33 Straight Months: AI-Exposed Entry-Level Jobs Keep Shrinking

Employment in America's most AI-exposed occupations has fallen for 33 straight months among workers aged 22 to 25. Not slowed — fallen, every single month since October 2023.

That number comes from payroll records, not a survey and not a forecast. And it sits next to a second number that complicates the panic: across workers of all ages, employment in those same high-exposure occupations was down only 0.2% year over year. So which is it, a collapse or a rounding error? The answer depends almost entirely on how old you are.

What the payroll records actually show

ADP Research's Canaries Dashboard, built with the Stanford Digital Economy Lab, tracks month-to-month employment using the payroll files of tens of thousands of private U.S. employers, covering roughly 4 million workers across more than 730 occupations. [Fact] Each occupation is matched to an established measure of AI exposure, and the dashboard compares hiring trends in the most-exposed occupations against the least-exposed ones, with the clock starting at the wide release of generative AI in late 2022.

The June 2026 reading, published on July 22, 2026, breaks down like this.

Employment in high AI-exposure occupations contracted 0.2% compared with a year earlier, while employment in the least-exposed occupations grew 0.6%. [Fact] For workers aged 22 to 25, employment in high-exposure jobs fell 4.3% year over year — the 33rd consecutive monthly decline — while employment for that same age group in the least-exposed jobs was flat. [Fact] For workers aged 26 to 30, high-exposure employment fell 2.6%, against growth of nearly 0.9% in the least-exposed roles. [Fact]

The gradient is the finding, not the headline

Line those three pairs up and something clean falls out of them.

The gap between low-exposure and high-exposure employment growth is 4.3 percentage points for 22-to-25-year-olds, 3.5 percentage points for 26-to-30-year-olds, and 0.8 percentage points across the workforce as a whole. [Estimate — our own subtraction of the growth rates ADP reports for each age band.] The penalty for being in an AI-exposed occupation shrinks by more than 80% between your early twenties and the all-ages average.

That shape matters more than any single figure in the release. A technology that was genuinely substituting for human labor inside a job would push employment down for everyone doing that job — the 45-year-old customer service rep and the 23-year-old alike. This pattern does something much narrower. It hits the door, not the building. Incumbents are largely holding their positions. The people who would have been hired into those positions are not being hired.

There is a quiet corollary here that nobody enjoys stating. If entry-level intake stays suppressed for another few years, the shortage does not show up in 2026 unemployment statistics. It shows up around 2032, when the mid-level bench that normally gets promoted from those entry roles is simply not there.

Which jobs sit on which side of the line

The June dashboard commentary does not name individual occupations. ADP chief economist Nela Richardson's related June 2026 note does, pointing to software development and customer service as examples at the high-exposure end and home health care as an example at the low-exposure end. [Fact]

That grouping cuts against the older automation story. A decade ago the standard prediction put physical and routine manual work first in line. What the payroll data shows is close to an inversion of it. The roles taking the hit are desk roles that produce text, code, and structured answers. The job that requires being in a room with a human body kept growing.

If you want the exposure and automation figures for these specific roles, we maintain them here: software developers, customer service representatives, and home health aides.

The part that is not proven

Here is the objection that deserves more weight than it usually gets.

The Canaries Dashboard measures employment in occupations rated as AI-exposed. It does not measure whether the employer in question deployed any AI at all. Nothing in the design separates "this firm adopted an AI tool and stopped hiring juniors" from "this firm over-hired juniors in 2021 and 2022, then corrected." [Claim]

And the occupations that sit at the top of most AI-exposure indices — software, IT, business support — are precisely the occupations that saw the most extreme pandemic-era hiring boom and the most extreme post-boom correction. Tech-sector contraction began in earnest in late 2022 and early 2023. The 33-month clock started in October 2023. Those two stories predict nearly identical payroll data.

ADP's own framing is careful about this. It describes AI as acting "from the bottom up, at the task level," and notes that the aggregate effect on jobs remains modest. The dashboard is an early-warning instrument, not a causal estimate, and reading it as a causal estimate is the most common error being made with this dataset right now.

A second limit is worth naming plainly: this is one payroll provider's client base. Large ADP clients are not a random sample of American employers, and firms that outsource payroll skew toward particular sizes and sectors. The direction of the finding is well supported. The precise magnitudes are not the population truth.

What this means if you are early in your career

The practical read is different from the panicked read.

If you are 22 to 25 and aiming at an AI-exposed occupation, the bottleneck in front of you is entry, not the occupation. Employment among people already inside those jobs is roughly stable. What is scarce is the first rung.

That argues for a specific targeting strategy rather than an abandonment of the field. Aim at the entry roles whose task bundle still contains something a model cannot do unsupervised — work that requires sitting with a client, carrying responsibility for a decision, or operating in a physical or regulated environment. Inside software, in practice, that has meant fewer pure implementation openings and more openings attached to a system somebody has to be accountable for. Where licensing exists, get licensed. A license is an artificial floor under an entry-level wage, and a model does not hold one.

And check the age gradient before acting on any headline about AI and jobs. A 0.2% decline across an entire workforce and a 4.3% decline inside one age band are the same dataset describing two very different worlds. Most coverage picks whichever of those two numbers matches the argument it already wanted to make.

What would settle this

There is a clean test available, and it is worth knowing what to watch for.

If AI adoption is the driver, the entry-level gap should keep widening in the occupations where deployment runs deepest, even as the broader hiring cycle recovers. If the 2021-22 over-hiring correction is the driver, the gap should close on its own as that correction runs out, regardless of how much AI those same firms buy. Two more years of this dashboard will separate the two. Until then, anyone telling you the 4.3% figure proves AI is taking entry-level jobs is telling you what they believe, not what this data has established. The honest position is that the warning light is on and the wiring has not been traced.

Sources

• ADP Research, "Canaries Dashboard: Employment in AI-exposed occupations contracted in June," published July 22, 2026 — https://www.adpresearch.com/research/canaries-dashboard-2026-june
• ADP Research, Nela Richardson, "Taking the guesswork out of AI prognostications," Main Street Macro, published June 16, 2026 — dashboard methodology, sample coverage, and named occupation examples — https://www.adpresearch.com/main-street-macro/taking-the-guesswork-out-of-ai-prognostications

About this analysis

This article was written with AI assistance. Every percentage attributed to ADP is taken directly from its published dashboard commentary. The percentage-point gaps by age band are our own arithmetic on those published figures and are labeled [Estimate]. The causal caveats and the career interpretation are ours, not ADP's.