Workers Borrow AI Expertise Across Job Roles
Job titles are a poor instrument for measuring what people actually do. They compress a shifting bundle of activities into a stable label, and the label then gets treated as if it described the work. That gap — between the name on an org chart and the tasks filling someone’s day — is where the most interesting change in AI-assisted work is currently happening, and it is precisely the gap that conventional workforce data was never built to see.
OpenAI’s economic research team has published a second installment in its “Work at the Frontier” series, and the finding worth sitting with is not that workers use AI for unfamiliar tasks. It is that they come back. Analyzing more than 1.5 million work-related messages sent through ChatGPT between April and July 2025, the researchers tracked what they call task crossover: workers reaching for activities historically attached to a different occupation. OpenAI, \"Work at the Frontier\” (second installment), 2025 The first report documented that this happens. The new one asks what follows, and the answer — partial, early, but consistent — is that some of it sticks.
Borrowing expertise rather than acquiring it
The most useful distinction in the research is not between occupations but between modes of asking. When workers prompt AI about tasks inside their own role, they write longer requests, ask for explanations, request specific formats, seek how-to guidance. Outside their role, the pattern inverts: prompts get shorter, and requests for teaching fall away. What rises instead is context. Workers supply examples, background documents, information lifted from a colleague, and they ask the model to check or verify something.
Read that combination carefully and a coherent behavior emerges. People are not enrolling in a correspondence course on marketing or contract law. They are arriving with a concrete problem and enough surrounding material to make the problem legible, then asking AI to apply knowledge from a field they do not command. The expertise is borrowed, not internalized — at least not yet. That is a meaningful difference from traditional upskilling, and it has consequences for how we think about what workers will and will not be able to do on their own five years from now.
The return rate is the real signal
Experimentation is cheap and mostly meaningless. Anyone can try a thing once. The question that separates a novelty from a workflow is whether the person shows up again.
Among roughly 6,200 workers observed consistently across the four months, cross-occupation tasks grew from 13.1 % of occupation-specific AI activity in April to 25.9 % in July. OpenAI, \"Work at the Frontier\” (second installment), 2025 That near-doubling is suggestive, though it needs a caveat the researchers are careful to supply: as workers try new things in any given month, the pool of tasks eligible to be called recurring expands, which inflates the trend somewhat. The cleaner evidence comes from a matched comparison. Workers returned to a cross-occupation task they had used the previous month 23.6 % of the time. OpenAI, \"Work at the Frontier\” (second installment), 2025 Comparable workers with no observed prior use of that same task returned to it 8.4 % of the time. That gap — roughly a factor of three — is the kind of number that survives scrutiny, because it compares like with like.
The variation across task types is where the picture gets genuinely informative. Customer discussions about goods and services recurred at 54 %. Promotional writing came back at 44 %, marketing materials at 37 %. Explaining financial information returned at about 15 %, against an 18.5 % average across all cross-occupation tasks. Something is sorting these activities, and the sorting is not random.
What the spread between 54 and 15 tells us
The tasks that stick are the ones embedded in cycles. A customer conversation happens because a customer exists and a question needs answering; it is not a project with a beginning and an end. Promotional writing recurs because campaigns recur. Financial explanation, by contrast, tends to be episodic — tied to a report, a filing, a specific moment — and it carries asymmetric downside if handled badly. A number explained incorrectly to an investor or a regulator is not a draft that gets revised; it is a liability.
So two forces are plausibly at work in the spread. One is rhythm: some activities sit inside repeating loops and some do not. The other is consequence: some domains tolerate a borrowed answer and some punish it. The study is honest that it cannot separate these, and neither should anyone reading it. What it can say is that recurrence clusters where the cost of a mediocre first draft is low and the frequency of need is high. That is a reasonable description of where AI has already proven itself in practice, and it suggests the frontier of adoption is being set less by capability than by tolerance for error.
The organizational blind spot
Here is where the findings turn from interesting to actionable, and where the gap between measurement and reality becomes a management problem rather than an academic one.
Companies implementing AI strategies tend to frame the challenge as access: which tools, which licenses, which departments. The study points somewhere else. Work design — how tasks are divided, who owns what, how responsibilities are drawn — deserves equal standing alongside tool availability. If a worker has quietly absorbed a recurring activity that belongs to no one’s job description, then the organization has a function running without an owner, an accountable person, or a review process. It works, until it does not, and when it fails there is no obvious place to look.
The same dynamic explains why the friction between identifying a problem and moving work forward can drop so sharply. A worker who needs a piece of adjacent expertise no longer has to locate the colleague who has it, negotiate their time, explain the context, and wait. The context goes into the prompt instead. That is a real gain — and it is a gain that arrives without the friction, the waiting, and the dependency that come with borrowing expertise from a colleague. What it does not arrive with is any corresponding change in how the organization maps its own work.
Broadening without renaming
The most defensible claim in the report is also the most modest. Job titles may not change while the mix of activities inside them does. A role can absorb adjacent responsibilities, become meaningfully broader, and still carry the same label on the org chart, the same salary band, and the same entry in the HR system. The transformation is real and simultaneously invisible to every instrument most organizations use to track work.
This is not a new phenomenon. Occupational boundaries have always been softer in practice than in classification. What is new is the speed and the individual initiative. Historically, task expansion happened through formal reassignment, team restructuring, or gradual drift ratified by a manager. Now it can happen in a single afternoon, at one person’s discretion, and leave no trace except a pattern in their prompts.
The research team is explicit about the limits of what it can see. Occupations are inferred from role or department information supplied during business onboarding. Messages are aggregated and anonymized, and the researchers never read individual conversations. The sample skews toward professional workers using a paid product, which is not the workforce. Four months is a short window, and the recurrence measure is confounded by the expanding pool of candidate tasks.
None of that undermines the central observation. Workers are not waiting for permission to widen their jobs. They are doing it in small increments, keeping what helps, discarding what does not, and returning to the parts that earn their place. The gain is genuine: less friction, faster movement from problem to progress, access to adjacent knowledge without a detour through someone else’s calendar. The distance between that practice and how organizations formally describe their work is where the next set of questions lives — and where the promise of AI at work will either be consolidated or quietly squandered. Whether borrowed expertise hardens into capability or settles into permanent dependency is the question the return rates cannot yet answer.
The evidence on how expertise transfers between domains under these conditions remains thin, and the occupational health literature on task expansion and role blur has been warning about overload for years. Whether borrowed expertise becomes genuine capability or remains a permanent dependency is the open question. The return rates say workers are betting it will be the former.
Sources
1. OpenAI