupation. A task is core when most people in the job rate it relevant and important.² Core tasks are three quarters of all tasks and take 85% of AI use.

A task central to a job can be support work for the business. Accounting is the core of an accountant’s job, but on a farm or in a factory it supports the business. The office jobs that do this support work draw less AI use than the average job, as a later section shows.

Countries use AI for the same work, in different amounts

The mix of tasks is almost the same in all seven markets. How much people and firms use AI differs, and the two measures rank the countries differently.

In the Economic Index, software and data work take about a fifth of Claude use in each of the seven markets. Office and administrative work take 7 to 8%.³ What changes from country to country is how much. Relative to their working-age population, people in France, the Netherlands and the US use Claude the most, and people in Italy the least.³ Among European firms with ten or more people, Denmark leads in the share using AI, and France comes fifth of the six countries.⁸

The two measures count different things. One counts conversations with one vendor’s AI. The other counts firms that use any AI technology at all. A country where many people use AI can have few firms using it, and France shows the widest such gap. Country figures also hide which languages people use AI in, as the Labs paper English is only half of AI use showed.

Almost half of AI use goes to tasks under four hours

The other half goes to longer tasks, and tasks of seven hours or more take a fifth.

The Economic Index estimates how long each task would take a person working without AI.³ breinz split the tasks into four groups by that time, under 2 hours, 2 to 4, 4 to 7, and 7 or more. It then measured each group’s share of AI use.

Success shows no pattern by task size. In every group, the task is completed in 64% to 75% of conversations.³ The AI True Cost Index measured how much of the promised time saving remains once failures, checking and rework are counted.

Office work draws less AI use than its share of desk work suggests

Jobs made mostly of desk work hold half of US employment and draw four fifths of AI use. Within them, jobs holding a fifth of all employment draw very little, and several of these are office and support work.

breinz set each US occupation’s share of AI use against its share of employment.^{3 4} An index of 1 means AI use in proportion to the number of people in the job. Software and data jobs reach about 6, and arts, design and media about 10. Two thirds of the tasks in office and administrative support are desk work, and its index is 0.64.

Seventy-four occupations made mostly of desk work draw AI use at less than a quarter of the average rate. Together they hold a fifth of US employment. Customer service representatives, accountants, general and operations managers, medical secretaries, HR specialists and insurance agents are among them. Several of these are support functions found in most businesses.

When a task from these jobs reaches an AI conversation, it succeeds about seven times in ten. That is the same rate as in the jobs that draw the most AI use. What differs is how often the tasks get there.

The data shows that these tasks rarely reach AI conversations and leaves the reasons open. Four explanations fit what is known. The work of these jobs runs inside customer records, accounting and ticketing systems, which a chat window does not reach. Employers seldom set out how to use AI. In the US, 49% of employees say their organisation has brought in AI, but only 26% say it has a clear plan for it.⁹ AI built into company systems, such as customer service chatbots, does not appear in conversation data at all. And AI adoption has only just reached the early majority, as the Labs paper Two years to first AI agent showed. Many people in these jobs may therefore not have started yet.

AI use in office work is likely to grow as more people start, gain experience and meet easier tools. The first users needed two years to get from their first chat to building AI agents. People starting now find agents already built into the products they use.

In the job types with the most AI use, job ads fell behind before AI use spread

Job ads fell from late 2022 while output grew, and in Europe employment grew in the occupations that use AI.

Indeed publishes job-ad data for six of the seven markets, all but Denmark. In late 2022, job ads in all six stood close to their highest level since 2020. In most of them, the hiring wave that followed the pandemic drove that level. From November 2022, the month ChatGPT launched, to autumn 2026, job ads fell by between 7% and 49%, depending on the market.⁵ Output grew in all six over the same period.⁷ By autumn 2026, job ads in most markets were at or above their level of early 2020. Where output grew fastest, in Spain, job ads fell least.

Job-ad data by job type is published only for the US, Germany and France. In these three markets, breinz compared the quarter of job types with the most AI use with the quarter with the least. Job ads for the high-use quarter fell further behind in all three.⁵ A gap in the same direction was already there between 2020 and 2022, before ChatGPT. breinz then took into account the growth of each job type’s own sector and that earlier trend. After that, the link between AI use and the fall in job ads remains only in Germany.

