The most useful question about AI and work may not be which jobs disappear. That is a lagging indicator. An earlier change is already visible: salespeople analyze data, designers troubleshoot software, and small-business owners draft copy or perform basic financial analysis before any company rewrites their job descriptions.
OpenAI Economic Research calls this pattern task crossover. In a study of more than 800,000 work-related ChatGPT messages from U.S. users, 16.8% of all work-related messages concerned tasks historically associated with another occupation. After broadly shared activities such as writing, summarizing, and scheduling were excluded, the share rose to 43.5% of occupation-specific messages.
That does not mean AI has completed 43.5% of another profession’s work. It means users are asking AI for help beyond the historical boundary of their stated role. The distinction turns a dramatic statistic into a useful leading indicator.
Reading time: about 9 minutes | Length: about 1,900 words
TL;DR
- Cross-occupation tasks account for 16.8% of all work-related messages and 43.5% only after generic work is removed.
- The study measures messages and attempted activities, not completed tasks, productivity, or job losses.
- Task movement is directional: design imports tasks, engineering exports them, and marketing does both.
- As execution spreads across roles, review capacity, acceptance authority, training, and accountability become the new organizational constraints.
What the 43.5% denominator includes
The OpenAI report divides the random sample of work-related messages into three groups:
| Classification | Share of all work-related messages |
|---|---|
| Generic | 61.5% |
| Within occupation | 21.8% |
| Cross-occupation | 16.8% |
Generic activities are common across many jobs and therefore reveal little about occupational boundaries. The report includes writing emails and scheduling meetings in this category.
When generic messages are removed, within-occupation and cross-occupation messages are rebased to 100%. Cross-occupation messages then account for 43.5% of that smaller, occupation-specific denominator.
Both numbers are valid. They answer different questions:
- 16.8% asks how much of all observed work-related use crosses a historical occupational boundary.
- 43.5% asks how much of occupation-specific use crosses that boundary after shared work is removed.
Writing “nearly half of work-related ChatGPT use is outside the user’s job” would be wrong. Nearly half of the non-generic, occupation-specific subset is outside the historical boundary.
What the study measured
The analysis covers more than 800,000 work-related messages from individual ChatGPT accounts belonging to U.S. users. Their occupation came from self-reported Department or Role information supplied through ChatGPT Business.
OpenAI mapped each selected message to one primary O*NET Detailed Work Activity. It then compared that activity with the historical task boundary associated with the user’s occupation. The study covered eight groups: customer experience, design, engineering, finance, human resources, legal, marketing, and sales.
The unit of analysis is a message. It is not an hour of work, a completed task, a project, or a job.
The report cannot observe:
- whether the AI output was used;
- whether the output was correct;
- how much time it saved;
- whether the user could have done the task without AI;
- whether a specialist reviewed it;
- whether the organization formally changed the user’s responsibilities.
OpenAI explicitly describes the findings as descriptive. The study does not estimate causal effects on employment or productivity, and its sample is not representative of the entire U.S. workforce.
Task networks move before titles do
The data still reveals something conventional labor statistics see late.
Job descriptions and occupational taxonomies record how work has been organized. AI usage records what people are trying now. A worker can experiment with a new task bundle long before HR changes a title or a company creates a new role.
This makes task crossover a plausible early signal of role change:
- AI lowers the execution cost of an unfamiliar task.
- The person closest to the problem tries it instead of handing it off.
- Repeated successful attempts become an informal responsibility.
- Review processes, incentives, and titles adjust later, if they adjust at all.
Only the first two steps are directly visible in the message data. The later steps are an organizational hypothesis, not a result proven by the study.
Task movement has a direction
Cross-occupation work is not a uniform exchange. Some roles import many tasks, while other roles supply tasks that spread across the organization.
The official OpenAI article highlights three patterns:
- Design imports tasks. About 35.2% of messages from designers involve work associated with another occupation, while design tasks account for only 1.7% of messages from workers in other fields.
- Engineering exports tasks. About 18.5% of engineering messages involve other occupations’ work, while engineering tasks account for 7.4% of messages among workers elsewhere.
- Marketing moves both ways. Marketers devote 24.3% of messages to other occupations’ tasks, while marketing tasks make up 8.9% of messages among workers in other fields.
These figures use different denominators and should not be subtracted from one another. Their value is directional.
Designers in the sample use AI to combine activities from several fields. Engineering tasks such as troubleshooting software travel outward. Marketing both absorbs outside work and supplies tasks that other roles perform.
This is closer to a changing task network than a list of occupations being replaced.
Small workplaces reveal the handoff economics
Among users in the middle 50% by message volume, cross-occupation messages account for 18.9% of work-related use in workspaces with two to five seats and 16.3% in workspaces with at least 101 seats. The same monotonic pattern does not appear among the heaviest users.
The report offers a plausible explanation: workers in small organizations have fewer specialists nearby, so they may use AI when they encounter a task that otherwise requires another function.
This remains an inference. Workspace seats are not total company size, and the difference is modest. But the pattern points to an economic mechanism worth testing.
