OpenAI Work at the Frontier: AI expands job roles

On 27 July 2026, OpenAI published the first report in its Work at the Frontier series, examining how AI changes not only how work is performed but the boundaries among roles. The analysis covers a random sample of more than 800,000 work-related messages from U.S. ChatGPT users. The authors call the observed pattern task crossover: a worker uses AI for an activity historically associated with another occupation.

The most frequently quoted figure, 43.5%, needs context. It is the share among messages classified as occupation-specific after generic work is excluded. Across the full sample of work-related messages, 16.8% were classified as crossing an occupational boundary, 21.8% as within the user’s role and 61.5% as generic work such as writing, summarising or scheduling.

The report does not show that 43.5% of jobs will disappear, establish a productivity increase or demonstrate that a user completed a specialist task correctly. OpenAI explicitly describes the research as an analysis of tool use, not an estimate of employment or productivity effects. That distinction matters for any company planning an AI strategy.

How the study was constructed

The researchers examined users in eight occupation groups: customer experience, design, engineering, finance, human resources, legal, marketing and sales. Occupation was derived from Department or Role information provided through ChatGPT Business, while the analysed work-related messages came from those users’ individual ChatGPT accounts.

According to the report, ChatGPT anonymously classified each message and researchers never read the underlying messages. The selected message and up to nine previous messages from the conversation were used as classification context. The unit of analysis was a message, not an hour of labour, completed project or business result.

Tasks were mapped to O*NET, the occupational and work-activity database sponsored by the U.S. Department of Labor. Classification was hierarchical: a message received an intermediate work activity, then a detailed activity, and that activity was compared with the user’s historical occupational boundary. Broad work appearing across many jobs was separated from “within occupation” and “cross-occupation” tasks.

This method is more concrete than asking a user whether AI helps them work outside their role, but it has limitations. Occupational boundaries are a model based on historical O*NET descriptions. One message may contain several tasks while classification chooses a primary activity. The classification model can also be wrong.

The principal results

Across the full work-related sample:

61.5% of messages were generic work;
21.8% fell within the user’s historical occupational boundary;
16.8% concerned work associated with another occupation.

After generic work was removed, the task-crossover share reached 43.5%. In five of eight groups, a majority of occupation-specific messages crossed the traditional boundary: 77% in customer experience, 75% in design, 69% in human resources, 56% in legal and 53% in marketing.

That does not mean three quarters of all design work was transferred to other occupations. The denominator contains only occupation-specific messages after the broad generic category is excluded. It is a strong example of how a headline without methodology can distort the study.

Which tasks travel among roles

OpenAI distinguishes two directions. A worker can bring in tasks from another occupation, while the traditional tasks of an occupation can also appear widely among people in other roles.

Marketing and engineering are frequent sources of tasks performed outside those functions. Repeated activities included calculating financial data, troubleshooting applications and systems, creating marketing materials and explaining regulations. Marketing work accounted for approximately 28–29% of non-generic messages among sales and design users and 26% among customer-experience users. Engineering tasks accounted for 28% among design users and roughly 20–22% among customer experience and finance.

Design shows the inward direction: 35.2% of all messages from designers involved work associated with other fields, while design work represented only 1.7% of messages from users in the seven other groups. Engineering was closer to the opposite: 18.5% of engineering messages crossed into other fields, but engineering tasks accounted for 7.4% of messages among workers elsewhere.

Marketing scored highly in both directions. Marketers devoted 24.3% of their messages to work linked with other occupations, while marketing tasks represented 8.9% of messages in other groups. The report therefore points less to simple automation and more to a recombination of responsibilities.

Small firms and AI as a generalist tool

Among typical-volume users—the middle 50% by message volume—the cross-occupation share declined from 18.9% in workspaces with 2–5 seats to 16.3% in workspaces with more than 100 seats. OpenAI offers a cautious interpretation: workers in smaller organisations may turn to AI when a specialist is not readily available.

The relationship was not monotonic among the highest-volume quartile. Workspace seats are also not total company headcount; the measure may proxy for industry, maturity, organisation design and purchasing patterns.

The finding is nonetheless operationally relevant to small businesses. AI can reduce the cost of a first attempt at data analysis, contract review, campaign design or technical troubleshooting. It does not automatically reduce the cost of error, particularly when a person operates outside their expertise and does not know what they have missed.

Task crossover does not transfer accountability

When a marketer writes a script, a salesperson analyses customer data or a business owner reviews a contract with AI, the organisation still needs an accountable person for quality and compliance. A model can help perform a task; it does not grant formal authority, a professional licence or domain expertise.

