OpenAI published research on 27 July analysing more than 800,000 ChatGPT messages from US business users, and reported that 43.5% to 44% of occupation-specific requests involved tasks associated with a different occupation. The paper is titled "Work at the Frontier: How AI is Expanding What People Do at Work."

The finding

Generic tasks — email, scheduling, summarising — were explicitly excluded from the headline figure, which is the methodological choice that makes it interesting: this is not people using a chatbot for admin, it is people in one role asking for work that belongs to another. The highest rates of crossing appeared in customer service, design and human resources. OpenAI's chief economist Ronnie Chatterji said the boundaries between jobs are "likely already becoming more flexible due to AI."

What it is not

It is not labour-market data. This is message classification: role-crossing is inferred from the content of requests, not verified against job titles, employment records or output. Nothing in the method establishes whether the designer asking a legal question got a usable answer, whether the work product was ever used, or whether a single job changed as a result. A count of attempts is not a measure of capability, and it is certainly not a measure of employment.

Who is holding the ruler

OpenAI measured its own product's telemetry and published it as evidence about the economy. The finding also happens to support the company's preferred narrative — that AI expands what workers do rather than displacing them — against the version in which it removes roles. That does not make the number wrong. It does mean the number was produced by the only party with access to the data and a strong interest in its interpretation, with no independent replication possible.

Why it is still worth reading

Because 800,000 messages is a genuinely large sample of how AI is used at work, and nothing comparable exists in public. Surveys ask people what they think they do; this counts what they typed. The right way to hold it is as the best available description of demand — what workers are reaching outside their role to attempt — rather than as a finding about what the labour market has become.

Timing note

OpenAI's own feed dates this publication to 03:30 UTC on 27 July, slightly ahead of our usual 24-hour window; it is included here because no equivalent dataset has been published.