AI-Exposed Jobs Rose 0.1% in July. The Gap Widened Anyway (34 Months)

Employment in America's most AI-exposed occupations grew 0.1% in July 2026 compared with a year earlier. After the June reading showed a 0.2% decline, that is the first positive all-ages number in this monthly series, and the headline ADP Research chose for it was "rose slightly."

Read one line further and the picture inverts. Employment in the least-exposed occupations grew 1.1% over the same year. The gap between the two groups did not close in July. It widened. And for workers aged 22 to 25 in high-exposure jobs, the count has now fallen for 34 consecutive months. [Fact]

This is the second monthly update in a series we started following with the June reading. The value of a monthly series is not any single month. It is what changes between them, and July changed three things at once, in three different directions.

What the July numbers say

The Canaries Dashboard is a joint project of the Stanford Digital Economy Lab and ADP Research. Each month it observes the payroll records of what ADP describes as millions of workers in more than 730 occupations at tens of thousands of private U.S. employers, links each occupation to an established measure of AI exposure, and compares employment trends in the most-exposed occupations against the least-exposed ones. The clock starts at the widespread release of generative AI in late 2022. [Fact]

The July 2026 reading, published August 19, 2026 under ADP chief economist Nela Richardson, reports the following. [Fact]

Across all ages, employment in high AI-exposure occupations was up 0.1% year over year, while employment in the least-exposed occupations was up 1.1%. Workers aged 22 to 25 saw overall employment contract 1.3% year over year. Within that age band, employment in high-exposure jobs fell 3%, the 34th straight monthly decline in a run that began in October 2023, while employment in the least-exposed jobs was flat. For workers aged 26 to 30, high-exposure employment fell 2.1% and least-exposed employment rose 1.8%.

That is the entire release. ADP's monthly commentary runs to a two-minute read and does not name individual occupations, sectors, or the exposure index it uses. Everything below that goes beyond those figures is our arithmetic or our interpretation, and is labeled as such.

June to July: three moves, three directions

Put the two months side by side and the "rose slightly" framing gets more complicated.

The all-ages gap between least-exposed and high-exposure employment growth was 0.8 percentage points in June (0.6% versus -0.2%). In July it was 1.0 percentage point (1.1% versus 0.1%). [Estimate — our subtraction of ADP's published growth rates.] High-exposure employment improved by 0.3 points month to month; least-exposed employment improved by 0.5 points. Both groups did better. The less-exposed group did better by more. So the first positive headline number in the series arrived alongside a wider penalty for being in an AI-exposed job, not a narrower one.

For 22-to-25-year-olds, the gap moved the other way. In June, high-exposure employment was down 4.3% against flat least-exposed employment, a gap of 4.3 points. In July the decline eased to 3% against flat, a gap of 3.0 points. [Estimate] That is a 1.3-point improvement in a single month, the largest move anywhere in the release.

For 26-to-30-year-olds, the gap widened. June showed high-exposure employment down 2.6% against least-exposed up nearly 0.9%, a gap of 3.5 points. July showed -2.1% against +1.8%, a gap of 3.9 points. [Estimate] The high-exposure decline eased by half a point, but least-exposed employment for this age group nearly doubled its growth rate, so the spread grew.

So in one month: the youngest band's penalty shrank, the next band's penalty grew, and the all-ages penalty grew. A single narrative does not fit all three. We think that is the honest reading, and it is worth resisting the pull to pick one.

Why the youngest band improved, and why to hold the applause

The 1.3-point easing for 22-to-25-year-olds is the number most likely to be read as a turn. There is a mechanical reason to be cautious about it.

These are year-over-year figures. The July 2026 number compares against July 2025, and by July 2025 this age band was already 22 months into its decline. A smaller year-over-year drop can mean the decline is slowing, or it can mean the comparison base has already fallen. The dashboard commentary does not separate the two, and neither can we from a two-paragraph release. [Claim] What the 34-month streak does establish is that the level has not recovered: 34 consecutive negative year-over-year readings means every July-to-July, June-to-June, and so on since October 2023 has come in below the prior year.

There is a second thing the July release adds that June did not. It reports that overall employment for 22-to-25-year-olds, across all exposure levels, fell 1.3%. [Fact] The June release gave no all-exposure figure for this age band. With it, the shape is clearer: this is not only a high-exposure problem for the youngest workers. Their least-exposed employment is flat, their high-exposure employment is down 3%, and the blended figure lands at -1.3%. The AI-exposed side is doing worse, but the whole cohort is treading water in a labor market where older workers are still adding jobs.

The 26-to-30 band is where the story got quieter and worse

The band that got less attention in June deserves more of it in July.

