OpenAI's July 27 Work at the Frontier report found that 43.5% of occupation-specific messages in its sample mapped to tasks historically associated with another occupation. The analysis covered more than 800,000 work-related messages from individual ChatGPT accounts of U.S. users whose roles were linked from Business account information; it did not measure productivity, output quality or employment effects.
OpenAI's July 27 report, Work at the Frontier: How AI is expanding what people do at work, found that 43.5% of occupation-specific messages in its sample concerned tasks historically associated with an occupation other than the user's own. Across all work-related messages, including generic work such as writing and scheduling, the cross-occupation share was 16.8%.
The study is evidence about how people ask ChatGPT for help, not proof that they completed specialist-quality work or that jobs were displaced.
What OpenAI measured
The researchers analyzed a random sample of more than 800,000 work-related messages sent from individual ChatGPT accounts of U.S. users. Occupations were assigned from self-reported department or role information linked from ChatGPT Business accounts, and task boundaries were based on the U.S. Department of Labor's O*NET data.
That account boundary matters. The analyzed messages came from the users' individual ChatGPT accounts, while the role information came from Business. OpenAI says the sample should not be generalized to ChatGPT Enterprise users or the entire U.S. workforce.
The researchers classified messages as generic, within occupation or cross-occupation. Generic work made up 61.5% of the sample; 21.8% was within occupation and 16.8% crossed an occupational boundary. When generic messages were excluded, cross-occupation tasks made up 43.5% of the remaining occupation-specific messages.
Where crossover appeared
OpenAI reported the highest occupation-specific crossover rates among customer-experience workers at 77%, designers at 75%, human-resources workers at 69%, legal workers at 56% and marketers at 53%. Marketing and engineering tasks appeared frequently in messages from people in other roles.
The report also found an asymmetry. About 35.2% of all messages from designers involved tasks associated with another occupation, while design tasks made up only 1.7% of messages from workers in other fields. Engineering showed the reverse pattern: 18.5% of engineering messages involved outside tasks, while engineering tasks appeared relatively often in other occupations.
Among typical-volume users, the all-message cross-occupation share was 18.9% in workspaces with two to five seats and 16.3% in workspaces with 101 or more seats. OpenAI treats workspace seats as different from total company size and describes the relationship as descriptive, not causal.
What the study cannot establish
Axios noted the authors' central caution: the data cannot determine whether AI created new cross-occupation work or helped users perform responsibilities they already had. The unit of analysis is a message, not a completed project or an hour of work. The study does not observe whether the output was used, whether it was correct, how much time it saved or whether a specialist reviewed it.
For data and ML teams, the practical signal is therefore about governance scope. General-purpose assistants are being asked to support work outside users' usual roles. Organizations need review and accountability processes that follow the risk of the task, especially for legal, financial, technical and customer-facing work, rather than assuming a user's job title defines every capability they will access through AI.
Key Points #
- 1OpenAI measured a 43.5% cross-occupation share among occupation-specific messages and 16.8% across all work-related messages in its sample.
- 2The study analyzed more than 800,000 messages from individual U.S. users whose roles were linked from ChatGPT Business account information.
- 3Message-level task crossover does not establish productivity, work quality, job displacement or whether a specialist reviewed the output.
Scoring Rationale #
The study offers large-scale descriptive evidence about how U.S. ChatGPT users seek help across occupational task boundaries. It is relevant to AI governance and job design, but it does not measure productivity, output quality or labor-market outcomes.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.