AI agents and jobs: what the data actually shows

AI agents can contribute to job losses, but the evidence available as of October 24, 2026, does not show work disappearing across the economy. It shows productivity gains in some settings, changes in how tasks are organized, and signs of weaker hiring in certain areas. Understanding the relationship between AI and employment requires separating what a system could automate, what a business actually automates, and what happens to its workers.

An exposed task is not an eliminated job

An occupation combines different activities: producing a document, checking information, negotiating an exception, and explaining a decision. An AI system’s ability to perform part of that work does not mean it can take over the whole job, including its constraints and responsibilities.

In its global index published on May 20, 2025, the International Labour Organization estimates that one in four workers holds an occupation with some exposure to generative AI. The highest exposure category accounts for 3.3% of global employment. The ILO emphasizes that jobs are more likely to be transformed than replaced entirely.

These figures describe technical potential assessed at the task level. They measure neither observed layoffs nor the share of jobs that will disappear by a particular date.

Indicator	What it measures	What it does not establish
Exposure	Overlap between occupational tasks and AI capabilities	Whether a business will use the technology
Adoption	Actual use within an organization	Whether that use delivers a net benefit
Productivity	Output produced with a given set of resources	Whether staffing will fall by the same proportion
Employment	Filled positions, hiring, and departures	Whether every change is caused by AI

The definition of an AI agent adds another distinction: it can carry out a sequence of actions using tools. But a measure of exposure to text generation does not automatically measure an agent’s ability to execute an entire business process.

Productivity can rise without a proportional drop in staffing

The study Generative AI at Work, published in February 2025, examines the gradual introduction of a conversational assistant among 5,172 customer support workers. Productivity, measured as issues resolved per hour, rises by an average of 15%.

The results vary by worker. Less experienced employees and those whose initial performance was lower benefit more from the assistance. The most experienced workers see small speed improvements and small declines in quality. This finding concerns an assistance tool in a specific setting, rather than the autonomous replacement of an entire department.

A productivity gain is not a job elimination rate. A business can use the freed capacity to shorten a backlog, extend service hours, handle more requests, or reduce its need to hire.

It can also decide to cut staffing. The outcome depends on factors including demand, integration costs, and required quality standards. It does not follow automatically from technical performance.

Consider a hypothetical document processing workflow. If extracting information becomes faster, but every file still requires a thorough review, the time saved on extraction is not the same as the time saved on the complete file.

Our article on AI agents for customer service helps define the scope of this assessment. The relevant unit is a service delivered correctly, rather than the number of responses generated.

Employment studies show a mixed picture

In Denmark, Anders Humlum and Emilie Vestergaard link adoption surveys with administrative records. Their working paper revised in March 2026 finds no statistically significant average effect on earnings or recorded hours. Their estimates rule out effects larger than 2% during the study window, two years after the chatbots concerned became publicly available.

They do, however, observe changes in work organization and new tasks related to AI. The absence of an average effect on those indicators therefore does not mean nothing has changed, nor does it provide a guarantee for subsequent years.

The picture differs in the United States. In the August 12, 2026, revision of their study on young workers, Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen analyze payroll data through June 2026. In exposed occupations, employment among people aged 22 to 25 is 19% below the path it would have followed had it moved in line with employment among their less exposed peers.

The gap arises mainly through reduced hiring. The authors present the findings as early descriptive indicators, without establishing a causal estimate. They specifically flag preexisting trends and differences between statistical sources.

These studies cover different countries, populations, and indicators. Their results can therefore coexist: a relatively stable average can conceal worsening access to particular jobs.

For a young applicant, a hire that never happens has a concrete impact, even without a visible wave of layoffs.

New studies require conclusions to be reassessed

A more recent Danish study adds another perspective. The working paper published by Danmarks Nationalbank on September 16, 2026 compares business adoption data with monthly employment records.

The authors estimate that, toward the end of 2025, firms that adopted AI in 2023 show employment growth 11% lower relative to firms that did not adopt it. The effects are concentrated in small businesses and occur mainly through fewer hires. According to the study, employment across the economy remains relatively little affected up to that point.

This figure does not mean those firms eliminated 11% of their jobs. It concerns a difference in employment growth between groups. It also does not isolate the specific effect of autonomous agents.

The period in which tools were studied matters as well. A randomized METR experiment published in July 2025 found that tasks took 19% longer with AI assistance. It involved 16 experienced developers completing 246 tasks in projects they knew well.

However, in its February 24, 2026, update, METR states that newer tools probably provide greater speed improvements. Selection biases prevent a reliable measurement of the size of that benefit.

