When AI Blurs the Lines Between Jobs
Two very different studies landed on the same conclusion from opposite directions. A controlled experiment inside Procter & Gamble shows why it happens: give people AI and the walls between specialties fall away. A sweep of hundreds of thousands of real ChatGPT-at-work conversations shows that it is already happening at scale: people spend a startling share of their time doing other occupations' jobs. Put together, they describe a single shift — the atomic unit of work is moving from the specialist, and from the team, to the individual-plus-AI — and, as Ethan Mollick puts it, they compel organizations "to think about division of labor in a new way."
Four beats + one verdict · two studies
01The lines are already blurring
OpenAI's "Work at the Frontier" report reads the world as it is, not a lab. Analyzing more than 800,000 work-related messages from U.S. ChatGPT Business users across eight occupation groups, it asks a simple question: when people bring a task to AI, does it belong to their own job — or someone else's? The answer is that boundary-crossing is now the norm, not the exception.
Of every work message that maps to a specific occupation, 43.5% concerned a task from a different occupation. Look across whole workloads and the crossover is everywhere: designers doing engineering, marketers who aren't in finance running financial calculations, non-marketers producing promotional copy. The org chart still exists; the work no longer respects it.
Crucially, the effect is uneven and revealing. Customer-experience, design, and HR roles spend the majority of their occupation-specific effort on other people's jobs, while engineering — the most technically gated function — crosses least. AI lowers the entry cost to a specialty exactly in proportion to how much that specialty depended on generalist judgment rather than deep technical training.
Among the report's task-migration examples: ~25% of marketing-related messages come from people who aren't marketers (mostly building promotional materials); ~10% of finance-related messages come from non-finance staff running calculations; engineers are no longer the only ones troubleshooting systems. The tasks are migrating toward whoever needs them done — and AI makes that person capable enough to do them.
Share of occupation-specific work spent on other occupations' tasks
02AI collapses the silo
Field data shows the lines blurring; it can't prove AI is the cause. The P&G experiment can. In a preregistered field experiment, 776 professionals worked on real new-product-development challenges under a clean 2×2 design — individually or in a team, with AI or without — so any difference is caused by the conditions, not by who happened to use the tool. This is the finding that gives the whole thesis its spine.
What people proposed, by orientation of the idea
Without AI, people stayed in their lane. R&D professionals proposed more technical solutions; Commercial professionals leaned commercial. You could tell someone's department from their ideas. That is what a functional silo is — expertise concentrated, and confined, by role.
With AI, that signature vanished. In the authors' words, "professionals using AI produced balanced solutions, regardless of their professional background." A single person plus a model could reason across the technical and commercial dimensions that used to require two different specialists in the room.
This is the causal core of Mollick's claim. The field report shows boundaries thinning; the experiment shows AI doing the thinning. When the tool supplies the expertise a role used to monopolize, the wall between roles stops doing much work.
03One person, a whole team
If AI can supply a missing specialty on demand, the most basic building block of org design — putting two specialists together so their expertise combines — starts to look optional. The experiment measured exactly this, scoring solution quality against a solo worker with no AI. The result is the single most quoted number in the study, and the one that most directly rewrites the division of labor.
A two-person team without AI beat the solo baseline by +0.24 SD — the measurable value of collaboration. An individual working with AI beat it by +0.37 SD: not just matching the team, but edging past it. Adding both — a team and AI — reached +0.39 SD, barely above the solo-with-AI result.
Read the gap between the second and third bars: almost all of the benefit a team provided, one person with AI now captures alone. The model behaves like a teammate — supplying the perspective, the push-back, and the missing expertise that a second human used to bring.
If a solo worker with AI performs like a cross-functional pair, the reason organizations bundle specialists into teams — to combine expertise — weakens. The unit of production shifts from the team to the individual + AI, with humans reserved for what AI still can't do: choosing which idea is actually right (see Beat 04).
Solution quality vs. a solo worker with no AI (standard deviations)
04Better, faster — and blind to it
The silo-collapse isn't a trade-off against quality — it comes with it. Working with AI, the same people produced better ideas, in less time, while feeling better about the work. But one gap is worth dwelling on: their own sense of how well they'd done didn't keep up with how well they actually did.
What AI added
- Higher average quality — and far more exceptional solutions in the top tier
- Longer, more detailed write-ups (hundreds more words per solution)
- A more positive emotional experience — closer to working with a person than a tool
- Balanced, cross-functional thinking from single specialists (Beat 02)
What humans still owned
- Judgment: AI lifted the quality of ideas generated; people still did the choosing
- Final solutions stayed closer to human work than to raw AI output
- Perception lagged reality — AI users under-rated their own (improved) results
- Deciding which of many good options was actually right for the business
AI's main contribution was to lift the quality of ideas generated, shifting the whole distribution upward — while the value of human judgment in selecting among them remained. That division — machine widens the funnel of options, human picks — is the shape of the new labor split the studies point toward: not human or AI, but a re-drawn line between generating and deciding.
The division of labor, rewritten
Two studies, two methods, one conclusion. The OpenAI report is a wide-angle photograph of work as it is actually being done — and it shows people already reaching across occupational lines at scale. The P&G experiment is a controlled close-up that explains the mechanism — hand someone AI and the walls between specialties fall away, so a single person now carries the expertise a team used to assemble. The field shows the footprint; the lab shows the foot. That convergence — an observational census and a randomized experiment agreeing — is exactly why Mollick treats the finding as more than a one-off.
Flatter orgs, fewer handoffs, empowered generalists.
If one person with AI can think across technical and commercial lines and match a team, work can be organized around problems rather than departments. Fewer coordination costs, fewer handoffs, faster cycles — and specialists freed to go deeper where depth still wins.
The design question becomes: which teams exist to combine expertise (now partly automatable) versus to combine judgment and accountability (still human)?
If anyone can do anyone's job passably, what happens to depth?
Porous walls cut both ways. When boundary-crossing is cheap, the incentive to cultivate deep expertise may erode — even though the studies show humans still own the judgment that separates a good option from the right one.
The task ahead isn't to remove the walls but to redraw them: around generating vs. deciding, breadth vs. depth, and the coordination that only people can be accountable for.