OpenAI Study: Cross-Occupation AI Tasks Becoming Routine

Workers who borrow skills from other occupations with ChatGPT keep coming back: among roughly 6,200 workers tracked consistently from April through July 2026, previously used cross-occupation tasks grew from 13.1% of occupation-specific AI activity in April to 25.9% in July. That finding comes from OpenAI's second Work at the Frontier report, based on an analysis of more than 1.5 million work-related ChatGPT messages over the same four-month window.

The research matters because it moves the AI-and-jobs debate past first contact. OpenAI's first Work at the Frontier report documented "task crossover": workers using AI for activities historically associated with another occupation. The new report asks what happens next, and the answer is recurrence. Workers return to tasks outside their occupational boundaries, and those activities become a larger share of their observed AI use over time. For companies working out how to adopt AI, the study offers an early window into how one-off experimentation turns into embedded workflow — and OpenAI argues that work design deserves a place alongside tool access in any AI strategy.

Prompting shifts when tasks fall outside the role

OpenAI found that workers prompt AI differently depending on whether a task fits their role. For tasks outside their occupation, they write shorter prompts on average than for tasks within it. They are also less likely to ask for explanations, how-to guidance, a specific response format, or advice.

They compensate in a different way. When asking AI for help outside their occupations, workers are more likely to provide examples or background — a document, a sample, information from a colleague. They are also more likely to ask AI to check or verify something.

OpenAI's interpretation: workers are using AI to borrow expertise. Rather than asking the model to teach them an entirely new field, they bring a problem plus relevant context and ask AI to apply knowledge associated with another field to it. That distinction — delegation of expertise rather than learning — shapes what organizations should actually train people to do.

Recurrence, quantified

The core question the report set out to answer is whether workers return to these out-of-role activities. Two analyses say yes.

First, the longitudinal sample. Among the roughly 6,200 workers observed consistently across the four months, the share of occupation-specific AI activity devoted to previously used cross-occupation tasks nearly doubled, from 13.1% in April to 25.9% in July. OpenAI calls that pattern consistent with workers incorporating cross-occupation assistance into ongoing workflows rather than simply experimenting once.

Second, a matched follow-up analysis. Among sampled matched one-month follow-up observations, workers returned to a cross-occupation task they had used in the previous month 23.6% of the time. Comparable workers with no observed use of that task in the previous month used it just 8.4% of the time. Similar gaps appeared for within-occupation and general tasks, suggesting prior use predicts future use across the board.

Some borrowed tasks stick; others don't

Recurrence rates vary sharply by task type. Workers returned the following month to some cross-occupation tasks at high rates: discussing goods or services with customers (54%), advertising or promotional writing (44%), and creating marketing materials (37%).

Other tasks fared worse. Workers returned to the task of explaining financial information only about 15% of the time. The average next-month return rate across all cross-occupation tasks was 18.5%.

OpenAI offers two possible explanations for the spread. The differences may reflect where AI fits naturally into recurring workflows. They could also reflect workplace norms, caution, or the perceived consequences of getting something wrong — a plausible reason explaining financial information, where errors carry real costs, lags marketing copy, where they mostly don't.

Job expansion before job titles change

Taken together, the first two Work at the Frontier reports sketch a mechanism for AI-driven job transformation that operates below the level of organizational charts. A worker experiments with an activity outside their traditional role, finds AI useful for it, and begins returning to that activity as part of their regular work.

The consequence, in OpenAI's framing: AI may reshape jobs well before their titles change. If cross-occupation activities become regular responsibilities, the mix of activities within a job could broaden even while its title stays the same. Job descriptions, role definitions, and possibly compensation structures would then be out of sync with what people actually do all day — a gap HR departments and managers will eventually have to close.

OpenAI says it will continue studying these shifts to build a clearer picture of how AI is changing the division of labor and what that means for workers, businesses, and the broader economy. The company has also published an AI Jobs Transition Framework modeling how AI capabilities, human roles, and demand could shape employment. For organizations still treating AI adoption as a procurement question, the data points elsewhere: the durable changes are showing up in how work itself gets divided, not just in who has a license.