Quick answer
No, not only programmers. In ADP payroll data, employment of software developers aged 22 to 25 fell about 28% further behind that of workers 35 and over than before late 2022. Customer service fell 22% further behind, the two most exposed groups 15% and 12%. Stock clerks show no change. Age is not seniority; no reason is shown.
Key takeaways
- Measured: a payroll employment index (Nov 2022 = 100), national, from one payroll processor's clients, by age band. No public dataset measures AI agents as such, and this one does not measure AI.
- A figure of -28% means employment of software developers aged 22 to 25 fell about 28% further behind that of developers 35 and over than it stood before late 2022. Customer service fell 22% further behind, so it is not unique to programming.
- The two groups of occupations scored as most exposed to AI in published research fell 15% and 12% further behind; the three least exposed groups show no such gap.
- Stock clerks show no change in the gap (0%) and home health aides moved the other way (+11%), but only three comparison occupations were checked.
- Age is not seniority, the data are national, the gaps are measured from a Nov 2022 base the source chose, and nothing here ties them to AI or agents.
Is AI only affecting entry-level programmers?
No: in published payroll data, the gap between workers aged 22 to 25 and older workers that opened after Nov 2022 is not unique to software developers. What was measured is an employment index (Nov 2022 = 100) built from one payroll processor's clients, national, by age band. No public dataset measures "AI agents" as such, and this one measures neither AI nor agents.
How to read the numbers: a figure of -28% means employment of software developers aged 22 to 25 fell about 28% further behind that of developers 35 and over than it stood before late 2022. The base month, Nov 2022, was chosen by the source's authors, and every figure is relative to it.
The gap is largest in software developers and nearly as large in customer service. It also shows in the two groups of occupations scored as most exposed to AI in published research, and it is absent or reversed in stock clerks, home health aides and the less exposed groups. That counts against the idea that this is only a programming story. It is not evidence about why it happens.
What did we measure, and how?
We compared how workers aged 22 to 25 fared against workers 35 and over, before and after a base date, in nine groups. The source is the Stanford Digital Economy Lab's "Canaries in the Coal Mine" dashboard, which uses ADP payroll records. We used the 17 September 2026 release: national, monthly, September 2021 to August 2026.
We started out assuming that only junior programming roles were affected, and the software direction (young workers down, older workers up) was already visible before we fixed our method, so that result was never a fair test. The other groups we had not seen. "Junior" is not something these files measure, since they report age bands. The "only programming" part can be checked: if the gap shows up outside software, it fails.
- The measure: each age band's index is set to 100 in Nov 2022, and every figure is relative to it.
- "Older": the average of the three older bands, 35-40, 41-49 and 50+. The 26-30 and 31-34 bands are not in the comparison.
- The comparison: the last 12 months (September 2025 to August 2026) against September 2021 to October 2022, for 22 to 25 versus older.
- Nine groups: software developers, customer service, stock clerks, home health aides, and five exposure groups, numbered 1 (least exposed) to 5 (most exposed). The dashboard says its exposure measures come from published research (it names Eloundou et al. 2024 and the Anthropic Economic Index); we did not check which one this file uses.
- Size: 54 series, each with 60 monthly readings. Months follow one another, so they are not 60 independent pieces of evidence.
Which jobs show the gap?
Employment of workers aged 22 to 25 fell furthest behind older workers in software developers (28%), then customer service (22%), then the two most exposed groups (15% and 12%), measured from Nov 2022. Stock clerks and the three least exposed groups show no such gap.
Plain meaning of the table: a negative figure means workers aged 22 to 25 fell that far behind the average of the three older bands (35-40, 41-49, 50+) compared with before; the base month, Nov 2022, was chosen by the source; age is not seniority, and this shows no cause. Last 12 months versus September 2021 to October 2022, national.
| Group | Change in the gap (percent equivalent) |
|---|---|
| Software developers | -28% |
| Customer service | -22% |
| Exposure group 5 (most exposed) | -15% |
| Exposure group 4 | -12% |
| Stock clerks | 0% |
| Exposure group 3 | 0% |
| Exposure group 1 (least exposed) | +3% |
| Exposure group 2 | +6% |
| Home health aides | +11% |
Before we looked, we set a gap of -0.10 log points (about -10%) as the margin for "present" (technical detail: log points are a way of measuring percent changes, and this margin is the only place we use them). Customer service and exposure groups 4 and 5 pass it, so the "only programming" part of the belief does not hold in this data. Keep in mind that the 22-25 band had risen slightly faster before Nov 2022, the base month the source chose. Software developers also sit well below the other three occupations: the difference is -24%, which counts the three equally, and one of them (home health aides) rose in the young band.
