Agentic AI and Jobs: The Entry-Level Effect

Overview

The clearest evidence on how AI is affecting employment does not come from surveys or forecasts. It comes from payroll records, and it identifies a specific group bearing the effect.

Research from the Stanford Digital Economy Lab, using high-frequency payroll data covering 4.6 million workers across more than 730 occupations, found that workers aged 22 to 25 in the occupations most exposed to generative AI experienced a relative employment decline of around 13%.

More experienced workers in the same occupations were largely unaffected, and in some cases saw their employment improve. The effect is concentrated by age within the same jobs, not spread across them.

That pattern sits awkwardly beside deployment data showing only 17% of organisations have actually deployed AI agents, with over 40% of agentic projects expected to be cancelled by the end of 2027. Both findings are credible, and reconciling them explains what is actually happening.

A 13% Relative Decline for 22-25s

The headline finding concerns a narrow demographic band. Workers aged 22 to 25 — those in their first years of employment — saw employment in AI-exposed occupations fall by around 13% relative to comparable roles after generative AI spread.

The word relative matters. This is not a 13% fall in absolute headcount but a decline measured against what employment in those roles would otherwise have been, isolating the effect from general labour market conditions.

Published figures for the effect vary somewhat between analyses and time periods, with some reporting around 16%. The direction and the concentration by age are consistent across versions.

Built on 4.6 Million Payroll Records

The methodology is what makes this finding stronger than most claims in this area. The researchers used high-frequency payroll records from one of the largest US payroll processors, covering 4.6 million workers across more than 730 occupations.

Payroll data records what actually happened. It is not affected by how respondents perceive AI, whether they attribute changes to it, or what they expect to happen — all of which contaminate survey-based evidence on this subject.

The breadth across occupations is equally important, because it allows comparison between roles with high and low AI exposure within the same labour market and time period. Without that comparison, any employment change could be attributed to the economy rather than to AI.

Experienced Workers Were Unaffected

The most analytically significant result is the contrast within occupations. More experienced workers in the same AI-exposed roles were largely unaffected, and some saw employment improve.

This rules out the simplest explanation. If AI were displacing entire occupations, employment would fall across all experience levels in those jobs. Instead the same occupation shows declining employment for the youngest workers and stable or improving employment for everyone else.

The interpretation offered is that firms are using AI to substitute for the junior tasks through which people traditionally gain their first foothold — and that experienced workers, whose value lies in judgement and context rather than in executing those tasks, become more productive with AI rather than replaceable by it.

Automation Versus Augmentation

The declines are concentrated in occupations where AI is more likely to automate human labour rather than augment it, and that distinction predicts the outcome better than exposure alone.

Augmentation means the technology makes a worker more effective — handling routine parts of a task while a person applies judgement to the rest. Employment holds or rises because each worker produces more.

Automation means the technology performs the task, and the employment effect is negative. The same tool can do either depending on the job it is applied to, which is why exposure to AI is not sufficient to predict what happens to employment — the nature of the work determines it.

Which Occupations Are Affected

The affected roles identified include software developers, customer service representatives, computer programmers and information systems managers — occupations where generative AI is used extensively.

Software development is the most striking inclusion. It has been among the most reliably well-paid and in-demand career paths for two decades, and it is precisely the field where AI code generation has advanced fastest.

Customer service fits a longer-established pattern of automation, but the mechanism has changed. Previous automation handled scripted enquiries and escalated anything unusual; generative systems handle a substantially wider range of interactions, which extends the automatable portion of the role considerably further.

Why Entry-Level Is the Vulnerable Tier

The concentration of the effect among the youngest workers has a straightforward explanation rooted in how work is distributed within organisations.

Junior roles are assembled from the most codifiable tasks — writing routine code, handling standard queries, producing first drafts, compiling reports. These are delegated to newcomers precisely because they are well-defined and low-risk, which is also what makes them the most automatable.

