AI and jobs: what AI agents really change, according to the data

So far, AI agents are changing the content of jobs task by task far more than they are wiping out entire occupations: that is what the main public statistical sources say today. The alarming figures that go viral usually measure how “exposed” tasks are to AI, not jobs actually lost, and official bodies themselves stress how uncertain their estimates are.

Replace, expose, transform: what are we talking about?

When people discuss AI and jobs, three different ideas tend to get mixed up.

• Replace: an entire occupation disappears because a machine does all the work.
• Expose: part of the tasks in an occupation can be done or assisted by AI.
• Transform: the job stays, but the way it is done changes, because some tasks are delegated and others take on more importance.

Most serious studies measure exposure, not replacement. A highly exposed occupation can keep hiring: that is what happens when AI speeds up part of the work without removing the need for a person to decide, check and take responsibility.

AI agents add a nuance. An assistant like ChatGPT answers a request. An agent can chain several steps on its own: look something up, fill in a form, send a message, update a spreadsheet. It takes automation one step further, but still task by task. If the difference feels fuzzy, our explainer on what an AI agent is covers the basics.

What the official data says

Three open, dated public sources give a reasonable picture. They do not measure the same thing, which partly explains why headlines disagree.

The ILO: one worker in four exposed, mostly in clerical tasks

On 20 May 2025, the International Labour Organization (ILO) published a global index of occupational exposure to generative AI (Working Paper 140, accessed on 22 October 2026 on ilo.org). Its main findings:

• globally, one worker in four is in an occupation with some degree of exposure to generative AI;
• 3.3% of global employment falls into the highest exposure category, with a clear gap between female employment (4.7%) and male employment (2.4%);
• exposure rises with country income: 11% of total employment in low-income countries, 34% in high-income countries.

Above all, the ILO concludes that because most occupations consist of tasks requiring human input, job transformation is the most likely impact of generative AI. It adds that clerical occupations continue to have the highest exposure.

The US Bureau of Labor Statistics: cautious projections, job by job

In February 2025, the US Bureau of Labor Statistics (BLS) explained how it factors AI into its 2023-2033 employment projections (Monthly Labor Review article, accessed on 22 October 2026). Across all occupations, it projects employment growth of 4.0% over the decade. Among the occupations covered in that article:

• software developers: +17.9%;
• personal financial advisors: +17.1%;
• paralegals and legal assistants: +1.2%;
• claims adjusters: -4.4%;
• customer service representatives: -5.0%.

The BLS does not present these figures as the effect of AI alone: they are projections that combine several factors. Above all, it warns that “the timing and scale of many potential impacts of GenAI are too uncertain to be reflected” in its projections, and that its methods are not designed to capture extremely rapid technological change.

Eurostat: one European business in five uses AI

In Europe, Eurostat measures adoption rather than the effect on jobs. According to its data, as recorded on 22 October 2026, 19.95% of EU businesses with 10 or more employees used at least one AI technology in 2025, up from 7.65% in 2021. The figure was 25.97% in Germany, 20.27% in Spain and 18.16% in France.

Eurostat’s page on AI use in enterprises, accessed on 22 October 2026, adds two useful points: 55.03% of large enterprises use AI, against 17% of small ones; and among businesses that considered AI without adopting it, 70.89% cite a lack of relevant expertise as a barrier.

Source	What it measures	Key figure	What it does not tell you
ILO, May 2025	Occupational exposure to generative AI	1 worker in 4 exposed; 3.3% highly exposed	How many jobs will disappear
BLS, February 2025	US employment projections 2023-2033	+4.0% all occupations; -5.0% customer service reps	The exact share due to AI
Eurostat, 2025 data	AI adoption by EU businesses	19.95% of businesses with 10+ employees	The effect on headcount

Why “exposed” does not mean “replaced”

A job is a bundle of tasks. Take a customer service representative: answering simple questions, finding a file, drafting an email, calming an angry customer, deciding on a goodwill gesture. The first three tasks are highly exposed to AI. The last two much less so, because they require judgment, empathy and accountability.

That is exactly what we describe in our article on handing customer service to an AI agent: the agent absorbs part of the requests, but complex cases and complaints go back to a person.

Three factors also slow the move from exposure to replacement:

1. Reliability. A model can produce a hallucination, that is, state something false with total confidence. On any task that commits the business, someone has to check.
2. Accountability. A decision affecting a customer, a patient or an employee must be explained and owned by a person.
3. Integration costs. Connecting an agent to a company’s tools, securing data and training staff all take time. The main barrier reported by Eurostat, a lack of expertise, makes that clear.

And AI agents: what do they change on top of that?

Assistants have mainly sped up writing, summarizing and research. Agents aim at something else: carrying out a chain of actions inside software. In practice, they can sort incoming emails, prepare a quote from a web form, chase unpaid invoices or compile a weekly report.

That reaches coordination and data-entry tasks that AI had barely touched until now. But agents are still limited: they make more mistakes on long tasks, they can be hijacked by instructions hidden in a web page, and they need clear rules. Companies that use them well generally keep a human in the loop who approves important actions. Our look at AI agent use cases in business details what already works and what still breaks.

What we do not know yet

It is worth being honest: a lot remains uncertain.

• The pace. Model capabilities are improving fast, while employment statistics arrive one or two years later. The figures available in October 2026 mostly describe 2025.
• Agents themselves. We do not yet know whether agents will become reliable enough to run whole processes on their own. None of the official statistics we consulted measures their effect on jobs separately.
• New occupations. Every technology wave has created some, but they cannot be counted in advance.
• Differences between countries. National studies are hard to compare, because they use different definitions and methods.

How to read an alarming AI jobs statistic

Before sharing a headline like “AI threatens X% of jobs”, ask yourself four questions:

1. Who produced it? A public statistics office, a consulting firm and a software vendor do not have the same interests.
2. What does it actually measure? Task exposure, business intentions, a projection, or jobs actually lost?
3. Over what period and where? A global figure says nothing about your local job market.
4. What caveats do the authors give? Serious studies always explain their limits.

Check yourself A news story claims that “34% of jobs will disappear because of AI”, citing the ILO. What do you make of it?

The ILO’s 34% figure is the share of employment exposed to generative AI in high-income countries, not jobs that will disappear. The ILO actually concludes that job transformation is the most likely impact. The story confuses exposure with replacement.

What you can do

For employees and managers alike, the best answer to uncertainty is practical: learn the tools on real tasks. Our hands-on guide to AI at work shows how to save time without taking risks, and our method to train your team on AI helps you launch a collective effort. To find out whether a specific task can be handed to an agent, try the Can I hand this to an agent? tool. Finally, our AI in numbers page gathers dated public data on AI adoption.