Quick answer
We found no relationship. Across 419 occupations, how San Francisco-area employment compared with the national count between two published snapshots (May 2022 and May 2025) showed no association with how much each occupation appears in one vendor's chat usage. A difference of up to about 10% between the least and most used occupations cannot be ruled out.
Key takeaways
- No relationship was found across 419 occupations between one vendor's chat-usage measure and how San Francisco-area employment compared with the national count between two published snapshots.
- The usage measure is the vendor's mapping of conversations to occupation tasks. It is national and was measured after the employment window.
- No public dataset measures AI agents as such, so this study does not.
- The five-occupation table is not a test and not a result about AI.
- A small relationship cannot be ruled out, and nothing here ties a difference to AI.
Are San Francisco jobs falling fastest in the occupations where AI chat is used most?
No. We found no relationship between an occupation's chat-usage share and how its San Francisco-area employment compared with the national count, across 419 occupations and two published snapshots. A difference of up to about 10% between the least and most used occupations cannot be ruled out. What we measured is the share of one vendor's chat conversations mapped to each occupation's tasks (Anthropic Claude.ai chat and Cowork, US, April-May 2026), not how much people in the occupation use AI, and no public dataset measures "AI agents" as such.
In plain words, the main estimate's 95% range runs from -10.0% to 11.9% across the usage range, so a much larger difference is unlikely and a small one cannot be ruled out. Noisy small cells and a coarse usage measure make a small relationship hard to see, so this is not proof that none exists. The comparison is between how San Francisco-area employment moved relative to the national count (May 2022 and May 2025). The table below shows five occupations people ask about.
What do five familiar occupations look like?
The table shows two published snapshots, May 2022 and May 2025, for five occupations. It is not a test, and no row is read as an AI result. Each figure is a single survey estimate with sampling error. The last column is BLS's relative standard error for the San Francisco-area cell, and a rough 95% interval is about twice the standard error either side. Data entry keyers (1,690 and 1,410 jobs) and receptionists are the least precise rows.
| Occupation | SF-area 2022 | SF-area 2025 | SF-area % difference | National % difference | SF-area relative standard error, 2022 / 2025 |
|---|---|---|---|---|---|
| Receptionists and information clerks | 9,780 | 8,510 | -13.0% | -10.0% | 5.0% / 8.8% |
| Customer service representatives | 25,770 | 19,930 | -23% | -9.9% | 2.7% / 3.1% |
| Data entry keyers | 1,690 | 1,410 | -16.6% | -19.3% | 20.1% / 11.2% |
| Bookkeeping, accounting and auditing clerks | 22,340 | 18,100 | -19.0% | -11.4% | 4.1% / 3.7% |
| Software developers | 67,010 | 69,030 | +3.0% | +10.0% | 3.0% / 2.3% |
The San Francisco-area estimate for customer service representatives was 23% lower in the May 2025 snapshot than in May 2022 (25,770 to 19,930); the national estimate was 9.9% lower. Both are survey estimates with sampling error (relative standard errors of 2.7% and 3.1% for the San Francisco-area cells).
If you run a San Francisco small business, look at an occupation's relative standard error in the table before you quote its local figure. Data entry keyers have by far the largest, and the receptionists' figure for 2025 is larger than the other three occupations'.
On the belief that only programmers or computer occupations are affected, the data say nothing for or against it, because no relationship was found with or without them. Junior versus senior roles cannot be tested in this data.
What did we measure, and how?
We lined up two things for each occupation: how the San Francisco-area employment estimate compared with the national estimate between two published snapshots, and where the occupation ranks in the vendor's chat data. Employment comes from the Bureau of Labor Statistics' OEWS survey; usage is the vendor's mapped share described above.
That usage measure is national, not San Francisco. It was measured after the employment window, covers one vendor only, and is not employer adoption. In the vendor's own file, a catch-all category, "Computer Occupations, All Other", carries a far larger share than "Software Developers" or "Computer Programmers", which suggests the measure follows how conversations are assigned to occupation codes.
The sample is the 607 occupations published for the San Francisco area in both snapshots, and 419 of them also have a published usage share. The other 188 have no published share. They sit below the vendor's publication threshold and are not necessarily zero (and many published shares tie at the bottom of the ranking), so we ran the analysis both ways, dropping them and counting them as zero. In every variant we tried, the interval included zero.
We started from an expectation: higher chat usage, lower San Francisco employment relative to the national count. The San Francisco counts for five occupations (receptionists, customer service, data entry, bookkeeping and software developers) were seen before the analysis plan was frozen, so the direction was visible beforehand for those five and was not a blind test. The other roughly 600 occupations were not seen.