Citing only the earlier slowdown to describe tools available in October 2026 would therefore be misleading. A study remains informative when its setting, period, and limitations stay attached to its findings.

How agents change the analysis

Much of the evidence above concerns generative AI broadly or AI assistants. Applying it directly to agents would skip a necessary step: their employment effects need to be measured within the organizations deploying them.

Our article on AI agents versus chatbots explains the functional difference. For employment, the key change is the move from occasional assistance to handling a sequence of steps: reviewing a file, preparing a response, updating software, and flagging exceptions.

That broader scope can make more work technically possible to automate. It also creates oversight needs: permissions, consistency across software systems, error handling, and human intervention when a situation falls outside the expected boundaries.

A hypothetical billing workflow illustrates the distinction. An agent prepares payment reminders from recorded payment data, while a person handles disputes, missing information, and commercial arrangements. The organization can reduce administrative time without eliminating the entire role.

The net result depends on both the work that remains and the new work created. Supervision can absorb part of the savings, while reliable integration may allow the same team, or a smaller one, to handle more cases.

The principle of a human in the loop therefore needs a concrete description: who checks the work, when they check it, and what power they have to intervene. An approval button alone does not establish that oversight is effective.

The guide to deploying AI agents in business helps frame these choices before attaching a staffing target to them.

Who bears the risk, and who benefits from the gains?

The debate extends beyond the total number of jobs. It also concerns access to initial assignments, career progression, earnings, and how work is divided between employees and outside contractors.

The US findings on young workers warrant particular attention to entry routes into occupations. If a business automates straightforward tasks, it also needs to consider how beginners will gain the experience required to handle complex situations.

This does not establish that all junior positions are at risk. It identifies an organizational risk worth examining: removing learning tasks can weaken the development of future professionals, even when experienced staffing levels remain stable.

The same reasoning applies to outsourced work. Within a particular business, a decline in assignments given to freelancers could precede a change in employee headcount. Assessing your own situation therefore means looking at assignment volumes as well as job postings.

For workers, learning to use a tool can help, but it does not provide insurance against job cuts. Domain knowledge, the ability to verify results, and the ability to explain a decision remain useful criteria for planning training.

For employers, the question also concerns how gains will be used: additional capacity, shorter turnaround times, better quality, or lower staffing costs. Those objectives should be explicit.

The AI tools organized by occupation can help identify possible uses. Their availability is neither evidence of profitability nor a forecast of employment.

Measuring a deployment without confusing speed with savings

For a small or medium-sized business, the most useful approach is to observe a real process before expanding deployment. The guide to evaluating AI agents recommends starting with representative tasks and explicit success criteria.

Begin by documenting the current process: time required, quality achieved, pending cases, and necessary interventions. Then compare tasks of similar difficulty, counting all the human work surrounding the agent.

The indicators serve complementary purposes:

• Total time: preparation, execution, verification, and corrections.
• Final quality: accuracy, accepted cases, and complaints.
• Shifted workload: additional work created for colleagues or customers.
• Total costs: software, integration, training, and maintenance.
• Changes in work: altered responsibilities, hiring, departures, and outside contracting.

Avoid immediately converting saved time into positions that could be eliminated. Minutes scattered across several roles do not necessarily add up to an entire position that can be removed.

Also examine how long the gain lasts. An encouraging initial result may depend on a narrow scope, clean data, or technical support that is difficult to sustain at a larger scale.

Finally, distinguish the finding from the decision. Measuring time savings is an operational result. Choosing to reduce staffing, provide training, or expand service is a business strategy decision.

Key takeaways

• An occupation’s exposure to AI measures technical potential; it does not directly predict job losses.
• Productivity gains vary with the tasks, workers, and tools studied.
• Recent evidence calls for monitoring hiring, particularly young people’s entry into certain exposed occupations.
• The specific effects of agents need to be assessed across complete processes, including oversight costs and the consequences for work.

Frequently asked questions

Will AI agents eliminate my job?

No study can predict what will happen to an individual job based on its title alone. Look at the tasks your employer is actually delegating, its hiring patterns, and the skills needed to check the work these systems produce.

Which jobs are most exposed to AI?

The ILO index published in May 2025 places clerical occupations among the most exposed. Exposure describes the technical potential to handle certain tasks, which varies across organizations; it is not a timetable for occupations to disappear.

Should I change careers because of AI?

A career change calls for evidence specific to your industry: available openings, changing responsibilities, and access to training. Learning to verify automated outputs can be useful, but training in a particular tool guarantees neither a job nor higher pay.

How can I tell whether AI actually improves productivity?

Compare work of similar difficulty against the same quality standards. Include preparation, checks, and corrections, then assess whether the time saved actually leads to more output or better service.