A handoff has costs:
- describing the problem;
- finding the right specialist;
- waiting in a queue;
- transferring context;
- reviewing the result;
- resolving responsibility when something goes wrong.
AI makes the first attempt cheap. When the cost of trying is lower than the cost of a handoff, more workers will attempt adjacent tasks themselves.
That is why AI can create more generalist behavior without eliminating specialists. Specialists may receive fewer routine requests while taking on more difficult review, exception handling, and standard-setting work.
Execution spreads, but acceptance remains scarce
Broader capability is not broader competence
The amplifier effect has a dangerous edge. AI can help a non-specialist produce an answer before that person has the knowledge required to evaluate it.
A salesperson can generate a SQL query without understanding data lineage. A founder can draft a contract without recognizing an unenforceable clause. A marketer can change a website without understanding its security model.
The task has crossed a boundary, but accountability has not disappeared.
Organizations therefore need two definitions:
- Execution permission: who may attempt a task with AI?
- Acceptance authority: who may decide that the result is safe and complete?
Low-risk drafts can have broad execution permission and lightweight review. Production changes, financial decisions, legal commitments, or access-control modifications need named acceptance authorities and explicit escalation paths.
Without this split, task crossover can create invisible role expansion. Employees take on more work, but evaluation criteria, training, pay, permissions, and liability remain anchored to the old job.
The new bottleneck is review capacity
AI reduces the cost of producing first attempts. It does not reduce the cost of every kind of validation.
If ten employees can suddenly draft analyses, queries, campaigns, and contracts, the organization may create more outputs than its experts can safely review. The bottleneck moves from production to acceptance.
This has three consequences.
First, expertise becomes more leveraged. A specialist can define templates, tests, policies, and reusable review criteria that guide many AI-assisted workers.
Second, weak review systems become more expensive. More attempts mean more opportunities for plausible errors to reach production.
Third, performance management must separate activity from accepted outcomes. Counting generated artifacts rewards volume exactly when judgment becomes scarce.
The useful organizational question is no longer only “Can this role do the task?” It is “Can the system detect when the task was done badly?”
What companies should measure
The OpenAI study is a map of usage, not an operating dashboard. Organizations can make the signal actionable by tracking a small set of internal measures:
Boundary-crossing rate: What share of AI-assisted tasks sits outside the worker’s documented role or training?
Acceptance path: Which outputs are self-approved, peer-reviewed, or approved by a domain specialist?
Rework and incident rate: Which cross-role tasks create corrections, rollbacks, customer complaints, or compliance issues?
Handoff displacement: Did AI remove a wait, or did it merely delay specialist involvement until later?
Capability conversion: Which repeated task crossovers have become reliable enough to justify training, permission changes, new performance expectations, or a redesigned role?
These metrics should be segmented by risk. A cross-role presentation draft and a cross-role production database change should not share one success threshold.
What this means for workers
The report does not prove that becoming a generalist is always better. It suggests that AI makes it easier to test adjacent responsibilities.
A durable strategy has three parts:
- learn enough of the neighboring domain to evaluate common failure modes;
- use AI to reduce execution friction;
- preserve access to specialists for high-risk review and edge cases.
The advantage comes from connecting domains while respecting their boundaries. Someone who can frame a business problem, generate a first analysis, and recognize when the answer requires a statistician has more leverage than someone who merely produces a plausible chart.
Task crossover rewards judgment, not indiscriminate scope expansion.
Conclusion
The study’s strongest result is not a forecast of job loss. It is evidence that the task list itself may be moving.
Across the sample, 16.8% of all work-related messages crossed a historical occupational boundary. Among the occupation-specific subset, the share was 43.5%. Marketing and engineering tasks traveled widely, some roles imported more activities than they exported, and moderate users in smaller workspaces showed somewhat more crossover.
The study cannot tell whether those attempts were successful, productive, or formally recognized. That limitation is also the management agenda.
As AI lowers the cost of attempting unfamiliar work, companies need to redesign review, permissions, training, incentives, and accountability around moving tasks. Jobs may be rewritten later. The operational change begins when the handoff stops.
The next action is to audit one recurring AI-assisted task that crosses a role boundary, then name its execution permission, acceptance authority, and failure signal.
FAQ
Does the study show that AI is replacing 43.5% of jobs?
No. The figure is the share of non-generic, occupation-specific messages classified as tasks associated with another occupation. The study does not measure job replacement or employment effects.
Why are both 16.8% and 43.5% reported?
16.8% uses all work-related messages as the denominator. The 43.5% figure removes generic activities and rebases the occupation-specific subset.
Does a cross-occupation message mean the task was completed?
No. The unit is a message. The study does not observe whether the output was used, correct, reviewed, or productive.
Which tasks spread most widely?
The report finds that marketing and engineering tasks appear frequently in other occupations’ messages. Financial calculation and technology troubleshooting also recur across multiple groups.
What should employers do first?
Identify cross-role AI use, classify it by risk, and define who may execute and who must approve. Then measure rework, incidents, and whether repeated crossover justifies formal role redesign.