Is it a straight line from least to most exposed?
No. It is a threshold: only the two most exposed groups show a gap, and the three least exposed show none or a small positive figure. Exposure groups 1 and 2 are positive, group 3 is 0%, and groups 4 and 5 are negative.
A straight line fitted across the five groups gives -5.6% per step, but groups 4 and 5 carry it, and it describes the pattern poorly. Do not read it as "each step of exposure moves the gap by about this much".
What else moves these numbers?
Many things besides AI move employment for young workers, and this data cannot separate them. Interest rates, the post-pandemic correction in hiring, the 2022 US tax change on software R&D costs, remote work and the size of graduating classes all coincide with the period. For example, the Federal Reserve began raising its target interest rate in March 2022 and kept raising it into 2023, and under Section 174 of the tax code, software development costs paid in tax years beginning after December 31, 2021 must be capitalized and amortized rather than deducted at once. These are context, not an explanation of the gap.
Also, an age band changes as people age into and out of it, and the data cannot tell that apart from changes in employment.
What can this data not show?
It cannot show that AI or agents changed anyone's job, and it says nothing about your area.
- It is national payroll data from one processor's clients: not a count of workers, and nothing about California or any other place.
- Nothing ties any gap to AI or agents. The exposure groups are published occupational scores, not measures of AI use.
- The make-up of the exposure groups is unpublished, so software developers and the named occupations may sit inside them. The groups overlap, so the exposure results are not a separate confirmation of the occupation results.
- Only three comparison occupations were checked. No widening in stock clerks or home health aides does not show there is none elsewhere.
- In software developers the 50+ band sits below 41-49, which the average of the older bands hides.
- A different base month than Nov 2022 would shift every figure.
For the other half of the picture, see how many small businesses use AI, a Census survey of use rather than of jobs.
How can I check these figures?
Open the Stanford Digital Economy Lab dashboard and download the release files of 17 September 2026. The page is the Canaries in the Coal Mine dashboard. The files we read are the age by exposure, software developers and customer service results, plus matching files for stock clerks and home health aides on the same page. We publish no data file; see the source page.
The source's index values are 100 in Nov 2022 by construction. Our gap figures compare the young band with the average of the older bands, late period versus early period, and express the change as a percent. A later release may give different numbers.
What should I do with this?
Do not change who you hire on the strength of these figures; use them as a prompt to look at your own payroll. They describe one national payroll index, not your applicants, and they say nothing about what any individual can do. If you hire for customer service or software work, the only thing they change is the assumption that this pattern is limited to coders. They do not say why it happens.
To see what is true for you, split your own headcount in these roles by age band (22-25, 26-34, 35 and over) at the end of 2022 and today, and compare the change. Do it for each role separately, and remember that age is not seniority. For what such tools do today, read AI agents for small business.
Analysis by Deivy Hernandez, from public data files.
FAQ
Does this prove AI is taking entry-level jobs?
No. The data are a payroll index by age band. They do not measure AI use, so they cannot separate AI from other forces. Interest rates, the post-pandemic correction, the 2022 US change in how software R&D is expensed, remote work and the size of graduating classes also move the same series.
Should I stop hiring young people?
These data cannot tell you that. They describe one national payroll index by age band, not your applicants, and they say nothing about what any 22-to-25-year-old can do. They show no cause, and in several occupations they show no change in the gap at all.
Does this describe my city or California?
No. The index is national. It cannot say anything about one state or metro area, and it covers one payroll processor's clients, not every employer.
Are workers aged 22 to 25 the same as junior employees?
No. Age is not seniority. A 24-year-old can hold a senior role and a 45-year-old can be new to a field. The source reports age bands only, so this article does too.
Why are there no margins of error?
The source publishes none, so any we calculated would look more precise than the data are. We print the readings as published and say plainly what they cannot show.