Responses come from an Epoch AI survey fielded on Ipsos’ KnowledgePanel, a probability-based online panel, from July 10–19, 2026. The survey covered 1,106 employed US adults. We removed three non-valid responses, leaving 1,103 respondents. All estimates are weighted.

Of the 1,103 respondents, 469 reported using AI at least partly for work in the past seven days; they form the base for this analysis (n = 469). The subscription question was asked of all 469. For each AI service a respondent used for work, it asked what type of subscription they used most of the time, with response options “Free plan (no paid subscription)”, “Employer-provided subscription”, “Personal paid subscription”, “Both employer-provided and personal paid subscription”, and “Not sure”.

Occupation comes from the survey’s self-reported occupation question, which asks what kind of work respondents do in their main job, using US Census Bureau occupation categories. Each category maps to a single 2018 SOC major group. 111 respondents chose “Other, please specify”. We coded 95 out of the 111 write-in answers into the survey’s occupation categories, using an LLM for initial coding and reviewing the assignments by hand. We were unable to assign the remaining free-text responses to an occupation group; these are included in the overall estimate but not in any occupation group.

We mapped each survey occupation category to its 2018 SOC major group and aggregated the major groups into the BLS high-level groups, with two deviations:

1. We broke out computer, engineering, and science occupations (SOC 15–19) as their own group, separate from the rest of BLS group 1 (management, business, science, and arts).
2. We combined BLS groups 4 and 5 (natural resources, construction, and maintenance; production, transportation, and material moving) into a single group.

For each group, we computed the weighted share of respondents in each subscription category. We report 90% Wilson confidence intervals for each estimate rather than a single margin of error, computed on Kish’s effective sample size to account for unequal weighting.

The full mapping, with the survey’s occupation categories listed verbatim from the questionnaire:

The occupation groups contain 49–207 respondents, so confidence intervals are wide. Treat the group-level point estimates as indicative of large differences, not precise levels.

Occupations are self-reported at the level of broad survey categories, and about 9% of respondents wrote in a free-text occupation that we coded ourselves. Coding errors could move individual respondents between groups, but our robustness checks suggest the overall pattern is independent of the coding.

The survey asked which subscription respondents used most of the time for work. Workers who use several tools are classified by their main one.

Although results are weighted to be representative, as with all surveys, non-respondents may differ from respondents in ways the weighting does not capture. Usage is self-reported and may be affected by recall error or misclassification.