OpenAI task crossover: 43% of work prompts cross roles

OpenAI published Work at the Frontier on July 27, and the finding it leads with is task crossover — work that historically belonged to one occupation showing up in the AI use of people in another. Across more than 800,000 work-related messages from U.S. ChatGPT users, 16.8% of work messages and 43.5% of occupation-specific messages describe a task that belongs to a different job. If you run a ten-person company, you already knew this. Your ops person writes the ad copy. What's new is that there's a number on it, and the number has a failure mode nobody's pricing in.

What actually happened

OpenAI matched the work described in individual messages against occupational tasks in the Department of Labor's O*NET database, then bucketed each message as generic, inside the user's occupation, or outside it. Generic work — drafting email, scheduling, summarizing documents — was stripped out first. Of what remained, nearly half crossed a job boundary.

The crossover isn't evenly spread. Marketing and engineering are the most borrowed task bundles, appearing in prompts from workers across other occupations. Customer-experience roles cross the most in the other direction: north of three-quarters of their occupation-specific prompts describe another profession's work. And per The Decoder, the pattern is sharper at small companies, where non-specialists pick up marketing work because there is no marketing department to hand it to.

One caveat worth holding. This is message-level classification against self-reported roles, not outcome measurement. It tells you what people asked for. It says nothing about whether the answer was right.

Why AI task crossover matters for your business

Crossover is where your quality bar disappears without anyone noticing. When your bookkeeper writes an Instagram caption, worst case is a bad ad. When your marketer writes a query against your production database, or your ops lead accepts an AI-drafted indemnity clause, worst case is a different category of expensive. The report's own hot spots — engineering as the most-borrowed bundle, legal work crossing in over half of prompts — are exactly the two places where confident, well-formatted, wrong output costs real money.

So put three gates on the crossings, not on the tool:

1. Read-only credentials for anyone querying data through an AI tool. A separate database role with SELECT on a read replica, not the app's connection string pasted into a chat window.
2. Legal and financial output gets a signature, not a skim. Someone with the actual credential signs off by name. "I reviewed it" is not a control.
3. Log the prompt and the output where the work lands — PR description, ticket comment, doc header. Six months out, when something's wrong, you need to know whether a person or a model wrote that sentence.

With those in place, crossover is exactly the leverage a small team wants. Without them, you've quietly distributed specialist risk to people who can't see it coming.

Sources: OpenAI, The Decoder.