In the sector that uses AI most, job ads and employment moved apart. Output in information and communication grew faster than the economy in all three markets.⁷ In Germany and France, the number of ICT professionals grew several times faster than employment overall. Across Europe, the occupations that use AI added jobs, while those that barely use it stood still.⁶ Job ads for software jobs fell further than job ads overall.

This data cannot separate the effect of AI on job ads from two other causes. These are interest rates and the correction after the hiring boom of 2021 and 2022. It does contradict the claim that AI cuts employment in the jobs where it is used, because those jobs grew in Europe. The picture for young workers is different. A Stanford study finds lower employment among workers aged 22 to 25 in the US jobs most exposed to AI.¹⁰

Office work shrank as well. In the EU, the number of clerks fell by 2.7% from 2022 to 2025, while employment overall grew by 3.0%.⁶ AI continues a decline in office work that digitisation started.

A large part of the economy is out of view

The statistics this paper combines leave out the self-employed and the smallest firms, and conversation data barely shows physical work. The share of each differs from market to market, so the same data says more about some economies than about others.

Farming shows the gap most plainly. Desk work makes up a third of a farmer’s tasks, from planning crops to filing subsidy claims. In the EU, one farmer in three spends more than six days a year on that paperwork alone.¹¹ That third is where AI can save farmers time, because AI use goes almost entirely to desk work. Yet the US count of employment by occupation includes only 6,500 of almost 800,000 farmers. It leaves out the self-employed, who are two thirds of farmers, and most farm employers.^{4 12} Counted on that payroll alone, farm work draws more AI use per worker than most jobs. Counted in full, it draws almost none. The conversations show farm tasks, but they cannot show whether farmers themselves are the ones using AI. Food service, transport and construction draw almost no AI use either.

Each of these gaps comes from a different measure. The US count of employment by occupation leaves out the self-employed.⁴ The EU survey of AI in firms leaves out firms with fewer than ten people.^{8 13} Physical work barely appears in conversation data.³ Of the three, the share of jobs in firms with fewer than ten people differs most from market to market.

Italy has the lowest measured use of AI on both measures. It also has the largest shares of self-employed and of jobs in firms with fewer than ten people. Denmark and Germany have the smallest.

Where firms are counted, adoption is a yes or a no. Eurostat counts a firm as using AI if it uses any one AI technology.⁸ A large firm that runs one chatbot counts the same as a firm built around AI.

The EU figure for AI in firms says as much about the size of a country’s firms as about how much they use AI. Firms with 250 or more people use AI three times as often as firms with 10 to 49 people. Yet most of the firms the survey counts have 10 to 49 people.¹³ So a country with many smaller firms scores low even if its firms use AI as much as same-sized firms elsewhere.

The firm count also leaves out the self-employed. In some US desk jobs, such as writers and designers, one person in five or more is self-employed. Each worker in these jobs uses AI about one and a half times as much as in desk jobs where more than four in five people are employees.^{3 12}

What follows

AI is used on desk work. Whether that use raises productivity depends on what people do with the time it frees.

A study of Danish workers found no change yet in their earnings or hours from using AI. Most of those who saved time spent it on other tasks.¹⁴ So time saved raises productivity only through the work it is spent on, and that work differs from one job to the next.

On a farm, at a food producer or in an energy network, office work supports the core of the business. So time AI frees there can go back to the work the business exists for. Only data from inside these firms can show whether it does.

The statistics cover some jobs and economies far better than others, so judgements about AI are skewed towards them. For the self-employed, the smallest firms and physical sectors, the data shows only which way things are moving. Any claim about AI’s effect there, good or bad, runs ahead of the evidence. The Labs paper Two years to first AI agent found that most Europeans had not yet started using AI in 2025. The use measured here comes from those who have.

Key takeaways

1. Desk work takes almost all AI use.
   Physical work draws almost none.
2. AI is used on tasks of every length.
   Almost half of the use goes to tasks under four hours, and success shows no pattern by task size.
3. Office work draws less AI use than its share of desk work suggests.
   When the tasks of desk jobs with little AI use do reach AI, seven in ten are completed, as in other jobs. This use is likely to grow as more people start using AI.
4. Job ads fell from late 2022, and in Europe employment grew in the occupations that use AI.
   In the job types with the most AI use, job ads were already falling behind before AI use spread.
5. A country’s AI ranking depends on what is measured.
   Use by people and use in firms rank countries differently, and the statistics cover some economies far better than others.
6. Where the statistics do not reach, claims about AI run ahead of the evidence.
   The self-employed, the smallest firms and farms are the least measured.