Quick answer
Neither confirmed nor refuted. Averages fell in California metros in the published estimates (selling -6.6%, software -11.0%). Selling's change was 4.9% higher than for four software occupations, 12.1% lower than for developers alone. By job totals in California, the software total changed by +3.6% and the selling total by -0.4% in the published estimates (sums of survey estimates, not headcounts).
Key takeaways
- What was measured: federal counts of jobs and median pay by employer-assigned job title, May 2022 and May 2025. No public dataset measures AI agents as such, and this one does not.
- Both groups' averages fell in California metros in the published estimates (selling -6.6%, software -11.0%, averages across occupation-and-metro pairs). The selling group's change was about 4.9% higher than the change for four software occupations (-2.0% to +12.3%), and about 12.1% lower than the change for software developers alone (-18.8% to -4.9%). Neither is a growth rate for sales jobs.
- By job totals in California, the software total changed by +3.6% and the selling total by -0.4% in the published estimates (sums of survey estimates, not headcounts), so counting jobs and counting pairs give different pictures.
- The pay result is tentative, nominal, and disappears when renamed metros are dropped. Usage of one AI chat tool shows no clear relationship with job changes.
Are sales and marketing jobs holding up better than software jobs?
Neither confirmed nor refuted: the answer depends on how software is defined and on whether you count jobs or occupation-and-metro pairs. These are gaps between two groups' changes, not growth or decline of sales jobs. By job totals in California, the software total changed by +3.6% and the selling total by -0.4% in the published estimates (sums of survey estimates, not headcounts).
| California metros, selling minus software | Employment | Median hourly wage (nominal) |
|---|---|---|
| Four software occupations | +4.9% | +1.9% |
| Range for that result | -2.0% to +12.3% | -0.4% to +4.2% |
| Software developers only | -12.1% | +1.6% |
| Range for that result | -18.8% to -4.9% | -1.3% to +4.7% |
A positive figure means the selling group's change was higher than the software group's, in the published estimates. The range is a 95% range; when it includes zero, the data cannot rule out no gap at all.
The two software definitions differ because pair averages give each software occupation in each metro equal weight, so programmers and web developers count as much as developers, while the software job total is dominated by software developers. That is why the two definitions can give different answers.
What did we measure, and how?
We compared two federal snapshots of jobs and median hourly pay by employer-assigned title, not "AI agents": May 2022 and May 2025, for 11 selling occupations and 4 software occupations, in California metros and in metros elsewhere. The source is the Bureau of Labor Statistics employer survey, OEWS. The selling group is marketing, sales and customer service titles; the software group is programmers, developers, QA testers and web developers. The unit is one occupation in one metro area, called an occupation-and-metro pair; later, just "pairs". We used 3,029 pairs across 360 metro areas, including 25 California metros.
We went in expecting selling work to have held up better than software work. We fixed the four-occupation definition first, and declared developers-only as an alternative. For the software side, the direction was partly visible before the plan was frozen.
All changes here compare two snapshots the BLS does not design for comparison: the two releases share no survey panel, and some metro names changed between them. The BLS "does not encourage the use of OEWS data for time-series analysis," and "each set of OEWS estimates is produced by combining six semiannual panels of survey data collected over a three-year period." We describe differences in the published estimates, nothing more.
"Selling" here is 11 occupation titles, not "sales jobs" in general. Product managers and UX researchers have no occupation code of their own, so they are in neither group. We did not measure AI, agents, skills, or founders.
Technical detail: in log points, the four-occupation employment gap is 0.05 log points (+4.9%; 95% CI -0.02 to 0.12 log points, or -2.0% to +12.3%; p = 0.168, Holm-adjusted p = 0.204) and developers only is -0.13 log points (-12.1%; 95% CI -0.21 to -0.05 log points, or -18.8% to -4.9%). California rests on 198 selling pairs and 46 California software pairs in 22 metros; metros are not independent draws of "California". The California-versus-elsewhere wage difference is 0.04 log points (+3.9%; 95% CI 0.01 to 0.06 log points, or +1.3% to +6.6%), Holm-adjusted p = 0.018. For usage, rho = -0.12 (95% CI -0.23 to -0.02, Holm-adjusted p = 0.079), r = -0.07 (p = 0.155), and with missing shares set to zero rho = 0.04 (n = 762). Only r = 0.15 or larger was reliably detectable.
Why do job totals and pair averages disagree?
The totals are weighted by jobs, so the biggest occupations dominate. Pair averages count every occupation-and-metro pair once, so many small pairs weigh as much as one large one, and the two can disagree. In California, the software total is dominated by software developers.
| Change, May 2022 to May 2025, in the published estimates | Selling | Software |
|---|---|---|
| California, job totals (state estimates) | -0.4% | +3.6% |
| United States, job totals | +1.7% | +4.5% |
| California metros, pair average | -6.6% (n = 198) | -11.0% (n = 46) |
| Metros outside California, pair average | +1.3% (n = 2233) | -10.3% (n = 552) |
By job totals, software's change was higher than selling's in both places in the published estimates. By pair averages, selling's change was higher than software's in California and outside it.
The California gap was about 7.1% lower than the gap elsewhere (95% range -14.0% to +0.3%). The range only just includes zero, and the sample was too small to reliably detect a difference of that size, so this is not evidence of no difference. California ranks 23 of 51 states, a middle rank.