Senior work is the opposite: deciding what should be done, judging whether an answer is right, managing relationships, taking responsibility for outcomes. Those tasks resist automation for the same reasons they resist delegation to a new graduate.

The Training Pipeline Problem

The long-term consequence is more serious than the immediate employment numbers, and it is slow enough to be easy to ignore.

Junior tasks were never valuable only for their output. They were how people developed the judgement that senior roles require — writing enough routine code to recognise what good code looks like, handling enough ordinary cases to recognise an unusual one.

Automating that tier removes the training pathway that produces experienced workers. The effect is invisible while experienced staff remain plentiful and appears years later, when the cohort that would have been promoted was never hired.

The same pattern is visible elsewhere. In the games industry, three quarters of students report concern about their prospects, citing a lack of entry-level roles specifically — an independent observation of the same mechanism from a different sector.

But Only 17% Have Deployed Agents

Against these findings sits deployment data suggesting agentic AI is far from widespread. Only about 17% of organisations have deployed AI agents, and more than 40% of agentic projects are expected to be cancelled by the end of 2027 over costs, unclear value or inadequate risk controls.

Market and adoption claims in this sector should be treated with corresponding caution. Assessments of the agentic vendor landscape have found only a small fraction of self-described agentic AI companies offer genuinely agentic capability, with the remainder rebranding existing assistants and automation.

So the picture is not one of autonomous agents running enterprises. It is limited, uneven deployment alongside a measurable employment effect on one specific group.

Why Both Facts Can Be True

Reconciling low deployment with a real employment effect is straightforward once you separate agents from generative AI more broadly.

The employment research concerns exposure to generative AI, not deployment of autonomous agents. A firm does not need an agentic platform to hire fewer junior developers — engineers using code generation tools, or a support team using AI drafting, changes hiring requirements without any agent being deployed.

Hiring decisions also respond faster than deployment statistics. A manager who believes a team can handle more work with AI assistance can slow hiring immediately, long before any formal project completes or appears in an adoption survey.

Which suggests the labour effect is running ahead of the technology's actual capability, driven by expectations as much as by demonstrated productivity — and that a substantial share of agentic projects being cancelled will not restore the entry-level roles that were not filled in the meantime.

Conclusion

The strongest evidence on AI and employment comes from payroll records rather than forecasts. Covering 4.6 million workers across more than 730 occupations, it shows workers aged 22 to 25 in AI-exposed roles experiencing a relative employment decline of around 13%, while more experienced workers in the same occupations were largely unaffected.

That contrast within occupations is the finding that matters. It indicates firms substituting AI for the junior tasks through which people traditionally enter a profession, while the same technology makes experienced workers more productive rather than redundant.

It is not happening because autonomous agents have taken over. Only 17% of organisations have deployed agents and over 40% of agentic projects are expected to be cancelled by 2027. Hiring decisions respond to expectations faster than deployment does, and a role not filled in 2025 is not restored when a project is cancelled in 2027.

The durable cost is to the pipeline. Junior work is how judgement is acquired, and an industry that automates its entry tier will discover the consequence about a decade later, when the people who would have become its seniors were never hired.

Data Source and Attribution

Stanford Digital Economy LabCanaries DashboardTIME

Employment findings come from 'Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence', a working paper from the Stanford Digital Economy Lab by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, using high-frequency payroll records covering 4.6 million workers across more than 730 occupations. Published estimates of the employment effect vary between roughly 13% and 16% depending on methodology and period, and this variation is noted in the text. Deployment and project cancellation figures come from Gartner research on agentic AI adoption.

FactsFigs reviews, cleans, and cross-checks every source dataset before shaping it into a data story. Each visualization is created and designed in FactsFigs Design Studio — an internal tool developed and owned by FactsFigs — and is the original work of a FactsFigs author, not an AI-generated copy of any existing graphic. Individual assets within a visual may or may not be produced with AI tools, but the design of the visual itself is solely FactsFigs' own.

Figures reflect research available at the time of publication. This is a working paper and findings may be revised. Nothing here is career or financial advice.

2026-07-20