Technical detail (skippable): the comparison could reliably detect a correlation of r = 0.14 or larger. The main model's slope was 0.00 log points (0.3%) from the lowest-use to the highest-use occupation, with a 95% CI of -0.11 to 0.11 log points (-10.0% to 11.9%), p = 0.951, R-squared 0.00, n = 419. A confidence interval is the range of results the data are consistent with. Spearman rho = 0.01 (95% CI -0.09 to 0.11). Outside computer and mathematical occupations (n = 400) the slope was 0.01 log points (95% CI -0.10 to 0.13). Counting unpublished shares as zero (n = 607) gave 0.05 log points (95% CI -0.05 to 0.14); other variants were also null. After Holm correction for five tests, all adjusted p-values were 0.65 to 1.00. The quadratic term gave p = 0.130, so there is no clear curve. The odds ratio for "fell more than the national count" was 1.11 (95% CI 0.55 to 2.26), with a cross-validated AUC of 0.47, which is no better than chance. Design effect 1.42; effective n = 295.
How did San Francisco compare with the national count overall?
Between two published snapshots, the San Francisco-area change was lower than the national change in 66% of the 607 occupations. Among the 329 occupations with at least 1,000 San Francisco-area jobs in May 2022 the share was 70%. All occupations together were 0.8% lower in the San Francisco-area 2025 snapshot, against 5.1% higher in the national one. Read that with care. BLS says it "does not encourage the use of OEWS data for time-series analysis", and the two estimates pool different three-year samples, so they are not the same workers. The BLS area title changed: San Francisco-Oakland-Hayward in 2022, San Francisco-Oakland-Fremont in 2025, with the same five counties in both (Alameda, Contra Costa, Marin, San Francisco, San Mateo). It is the metro area, not the city, and each occupation counts equally regardless of size. We do not explain the gap, and the study cannot say why it exists; the usage result does not explain it either way.
What can't the data show?
They cannot show why any occupation differs, and nothing here ties a difference to AI. The limits:
- Two snapshots, not a panel. Each pools three years of survey data and the two share no panel.
- The area is the five-county metro, not the city, and the usage measure is not specific to San Francisco.
- Usage was measured in April-May 2026, after the employment window, for one vendor's Claude.ai chat and Cowork on Free, Pro and Max plans only.
- 188 occupations have no published share, 27 small occupations (under 100 jobs in 2022) and 25 suppressed cells were excluded, so the sample is selected.
- OEWS covers wage-and-salary jobs, not the self-employed, and says nothing about individual firms, tasks or who performs them.
- Occupations differ in wages, industries and local demand for reasons the study did not model. It is observational and covers one metro area.
How can you check it?
Every input is a public file. The metro and national OEWS files for May 2022 and May 2025 are on the BLS data tables page, and the time-series caution is on the OEWS FAQ. The area titles and county lists are on the BLS definition pages for May 2022 and May 2025. Usage shares come from the Anthropic Economic Index report, "Cadences" (June 2026), published on Hugging Face, whose README says "Data released under CC-BY".
Related pages measure other things: AI use by city covers Census survey answers from businesses, sales and marketing jobs against software jobs compares employment and pay by occupation group in California, and entry-level jobs and AI uses payroll data by age.
What can you do with this?
Use it as a checklist when you read a claim about local jobs and AI. The data back these rules:
- Ask which measure the claim uses and what it was counted from.
- Ask whether the two numbers are from the same workers. OEWS snapshots are not.
- For one occupation in one metro area, look at the standard error before trusting the percent. A rough 95% interval is about twice it either side.
- Treat a claim that links a single row to AI as unsupported by this study, whichever way it points.
Analysis by Deivy Hernandez, from public data files.
FAQ
Why compare San Francisco with the national figure?
An occupation can change everywhere at once. Setting the San Francisco-area change beside the national change for the same occupation keeps a general shift from being read as a local one.
Why two snapshots and not a trend line?
BLS says it "does not encourage the use of OEWS data for time-series analysis". Each estimate pools three years of survey responses, so two snapshots are two separate pictures, not points on a line.
Why not measure AI agents directly?
No public dataset measures AI agents as such. The only usable measure was one vendor's chat conversations, so the findings carry that narrower name.
Why use the metro area and not the city of San Francisco?
BLS publishes these estimates for the five-county metro area. There is no city-level OEWS estimate for each occupation to use instead.