Companies need at least three operating tiers:

self-service for reversible work that is easy to verify;
mandatory specialist review for material financial, legal, security or privacy risk;
no autonomous execution for regulated or irreversible decisions.

A marketing draft can be corrected before publication. A script that modifies production data, an interpretation of a legal obligation or an employment decision needs a different gate. AI’s value does not have to come from removing the specialist. It may instead move the specialist away from every first draft and towards difficult-case review.

Security risk when roles expand

Task crossover directly affects access-control design. A user begins to perform activities that their IAM, DLP and approval processes did not anticipate. AI may generate code, but the user should not automatically receive a production token. It may support financial analysis, but that does not require access to the entire accounting system.

Common risks include:

placing data in an unapproved tool;
executing code without review or dependency scanning;
relying on a fabricated source;
exposing information outside its established team boundary;
bypassing procurement or legal review;
accepting an output beyond the user’s competence.

Controls should follow the activity and data, not only the job title. Useful mechanisms include tool policies, isolated test environments, source allowlists, output scanning and action logging without excessive content retention. Our analysis of OpenAI privacy controls, PII and GDPR covers the data layer.

How a company should measure impact

The report does not measure productivity, correctness or whether a generated result was used. A company should not cite 43.5% as its own ROI. It needs local metrics for defined workflows:

time to task completion;
outputs accepted without substantial correction;
defects found by a specialist;
escalation and blocked-action rates;
review cost;
security or privacy incidents;
quality against the pre-AI baseline.

The strongest pilot selects one activity, includes a comparison and defines a stopping criterion in advance. A broad “number of AI users” metric can rise while the business result deteriorates. An AI adoption scorecard should combine usage with quality, risk and savings.

A new skills model

If task boundaries are moving, AI training cannot stop at prompt writing. Workers need to recognise when an activity leaves their domain, how to verify evidence, when to call a specialist and which data must not be submitted.

Specialists, in turn, need efficient ways to review material prepared by someone outside their function. That differs from producing it from scratch: reviewers need checklists, provenance, change history and explicit accountability.

An organisation can create a catalogue of approved use cases containing an owner, model, data class, required reviewer and exit criterion. Such a catalogue helps scale AI-agent business automation without turning every employee into an uncontrolled administrator across several domains.

An example of safe task crossover

Suppose a salesperson wants to analyse a CRM export without waiting for an analyst. A safe workflow does not begin by uploading the full file to a public chat. The system creates a constrained view without unnecessary fields, removes personal identifiers and makes it available inside an approved AI environment.

The model may propose code and a visualisation, but the code runs in a read-only isolated sandbox. The result exposes the query, transformations and column provenance. If the analysis will affect pricing, customer segmentation or an automated campaign, the data owner or analyst approves the rule before use.

This process lets the salesperson perform part of an analytical task without administrative access or the removal of specialist oversight. The business receives a faster first answer, while the reviewer receives reproducible material. That is the difference between expanding a role and transferring accountability without control.

The same pattern can be adapted for contract drafts, administrative scripts or HR materials: minimum necessary data, a low-impact environment, visible sources and a gate before any high-impact decision.

Limitations that matter

OpenAI identifies several important constraints:

the sample is not representative of the entire U.S. workforce;
it covers eight selected occupation groups;
role information is self-reported;
ChatGPT Business and Enterprise are distinct product populations, so the results should not be generalised automatically to Enterprise;
a message does not establish completion or correctness;
the study is descriptive and cannot show how the same person would have worked without AI.

This is valuable behavioural evidence from a production tool, but it is not a causal experiment. It shows that people ask AI to support a wider range of work than their historical occupation description would predict. It does not establish whether organisations will retain that division, employ more generalists or require more specialist review.

Bottom line

Work at the Frontier provides unusual data on how AI moves tasks among roles. The central lesson is not the 43.5% figure alone, but the gap between attempting an activity and being accountable for the result. AI makes it easier to enter another domain; it does not automatically supply quality, authority or risk awareness.

For companies, that means redesigning workflows, permissions and training around activities. The useful question is not “how many people use AI?” but “which tasks have moved between roles, who verifies them and is the result better without increasing risk?”.

Breachroad helps design secure AI workflows, test agents and control data flows before broad deployment.

Primary sources: OpenAI — How AI is expanding what people do at work, OpenAI Economic Research — full Work at the Frontier report, O*NET — occupational and activity data.