When a customer service worker uses ChatGPT to run a financial calculation that would once have gone to an analyst, or when an HR professional reaches for the platform to troubleshoot a technical system they would have sent to IT, the shift feels like a convenience. OpenAI's new economic research released Monday says it is something larger: a measurable, large-scale inversion of the division of labor itself. A new OpenAI study drawn from more than 800,000 ChatGPT messages from U.S. business users found that 43.5% of occupation-specific AI requests involved tasks that traditionally belong to a different job role. The phenomenon — which the researchers call "task crossover" — is the first large-scale empirical evidence that the specialization trend that has organized work since Adam Smith's pin factory may be running in reverse.

"The boundaries between jobs are likely already becoming more flexible due to AI," Ronnie Chatterji, OpenAI's chief economist, said in an interview with Axios ahead of the report's public release.

The report, titled Work at the Frontier: How AI is Expanding What People Do at Work, is the first installment of an ongoing series OpenAI is framing as a deliberate effort to give policymakers and employers empirical data rather than speculation. Alongside the report, OpenAI opened an External Research Exchange, inviting independent labor economists to propose studies using platform data.

How Researchers Separated the Signal From the Noise

The core methodological insight of the study is a distinction that sounds simple but turns out to be analytically significant: not every ChatGPT message at work tells you anything interesting about occupational change.

The research team first filtered out what they classified as "generic" work requests — the 61.5% of all work-related messages that could have come from any employee in any role, including drafting emails, summarizing documents, and scheduling tasks. These tell you that people are using AI, but not what their use reveals about their professional identity. Once that layer is stripped away, the remaining 38.5% of messages are "occupation-specific" — tied to knowledge or skills associated with a particular professional domain. It is within this narrower, more informative slice that the 43.5% crossover rate emerges.

The classification relies on O*NET, a U.S. Department of Labor database that describes occupations by their specific work activities and detailed task requirements. An LLM-based classifier reads each message transcript, summarizes what the AI was doing for the user, and maps it to O*NET activity codes. When those activity codes correspond to a different occupation than the one inferred for the user, that message is counted as task crossover. The methodology surfaces something that conventional labor statistics — which measure employment counts, wage changes, and job postings — cannot: what workers are actually doing moment-to-moment, and whether that activity matches their formal role.

Which Occupations Borrow Most — and Which Are Borrowed From

The crossover is not evenly distributed. Customer experience workers show the highest rate: once generic requests are stripped out, 77% of their occupation-specific prompts involve tasks from another professional domain, according to Axios. Designers come in at 75%, HR professionals at 69%, legal workers at 56%, and marketers at 53%. Sales and finance workers clock in around 40%, and engineering workers are the lowest in the study at 28%.

But the share of crossover requests a worker sends is only half the picture. The other half is directionality — whether a given occupation's tasks travel to other workers, or whether those workers borrow from others while rarely being borrowed from.

Design illustrates what the researchers call an "inward" pattern: designers draw broadly from outside disciplines — about 35% of their messages involve tasks from other occupations — but design tasks themselves appear in only 1.7% of other workers' messages. Designers borrow widely and are rarely borrowed from. Engineering is roughly the inverse: engineers stay largely within their own domain, with only 18.5% of their messages involving outside-role work, but engineering tasks spread widely across the organization — showing up in 7.4% of messages from workers in other fields.

Marketing occupies a distinct position, exhibiting high crossover in both directions. Marketers draw from multiple domains at 24.3% of their messages, and marketing tasks appear in 8.9% of messages from workers in other fields — the highest outward share in the sample.

Why Small Businesses Are the Canary

Company size turns out to be a meaningful variable. Among typical ChatGPT users, the share of outside-occupation task requests falls from 18.9% at businesses with just two to five seats to 16.3% at organizations with more than 100 seats, according to Axios. The gap is modest but consistent, and its explanation is intuitive: in a large firm, a worker with a marketing question can ping the marketing team; in a five-person company, that worker may be the marketing team today.

"AI may be especially useful as a generalist tool where specialist resources are scarce," the report notes. That framing suggests a new economic dynamic: AI functioning as an on-demand specialist, enabling small-business workers to cover expertise gaps that would otherwise require a hire or an outside contractor. If the crossover rate at small businesses is meaningfully higher than at large ones, it raises a downstream question that the current data cannot yet answer: are those small businesses adjusting their hiring decisions as a result?

What This Tells — and Does Not Tell — Labor Economists

The authors are explicit about what the study cannot establish, and those caveats matter. The data does not determine whether AI is creating genuinely new cross-occupational work for these employees or whether workers are simply doing things they were already responsible for, now with better tools. It does not measure whether AI-assisted crossover work matches the quality of specialist output. It does not track whether the pattern drives productivity gains. And it does not establish a link to downstream hiring decisions or wage changes.

These are not minor gaps. A marketer who can now debug a website independently is doing something newly possible, but whether that produces an outcome equivalent to a developer's work — and whether it eventually means fewer developers are hired — is a separate question that usage data alone cannot answer.

What the data can do, the researchers argue, is function as a leading indicator that conventional labor statistics will confirm only later. Job descriptions, wage surveys, and employment counts are all lagging signals — they measure what has already settled into organizational structure. AI chat logs, by contrast, capture workers experimenting with new task combinations in real time. The 43.5% crossover rate is a signal of organizational change in progress, not a final measurement of organizational change achieved.

