A message can cross a job line
A salesperson asks for help exploring a customer dataset. A designer checks a contract. Someone in human resources troubleshoots a software problem.
None of that proves a job has changed. It does show how easily a task can move when a general-purpose assistant is sitting in the same window as the rest of the work.
OpenAI has tried to measure that movement. Its new report analyses more than 800,000 work-related messages from US ChatGPT users. It covers eight broad occupation groups: customer experience, design, engineering, finance, human resources, legal, marketing and sales.
The researchers matched each message to a work activity in O*NET, the US Department of Labor's occupation database. They then asked whether that activity sat inside the user's stated occupation, outside it or was too common across jobs to count either way.
OpenAI calls the outside category task crossover. The name is new. The behaviour is familiar to anyone in a small team who has fixed the website, checked a number or written a bit of copy because the specialist was not there.
Most work messages were ordinary shared work
The headline number needs its denominator.
Across the whole sample, 16.8% of work-related messages were classified as cross-occupation. Another 21.8% stayed within the user's occupation. The largest group, 61.5%, was generic work such as writing, summarising or scheduling.
Once that generic work is removed, 43.5% of occupation-specific messages fall outside the user's own field. The share reaches 77% for customer-experience users, 75% for designers and 69% for people in human resources.
Those percentages describe the messages that were specific enough to map to an occupation. They do not mean three quarters of a designer's day has moved into another job. A message is not an hour, a project or a share of someone's salary.
The study used role information people had self-reported through ChatGPT Business and work messages from their individual ChatGPT accounts. OpenAI says a model classified the messages anonymously and researchers did not read the underlying conversations.
Some tasks travel almost everywhere
The useful part is not just how often people cross a boundary. It is what they reach for on the other side.
Financial calculation ranked among the three most common finance-related tasks for all seven non-finance groups. Computer troubleshooting did the same across every group outside engineering.
Marketing work also travels widely. People asked for ads, social posts, flyers, product sheets and other promotional material. Engineering tasks showed up often in design, customer experience and finance messages.
There are different patterns inside that traffic. Designers drew heavily on tasks associated with other occupations, while design work rarely appeared elsewhere. Engineering supplied tasks that travelled out. Marketing did both.
That is a more grounded picture than saying AI simply automates jobs. A job can keep its name while the bundle of work inside it becomes a little wider, messier and harder to classify.
Smaller workspaces show a modest difference
OpenAI also looked at the number of seats in a user's ChatGPT workspace.
Among users in the middle half by message volume, cross-occupation messages made up 18.9% of work use in workspaces with two to five seats. The share was 16.3% in workspaces with more than 100 seats.
One plausible explanation is simple: smaller organisations have fewer specialists to hand work to. But the study cannot establish that cause. Workspace seats are not the same as company size, and they can also reflect industry, maturity or how widely a company has rolled out ChatGPT.
The pattern was not a steady decline among the heaviest users. People with established AI routines may use the tool differently, or they may send many messages while staying closer to their core work.
So the small-business result is a clue, not a rule. The gap is modest and descriptive.
Trying a task is not the same as doing it well
Several findings are confirmed within this dataset: work messages often combine tasks from different occupational traditions, marketing and engineering tasks travel widely, and the sample shows a small-workspace gradient among typical-volume users.
OpenAI's broader claim is that AI may broaden roles before job titles or official statistics catch up. The evidence supports that as a possibility, not a settled labour-market outcome.
The study does not observe whether an answer was used, whether it was correct, how much time it saved or whether a specialist checked it. It does not measure productivity, pay, hiring or job losses. The sample is also not representative of the entire US workforce.
One message can contain several overlapping tasks, yet the method assigns it one primary activity. Occupational boundaries are built from historical O*NET descriptions, which are useful but not perfect descriptions of work happening now.
The open question is less dramatic and more practical: when people take on work outside their expertise, who checks the result? AI can make the first attempt easier. It cannot quietly supply the accountability that used to arrive with the handoff.
Sources
- OpenAI Economic Research — How AI is expanding what people do at workPrimary research announcement published 27 July 2026. Source for the study's headline findings and OpenAI's interpretation of task crossover.
- Chin and Richmond — Work at the FrontierPrimary 16-page report. Source for the sample, classification method, occupation and workspace comparisons, privacy description and methodological limitations.
