OpenAI Finds Workers Using ChatGPT Beyond Their Job Titles

OpenAI’s economic research team says roughly 44% of the occupation-specific requests its US business customers send to ChatGPT involve work that normally belongs to a different profession. The finding rests on a sample of more than 800,000 work-related messages and was first shared with Axios.

Strip out the generic office work that every job contains, such as drafting emails and booking meetings, and 44% of what remains sits outside the requester’s own discipline. The rate splits hard by function. Customer experience staff came in highest at 77%, followed by design at 75% and human resources at 69%. Legal registered 56%, marketing 53%, and sales and finance 40% each. Engineering was the low outlier at 28%, the one function whose ChatGPT use mostly stays in its own lane.

“The boundaries between jobs are likely already becoming more flexible due to AI,” Ronnie Chatterji, OpenAI’s chief economist, told Axios. The specific tasks cited are the kind a company would normally route to a specialist: building marketing collateral, debugging software, running financial calculations, interpreting regulations.

What the data measures

Two different denominators are in play, and they are easy to conflate. The 44% figure covers only occupation-specific requests. A separate cut, drawn from typical ChatGPT users rather than the occupation-tagged sample, found that about 19% of all work-related requests at the smallest businesses crossed occupational lines, against roughly 16% at larger firms. OpenAI reads that gap as small companies leaning on the model where there is no specialist on the payroll.

The authors are explicit about the limits. The analysis cannot separate two very different stories: AI handing workers tasks they never had, or AI helping them do things already sitting in their remit. It also says nothing about whether the output was any good, whether the work got done faster, or whether any employer changed a hiring plan because of it.

This is usage telemetry produced by the company that sells the seats, and it measures what people typed rather than what the output was worth. OpenAI has been making a version of this argument commercially for a while, and the enterprise business it supports is now large enough that the framing carries weight with buyers.

The seat-pricing argument

ChatGPT for business is sold per person. A general assistant priced that way is worth considerably more to a buyer if every seat is used across functions than if it is a writing tool that marketing opens twice a day. Cross-boundary usage is the strongest available case for broad deployment rather than departmental pilots, and it is the case OpenAI has just put a number on.

Set it against the company’s own B2B Signals report from May 6, 2026, and the picture gets more complicated. That analysis found enterprise usage clustering tightly around each function’s core work: 78.3% of sales messages were writing and communication, 60.2% of IT and security messages were procedural how-to, and 48.9% of software development messages were coding. The two readings are not strictly contradictory, since broad task categories and fine-grained occupational mapping can capture different things in the same message. But they point buyers in different directions, and only one of them argues for expanding the license beyond the teams already using it.

Scale is not the constraint here. OpenAI said in July 2026 that Codex and ChatGPT Work had reached 10 million users, and the deployment question has moved from access to what people do once they have it.

Where this shows up before the labor statistics do

OpenAI’s AI Jobs Transition Framework, published April 25, 2026, sorted 921 occupations covering roughly 148 million US jobs into four buckets and put 24% of them in the category most likely to reorganize rather than disappear. Cross-boundary prompting is what that reorganization looks like in telemetry, months or years before it surfaces in unemployment data or job postings.

For companies actually running these deployments, the practical questions are unglamorous and immediate:

• Who signs off when a customer-support rep runs a financial calculation or a designer drafts contract language, and what review sits between the output and the customer
• Whether departmental tool budgets survive if the general assistant absorbs the tasks the specialist software was bought for
• Whether job descriptions and internal capability mapping get rewritten to match what people are already doing
• Whether the pattern holds outside OpenAI’s own usage data, which no external researcher can currently audit

The last point is the one to watch. OpenAI’s Signals research page is where its economic work has landed since February 2026, and the underlying report for this analysis has not appeared there yet. Until it does, a set of numbers that speaks directly to how firms should structure hiring exists only as a chart in one news story, sourced entirely to the vendor whose product it measures.