Do the sub-groups agree with the main answer?
No. Marketing, sales and customer service point in different directions, and these three looks are exploratory. On employment, customer service leans against selling (-4.6%, range -13.4% to +5.0%, which includes zero). They were not corrected for repeated testing; treat them as leads. Each compares California pairs of one sub-group with the 46 software pairs.
| Selling sub-group vs software, California employment | Result |
|---|---|
| Marketing (48 pairs) | +14.8% (+2.6% to +28.4%) |
| Sales (125 pairs) | +3.3% (-3.9% to +10.9%) |
| Customer service (25 pairs) | -4.6% (-13.4% to +5.0%) |
Customer service's nominal wage change was higher than software's, at +12.0% (95% range +7.8% to +16.4%).
Did selling pay move differently from software pay?
Only tentatively, and only in nominal pay. Nominal means not adjusted for inflation. California software medians rose +8.5% across 45 pairs in the published estimates, against +13.3% across 534 pairs in metros elsewhere. Selling medians were +10.5% in California (198 pairs) and +11.2% elsewhere (2213 pairs). The gap is mostly software medians rising less in California in these estimates, with customer service pay rising more.
The California-versus-elsewhere wage gap is +3.9% (+1.3% to +6.6%). It is fragile: with metros whose names changed dropped, it becomes +2.3% (-0.6% to +5.3%). With software plus UI designers, it becomes +2.4% (-0.7% to +5.5%). Both ranges include zero. California ranks 10 of 51 states on the state version of this comparison, so a gap this size is not rare among states.
Is more AI chat usage linked to job changes?
No clear relationship. We compared each California occupation's share of Claude.ai chats (April and May 2026) with its employment change from 2022 to 2025. That covers only 364 occupations: 398 of 762 occupations have no published share, and the published ones are larger. Usage was measured after the employment window, and it is one vendor's chat usage, not business adoption or agents.
Setting missing shares to zero reverses the sign, so the result depends on which occupations are included.
What can't this data show?
It cannot show who sells inside a firm, whether AI was used, why employment changed, or whether jobs moved to other titles.
- Job counts are by employer-assigned title. They say nothing about tasks or moves to titles outside the two groups, such as "growth" roles.
- The sample leaves out nonmetropolitan areas, the self-employed, owner-operators, founders, contractors and platform workers. It also drops 36 metro areas present only in 2022 and 33 present only in 2025 (including Madera, California), and pairs under 100 jobs or suppressed in either year.
- The 2022 and 2025 releases share no survey panel. Metro boundaries and names changed. With the eight of 25 California metros that were renamed dropped, the employment gap is +6.1% (95% range -3.4% to +16.6%), still positive but with a range that includes zero.
- Wages are nominal. Pairs with an unavailable or top-coded wage are excluded.
- Several looks were taken at one data set: the sub-groups, the wage result and the alternative definitions. Only the five main hypotheses were corrected.
- The usage data cover Claude chat and Cowork (Free, Pro, and Max plans) only.
How can you check it?
Every input is a public file. The OEWS state and metropolitan files for May 2022 and May 2025 are at the BLS data tables page; the time-series caution is on its OEWS FAQ. Usage shares come from the Anthropic Economic Index report of June 2026, published on Hugging Face, whose README says "Data released under CC-BY". Related: entry-level jobs and AI and junior developer jobs by age.
What can you do with this?
Run three checks on a "sales is safe" or "software is doomed" headline before you trust it.
- Ask how each group is defined. Here, four software occupations and developers alone gave opposite answers.
- Ask whether it counts jobs or occupation-and-metro pairs. Here, job totals and pair averages disagreed.
- Ask which two years are compared, and whether the source is built for comparing years. OEWS is not.
If you are deciding whether to hire for sales or to build software yourself, these figures do not tell you which is safer; they tell you the answer changes with how the question is asked. If you are deciding what to hire or automate, start from your own tasks. Our post on how many small businesses use AI shows what businesses of your size report doing.
Analysis by Deivy Hernandez, from public data files.
FAQ
Are sales jobs safe from AI?
These data cannot say. They count jobs by employer-assigned title and do not record whether AI was used. In California, the change in the published estimates for selling occupations was mixed, and one sub-group, customer service, leaned against selling on employment (-4.6%, range -13.4% to +5.0%, which includes zero).
Are marketing jobs doing better than software jobs?
In this exploratory look, the gap favoured marketing against the four software occupations: +14.8% (+2.6% to +28.4%). It is one of three sub-group looks that were not corrected for repeated testing, so treat it as a lead, not a finding.
Do product managers and founders count as software jobs here?
No. Product managers and UX researchers have no occupation code of their own, so they sit in neither group. Founders, owner-operators, contractors and the self-employed are outside the survey too.
Did software pay fall in California?
In the published estimates, nominal median pay for the four software occupations is +8.5% across California occupation-and-metro pairs, less than the +13.3% in metros elsewhere. Nominal means not adjusted for inflation.
Did customer service jobs hold up better than software jobs?
In this exploratory look it leans the other way, but the range includes zero: -4.6% (range -13.4% to +5.0%) for customer service against the four software occupations in California metros, on employment. Its nominal pay change was higher than software's.
Can I use this to decide whom to hire?
Not directly. It describes groups of occupations across metro areas, not the labor market for the role you need. Use it as a reason to be cautious about stories that say one kind of job is safe and another is not.