Workers aged 26 to 30 are not entry-level. They are typically three to eight years into a career, past the first rung. In June their high-exposure employment was down 2.6%; in July it was down 2.1%. That is improvement. But their least-exposed peers went from 0.9% growth to 1.8%. [Fact] The distance between the two career paths for the same age group is now 3.9 points, wider than the 3.0 points separating the same paths for 22-to-25-year-olds. [Estimate]

That ordering is new. In June the youngest band had the widest gap, which fit the "AI hits the door, not the building" reading we offered then. In July the widest gap sits one band up. One month is not a trend, and we will say so. But if it holds, the story shifts from "firms stopped hiring juniors into AI-exposed roles" toward "firms are also not retaining or advancing the people they hired into those roles two to five years ago." Those are different problems with different fixes, and the July data is the first month in this series to hint at the second one.

What this dataset cannot tell you

The same caveats we set out for June apply, and the July numbers add one.

The dashboard measures employment in occupations rated as AI-exposed. It does not observe whether any given employer deployed AI, what tool, or in which task. A firm that adopted a coding assistant and slowed junior hiring and a firm that over-hired in 2021 and 2022 and is still correcting look identical in this data. [Claim] The occupations at the top of most exposure indices, software, IT, business and administrative support, are also the occupations with the most extreme pandemic-era hiring swing, and the 34-month clock started roughly a year after the tech-sector correction began. Two stories, one payroll series.

The sample is one payroll processor's client base. ADP is large, but firms that outsource payroll to ADP are not a random draw of U.S. employers by size or sector. The directions here are well supported. The precise percentages are not population estimates. [Claim]

And the new caveat: the all-ages high-exposure figure moved from -0.2% to +0.1%. That is a 0.3-point swing in a series measured in tenths. We do not know the dashboard's month-to-month noise floor, because ADP does not publish confidence intervals in the monthly note. A move of that size could be signal, revision, or seasonal composition. Treating it as "AI-exposed employment is now growing" is a stronger claim than the data can carry. [Claim]

We should also be direct about our own instrument. Our occupation pages assign each occupation an exposure profile drawn from a single class of task-based indices. The Canaries Dashboard uses the same family of measures. If those indices systematically mislabel which occupations are actually being changed by AI, both the dashboard and our pages inherit the error together. That is not a reason to discard either. It is a reason not to treat agreement between them as independent confirmation.

Which occupations sit where

ADP's monthly notes do not name occupations, and the July note is no exception. The June commentary from Richardson pointed to software development and customer service as examples at the high-exposure end and home health care at the low-exposure end, and nothing in July changes that grouping. [Fact]

For readers in those roles, we keep the underlying exposure and task profiles on the occupation pages for software developers, customer service representatives, and home health aides. The mapping from ADP's exposure bands to those specific pages is ours, not ADP's. [Claim]

What to do with a monthly series

If you are 22 to 25 and heading into an AI-exposed field, July does not change the June advice. The bottleneck is entry, not the occupation. High-exposure employment across all ages is flat to slightly up; the shortage is in the first rung. Aim at entry roles whose task bundle includes something a model cannot do unsupervised: client contact, accountability for a decision, a regulated or physical environment, a license.

If you are 26 to 30 and in an AI-exposed role, July is the first month in this series that speaks to you specifically, and it is not reassuring. Your least-exposed peers are pulling ahead faster than your own cohort is recovering. The practical response is the same one that has always protected mid-career workers in a shifting field: make sure the part of your job that is visible to your employer is the part a model does not do. If the visible part of your work is producing the artifact, code, copy, tickets, reports, the exposure is real. If it is owning the outcome the artifact serves, it is much less so.

And for everyone reading monthly AI-and-jobs headlines: the July release contains a positive all-ages number, a widening all-ages gap, a shrinking youngest-band gap, and a widening next-band gap, all at once. Any article that reports only one of those four is telling you which conclusion it started with.

What August would need to show

Two months give us a direction of travel for each band. A third will tell us whether July was noise.

If the 22-to-25 gap keeps narrowing while the streak continues, the base-effect explanation gains weight: the decline is aging out of the comparison window without the level recovering. If the 26-to-30 gap keeps widening, the retention story moves from hint to pattern. If the all-ages high-exposure figure goes back below zero, the July "rise" was within the noise. We will check the August note when ADP publishes it and update this series either way.

Sources

• ADP Research, Nela Richardson, "Canaries Dashboard: Employment in AI-exposed occupations rose slightly in July," published August 19, 2026 — https://www.adpresearch.com/research/canaries-dashboard-employment-in-ai-exposed-occupations-rose-slightly-in-july
• ADP Research, Nela Richardson, "Canaries Dashboard: Employment in AI-exposed occupations contracted in June," published July 22, 2026 — June comparison figures — https://www.adpresearch.com/research/canaries-dashboard-2026-june
• Our June analysis: 33 Straight Months: AI-Exposed Entry-Level Jobs Keep Shrinking

About this analysis

This article was written with AI assistance. Every percentage attributed to ADP is quoted from its published monthly commentary; the underlying dashboard is hosted by the Stanford Digital Economy Lab and we did not access its raw data for this article. The percentage-point gaps and month-to-month changes are our own subtraction of ADP's published figures and are labeled [Estimate]. The base-effect, retention, and noise-floor caveats, and the mapping to our occupation pages, are our interpretation, not ADP's. ADP Research material is quoted briefly under its copyright; no figures or charts are reproduced.