A signal from everyday AI use

The debate over artificial intelligence and employment is often framed as a contest between workers and machines: which jobs will disappear, and how quickly? New research from OpenAI points to a more immediate organisational change. AI is helping people perform tasks that have traditionally belonged to other functions, potentially changing the practical boundaries of jobs well before employers rename positions or revise formal job descriptions.

OpenAI Economic Research analysed a random sample of more than 800,000 work-related messages from US users whose self-reported roles were linked through ChatGPT Business. It found that 16.8% of all messages concerned tasks historically associated with another occupation. Once broadly shared activities such as writing, summarising and scheduling were excluded, 43.5% of occupation-specific messages were classified as crossing an occupational boundary.

That does not show that AI has replaced a job, or even that a worker completed the task without expert assistance. It does, however, show where workers are seeking help. A sales employee may explore customer data, a designer may create promotional material, or a customer-experience worker may troubleshoot a software problem. The practical effect can be to reduce the number of handoffs required to advance a piece of work.

Jobs are bundles of tasks, not fixed containers

The distinction matters because occupations contain varied activities. A legal professional may draft, research, communicate, manage documents and coordinate with clients. A marketer may analyse data, develop creative material, plan campaigns and work with digital tools. The effects of AI on each part of those roles can differ sharply.

OpenAI’s research uses the US Department of Labor’s O*NET task descriptions as a historical baseline. It identifies whether the task in a message is generic, within the user’s stated occupation, or most closely associated with another occupation. By that measure, customer-experience, design and human-resources users showed especially high shares of cross-occupation activity among non-generic messages. Marketing and engineering tasks appeared especially widely in the AI use of people in other fields.

The pattern supports a task-based view of technological change. AI can automate selected activities, improve the speed or quality of others, and make it easier for a non-specialist to attempt work that previously required a specialist’s time. None of those outcomes necessarily eliminates the underlying occupation. But together they can change the mix of work expected from an employee.

This is also why a job title is becoming a less reliable shorthand for what an individual actually does. A firm may retain the same formal roles while shifting more first-draft analysis, routine troubleshooting, content production or information gathering towards employees closest to a customer or operational problem.

Smaller teams may feel the change first

The research suggests that task crossover is more common among typical-volume users in smaller ChatGPT Business workspaces. For these users, cross-occupation tasks accounted for 18.9% of messages in workspaces with two to five seats, compared with 16.3% in workspaces with more than 100 seats.

The difference is modest and does not establish a cause. Still, it fits a familiar business reality: smaller organisations often have fewer dedicated specialists and less capacity for internal handoffs. A general manager may need basic marketing copy, a financial calculation, a policy explanation or technical guidance before an external adviser or specialist can be engaged. Generative AI can lower the cost of producing an initial answer or draft.

For small businesses, this may make employees more versatile and shorten response times. For larger companies, the same tools could reduce bottlenecks between functions. But the benefit depends on whether managers redesign processes around the technology rather than merely adding another tool to existing work.

Broader scope raises new management questions

Task crossover should not be mistaken for permission to dissolve professional boundaries. Some work is relatively safe to broaden: drafting a first version of a customer email, turning meeting notes into a project plan, or generating possible marketing concepts. Other work involves regulated advice, confidential information, security-sensitive systems or decisions that require formal authority and accountable judgement.

The central management question is therefore not simply whether AI can produce an answer. It is which employee may use the answer, under what controls, and who remains responsible for the decision. A worker using AI to prepare a contract summary is not necessarily qualified to give legal advice. An employee who can generate a financial model still needs to understand its assumptions, validate inputs and escalate issues appropriately.

Organisations will need clearer operating rules in at least four areas:

  • Decision rights: Define which decisions can move closer to the frontline and which require specialist review.
  • Quality assurance: Build checks for factual accuracy, calculations, bias, security and compliance before AI-assisted work is used externally or in consequential decisions.
  • Skills development: Train workers not only to prompt AI, but also to frame problems, recognise weak outputs, verify evidence and know when to involve an expert.
  • Performance and pay: Reassess how roles are evaluated when employees take on a wider mix of tasks, including whether broader responsibility is recognised in workload, progression and compensation.

Without these changes, AI may simply transfer hidden work and risk onto employees. A company can appear more efficient because fewer requests reach specialists, while quality-control burdens and ambiguity rise elsewhere in the organisation.

Productivity is possible, but not automatic

The wider evidence supports caution. Research reviewed by the OECD finds that generative AI can improve performance on particular tasks, especially where work is text-intensive and clearly defined. Yet outcomes depend on the tool’s capability, the complexity of the work, the user’s skill and the way AI is incorporated into a process. A faster draft does not automatically create a faster or better end-to-end service.

The International Labour Organization has similarly argued that job transformation is the more likely aggregate outcome of generative AI exposure, since most occupations combine tasks that continue to require human input. Its more recent review of empirical research notes that large-scale displacement has so far remained limited, while measurable productivity, employment and earnings effects at scale have been uneven.

That does not rule out job losses in particular occupations or firms. US employment projections, for example, expect automated systems, including AI, to contribute to lower employment in office and administrative support work over the 2024–34 period, while demand for roles involved in developing and integrating AI is expected to support growth in computer and mathematical occupations. The labour-market outcome will vary with the task, sector, business model and pace of adoption.

Treat usage data as an early indicator, not a verdict

OpenAI’s findings are useful because they observe requests made in the course of work rather than only estimating what an AI system could theoretically do. They may capture the experimental phase of job redesign: workers are testing new combinations of tasks before organisations make those changes official.

But the study has important limits. Its sample is not representative of the entire US workforce, covers eight occupational groups, relies on self-reported role information and is limited to a specific population of ChatGPT Business-linked users. It measures messages rather than completed projects, working hours, quality, time saved or employment outcomes. It cannot establish whether a worker would have done the task without AI, whether a specialist later reviewed it, or whether the AI-generated output was adopted.

The appropriate conclusion is not that occupational boundaries have already disappeared. It is that they are becoming more permeable in some AI-enabled workflows. For businesses, the strategic challenge is to decide where that permeability creates useful autonomy and speed, and where it creates unacceptable risk.

The companies most likely to benefit will not be those that treat AI as a simple substitute for labour. They will be those that deliberately redesign task allocation, preserve expert oversight where it matters, and help workers build the judgement needed to operate across a broader range of work.

Sources