OpenAI says AI is creating jobs humans couldn’t do before

Is AI changing the kind of work people do, and not just how fast they do it? New research from OpenAI's Economic Research team suggests it is changing who does what.

The report, Work at the Frontier: How AI is Expanding What People Do at Work, was published on July 27 by Caroline Chin and Alex Martin Richmond. It is the first in a planned series.

Workers are taking on tasks from other roles

The study analysed a random sample of more than 800,000 work-related messages from US ChatGPT users, matched to self-reported occupations from their ChatGPT Business accounts. OpenAI found that 16.8% of work-related messages, and 43.5% of occupation-specific messages, concerned tasks historically associated with a different occupation.

The company calls this task crossover. Note the unit: these are individual messages, not whole conversations. The remaining split is 61.5% generic work such as writing emails and scheduling, and 21.8% squarely inside the user's own role.

The tasks that travel furthest are financial calculation and computer troubleshooting. Each ranks among the three most common outside tasks in every one of the seven other occupation groups studied.

Which teams cross over most

Once generic work is excluded, outside-occupation tasks account for 77% of occupation-specific messages from customer experience workers, 75% from designers, 69% from human resources, 56% from legal teams and 53% from marketers. Cross-occupation work forms a majority in five of the eight groups measured.

Some jobs absorb work, others export it

The more interesting pattern is directional. About 35.2% of messages from designers involve work usually associated with another occupation, but design tasks make up only 1.7% of messages from workers in other fields. Designers pull in work from everywhere and almost nobody pulls in theirs.

Engineering runs the opposite way. Only 18.5% of engineering messages involve outside tasks, but engineering work accounts for 7.4% of messages from other occupations. Marketing does both, at 24.3% inbound and 8.9% outbound, the highest outward share in the sample.

Smaller workspaces, with caveats

Crossover is somewhat higher in small workspaces, but the effect is narrower than it first appears. Among users in the middle 50% by message volume, the cross-occupation share falls from 18.9% in workspaces with two to five seats to 16.3% in workspaces with more than 100 seats. That is roughly 2.5 percentage points.

Among the heaviest users, the pattern does not hold at all. OpenAI also notes that workspace seats are not the same as company size.

A separate study on agentic tools

Related numbers have been circulating from a different OpenAI paper, The Shift to Agentic AI: Evidence from Codex, published a month earlier on June 25. It found that by May 2026, 70.2% of sampled Codex users had submitted at least one task estimated to take a person more than an hour, and 25.6% had submitted one estimated at more than eight hours. Inside OpenAI, legal, finance and recruiting teams adopted Codex as their main AI tool around April 2026, using it for automation, data clean-up and analysis.

Those thresholds are model-estimated rather than measured, and drawn from a 0.1% random sample of users who opted in to training. They belong to a different dataset and should not be read as part of the crossover findings.

What the research does not show

OpenAI is explicit about the limits. The study does not observe whether AI output was used, how good it was, how much time it saved, whether the user could have done the task without AI, or whether a specialist reviewed it. The sample is not representative of the US workforce, covers only eight occupation groups, and excludes ChatGPT Enterprise users entirely.

It is also worth stating plainly: this is OpenAI studying its own product, using its own customers, and it has not been peer reviewed. The findings are descriptive, and India is not in the sample at all.

What businesses should do next

Within those limits, the signal is still useful. If employees are routinely handling adjacent work rather than passing it along, organisations need review processes that match. OpenAI's own conclusion is that workers may need training to evaluate AI-assisted work outside their expertise, and firms will need clear lines of accountability for it.

For workers, the skills that follow are unglamorous but durable: knowing when output is wrong, and knowing when to call the specialist anyway.