OpenAI's Own Usage Data Shows 1 in 6 Work Messages Crossing Job Lines

16.8% of all work-related ChatGPT messages — and 43.5% of messages tied to a specific occupation — perform tasks that OpenAI’s own taxonomy assigns to a different job than the sender’s. That’s the headline number from “Work at the Frontier,” a new OpenAI report analyzing more than 800,000 US ChatGPT messages for what it calls task crossover: work historically owned by one role showing up inside another role’s AI usage. The crossover isn’t evenly spread — it’s sharpest in customer experience (77% of messages), design (75%), HR (69%), legal (56%), and marketing (53%), functions where a lot of the underlying work is language- and judgment-heavy rather than deeply technical.

The most telling split in the data is by company size: workspaces with 2-5 seats show 18.9% crossover versus 16.3% for workspaces over 100 seats. Smaller companies don’t have a dedicated legal team or a design department to route work to, so the crossover isn’t really about AI making individual employees more versatile — it’s AI standing in for specialist headcount those companies never had.

OpenAI’s own small-business pilot data backs that reading with a number: 42% of participants in its “Small Business AI Jams” reported saving more than five hours a week, and 78% left having built a working AI workflow in a single session. And the pattern isn’t confined to small operators — OpenAI’s newsroom case studies show the same crossover at organizations with real specialist teams: the Seattle Times built an AI prospecting agent that cut sales research “from hours to minutes,” while the Daily Beast’s “Data Scouts” agents surface business opportunities that would otherwise sit with a dedicated ops function.

Read together, the three reports describe the same mechanism at different scales: AI doesn’t so much expand what one person can do as let companies quietly skip hiring for the roles it substitutes for. For companies evaluating where AI actually changes headcount plans, task crossover — not raw adoption — is the number worth tracking.