OpenAI Releases Second Work at the Frontier Report on New Ways of Working

The team analyzed more than 1.5 million work-related ChatGPT messages sent by US-registered users over four months from April through July 2026. Occupations came from the role or department information users entered when onboarding to ChatGPT Business, which was linked to the same person's individual-account messages. Messages sent through Business accounts were not included, and users who opted out of training and training-disabled messages were excluded. Messages were automatically scrubbed of personal information before models assigned task labels, and the report states that researchers did not read individual conversations.

Whether a task lies inside or outside an occupation was decided with O*NET, the US Department of Labor's occupational database. Each message was mapped to the single detailed work activity that best matched it, and if the occupation traditionally associated with that activity differed from the sender's occupation, it counted as cross-occupation work. Comparisons covered nine occupation groups, and the team wrote that the sample consists of ChatGPT users with linked occupation information and is not representative of the US workforce.

People write prompts differently when asking for help with work that is not their own. Cross-occupation requests were about 20 characters shorter on average than within-occupation requests, and less likely to ask for explanations or how-to guidance, a specific format, or advice. They were more likely to include examples or background material and to ask for checking or verification. Alex Martin Richmond and Caroline Chin of OpenAI's Economic Research team, the report's two authors, wrote, "AI lets workers 'borrow expertise' by drawing on assistance associated with another field." The interpretation is that workers are not trying to learn a new field but bringing a problem and its context and asking AI to apply someone else's knowledge.

The key figure is recurrence. Among roughly 6,200 workers observed throughout the four months, the share of occupation-specific AI activity taken up by cross-occupation tasks they had already used nearly doubled, from 13.1% in April to 25.9% in July. The team read this as workers building the assistance into their workflows rather than trying it once. The calculation covers only workers with at least 10 sampled messages in each of the four months, and a footnote notes that the number of tasks that can count as recurring grows as the months pass.

Measured against a comparison group, the direction is the same. Workers who used AI for a cross-occupation task in one month used the same task the next month 23.6% of the time, compared with 8.4% for workers in the same broad occupation with similar activity levels but no recorded use of that task the previous month. That is a gap of 15.2 percentage points, and similar gaps of 15 to 17 points appeared for within-occupation and generic tasks. The coworker effect was far smaller. When someone in a workspace performed a cross-occupation task, coworkers who had not used it that month used it the following month in 3.1% of cases, slightly above the 2.5% among comparable workers in other workspaces.

Seen through a sociologist's eyes, the report is a map of where the boundaries of the division of labor give way and where they hold. The share of workers who returned to the same cross-occupation task in the month after first using it averaged 18.5%, roughly one in five. Discussing goods or services with customers came back 54% of the time, advertising or promotional writing 44%, and creating marketing materials 37%. By contrast, explaining financial information to customers returned 15%, presenting business information 15%, and legal research 10%. The team wrote that these differences may reflect where AI fits naturally into recurring workflows, or differences in workplace norms, caution, or the perceived consequences of getting something wrong. A picture in which marketing crosses over while legal and finance stay put fits the long-standing observation that boundaries are firmest where professional credentials and liability are at stake.

The team laid out three paths that cross-occupation experimentation could take: it could prove unproductive or unwelcome within the organization and recede within months, stabilize at a steady share of overall activity, or prove rewarding enough that its share of AI use keeps growing. Four months of data pointed to the third. In an appendix that takes the 2,100 highest-activity workers, the top third, and counts generic and unmapped tasks in the denominator, the share still rose from 7.0% in April to 10.6% in May, 11.8% in June, and 12.1% in July. The level was lower but the direction was the same.

The report's conclusion is that jobs change before titles do. When a worker experiments with an activity outside a traditional role, finds AI useful for it, and keeps returning to it, the mix of activities within the job broadens even while its title stays the same. The two authors wrote, "Our findings suggest that work design deserves a place alongside access to AI tools in how organizations implement their AI strategies." The report's introduction notes that the rise of the modern corporation created specialized departments such as finance, legal, marketing, and human resources, so that completing a task often requires several people, and that generative AI is an opportunity to change that structure by helping workers handle more of the work themselves.

METAL has reported that one in five US workers delegates tasks to AI rather than to colleagues; this report asks whether that delegation carries into the following month, tracking the same people for four months. METAL has also reported on OpenAI's launch of an agent that answers questions about company data. What this report measures is whose work such tools broaden inside a company, and by how much.

METAL reviewed the original report PDF, which runs to 13 pages and presents the figures above in five charts plus one appendix chart. Its methodology section notes that all activity measures refer to observed AI use rather than total work or time spent working, and that repeated use in unsampled conversations may be missed. The team said it will continue to study how AI is changing the division of labor and what that means for workers, businesses, and the broader economy.

Occupational boundaries are being redrawn not by org charts but by what people return to each month. This report is the first to count those returns, and it leaves a numerical record of which boundaries open first. The sequence in which work changes before job titles do has already begun.