The broader independent research landscape supports treating the finding as significant without overstating it. The Peterson Institute for International Economics noted in early 2026 that AI labor market research is "still in the first inning," with methodology debates still unresolved. A parallel study using Anthropic's Claude data from June 2026 found that more than a third of 9,700 surveyed workers expected their job responsibilities to change significantly within a year — while only about 10% feared losing their role — suggesting that role reorganization, not wholesale job elimination, is the near-term story. Goldman Sachs research published in April 2026 estimated AI was eliminating roughly 16,000 net U.S. jobs per month, with substitution concentrated in data entry, customer service, and legal support.

There is also a self-serving dimension to the data worth acknowledging. OpenAI's research function has a built-in incentive to frame its platform's impact favorably. Inviting independent economists to propose studies using platform data — the External Research Exchange — is a partial response to that concern, but the present study remains internally produced and has not yet been peer-reviewed.

What Economists Have Been Predicting — and What This Data Tests

The finding lands inside a specific academic debate. Since David Autor, Frank Levy, and Richard Murnane established the task-based framework for analyzing automation in 2003, labor economists have understood that technology substitutes for specific tasks rather than entire jobs. Daron Acemoglu and Pascual Restrepo extended that framework to distinguish between the "displacement effect" — automation replacing tasks — and the "reinstatement effect" — new tasks created where humans have a comparative advantage.

The OpenAI "task crossover" finding is an empirical observation of the task reallocation mechanism in action. Workers are not waiting for job descriptions to change; they are reallocating task sets themselves, in real time, through AI. The question the task-based framework would then ask is: does this task expansion represent genuine new work (the reinstatement channel) or task substitution that will eventually reduce headcount? The current data cannot distinguish between the two.

The historical context sharpens why that distinction matters. Since Adam Smith's Wealth of Nations identified specialization as the engine of industrial productivity, the dominant trajectory of work has been toward deeper specialization — workers doing fewer things more expertly. Every major wave of technology has, on net, deepened that trend by enabling larger-scale production that rewards expertise. If AI is reversing the direction — enabling individual workers to access adjacent expertise at low cost, reducing the value of specialization as an organizational principle — that would represent a structural shift in how companies should be designed and how workers should invest in their skills.

The 43.5% crossover rate is not proof of that reversal. It is evidence that the mechanism for that reversal exists and is operating at scale. Whether it will ultimately produce shallower specialization across the economy, or whether it will be offset by new forms of expertise deepening, is the research question this data opens rather than closes.

What It Means for Employers and Workers Right Now

The practical implication for employers is pointed: job descriptions written around stable task lists may already be disconnected from what employees actually do. If nearly half of occupation-specific AI use is crossing role boundaries, then the task inventories that underpin most HR systems — and most hiring decisions — may be underspecifying how work is actually organized in AI-assisted environments.

For workers, the data suggests that AI use is measurably expanding the effective scope of their roles. Whether that expansion is recognized in compensation, titles, or promotions is a question the study does not address — but it is the question workers asking their managers about career development should be asking with data in hand.

For policymakers and labor economists tracking AI's effects, the "Work at the Frontier" series represents a new category of evidence: platform-level behavioral data that captures task-level changes as they happen, before those changes propagate into the official statistics that governments use to measure employment. If the crossover trend persists and deepens, it will eventually appear in wage surveys, job posting data, and occupational employment statistics — but it will arrive there as a confirmed trend, not a current one. OpenAI's data offers a preview that conventional measurement cannot.

Frequently Asked Questions

What does "task crossover" actually mean, and why does it matter?

Task crossover is what happens when a worker in one profession uses AI to do tasks that traditionally belong to a different profession. A financial analyst using ChatGPT to write marketing copy, or a legal worker using it to perform financial calculations — both are examples. It matters because at 43.5% of occupation-specific AI use, it suggests AI is enabling workers to expand their effective skill sets across role boundaries, which could eventually reshape how companies write job descriptions, build teams, and make hiring decisions. The findings are detailed in the OpenAI Work at the Frontier report published July 27, 2026.

Which job roles are most affected by AI task crossover?

Customer experience workers lead, with 77% of their occupation-specific AI prompts involving tasks from another profession. Designers (75%) and HR professionals (69%) follow. Legal workers come in at 56% and marketers at 53%. Engineers show the lowest rate at 28%, though engineering tasks spread most widely to other workers — showing up in 7.4% of non-engineer messages, according to Axios.

Does this mean AI is eliminating specialist jobs?

The study does not establish that. It confirms that workers are using AI to do specialist-adjacent work, but cannot determine whether that output matches specialist quality, whether it is changing hiring decisions, or whether companies are reducing specialist headcount as a result. Independent research from the Dallas Federal Reserve Bank found that wages in AI-exposed sectors — like computer systems design, which saw 16.7% pay growth against a 7.5% national average — have held or risen even as employment in those sectors slightly declined. The picture is task reorganization rather than wholesale specialist displacement, at least in the current data.

Is AI reversing the historical trend toward deeper job specialization?

That is the larger structural question the OpenAI data raises without yet answering. Since Adam Smith identified specialization as the engine of industrial productivity, every major technology wave has on net deepened specialization. The task crossover finding is evidence that AI may be creating a mechanism to reverse that trend — enabling workers to access adjacent expertise at low cost, reducing the organizational value of specialization. Whether it will ultimately produce shallower specialization across the economy, or be offset by new forms of expertise deepening, is the research question this data opens rather than closes.