AI helped solo professionals match two-person teams in a 776-person experiment, suggesting success in an AI-shaped workplace may depend on learning to cross expertise boundaries

The most useful question about succeeding in an AI-shaped workplace may not be whether you can write a clever prompt. It may be whether you can use the system to think beyond your usual professional lane without surrendering the effort that turns assistance into learning.

A large field experiment offers unusually concrete evidence for that distinction. In a pre-registered study of 776 professionals at Procter & Gamble, researchers asked participants to work on real product-development challenges. People were randomly assigned to work alone or in pairs, and with or without access to a generative AI tool.

The headline finding was striking: individuals working with AI matched the performance of two-person teams working without it. But the more interesting result was what happened to expertise. Commercial specialists normally leaned towards commercial ideas, while research and development specialists favoured technical ones. With AI, both groups produced more balanced proposals.

That points to a less obvious version of success in an AI-heavy world. The advantage may go not simply to people who produce more, but to people who can reach across the borders of what they already know.

AI acted less like a shortcut and more like a second perspective

Teams traditionally matter because another person brings knowledge, questions and assumptions that we do not have. The P&G experiment suggests an AI system can reproduce some of that benefit in a tightly defined setting.

According to the researchers, teams without AI scored 0.24 standard deviations above individuals without AI on solution quality. Individuals using AI improved by 0.37 standard deviations, while teams using it improved by 0.39. People also worked faster: individuals with AI spent 16.4% less time than the control group, and AI-assisted teams spent 12.7% less.

The results were not just about speed. The tool appeared to help people combine technical and commercial thinking, a role usually filled by cross-functional colleagues. The authors describe it as a “cybernetic teammate”, although that phrase should not be mistaken for proof that software can replace the social, political and practical work of a real team.

The Harvard Business School AI Institute’s summary notes that AI users also reported more positive emotion and less anxiety or frustration during the task. That is intriguing, but it was self-reported and tied to one experiment. It does not tell us what months or years of AI-mediated work will feel like.

The scarce skill may be knowing how to widen a problem

Much of the public conversation treats AI fluency as a technical skill: choosing a model, constructing prompts or automating a workflow. Those abilities matter. Yet the P&G result suggests another skill is becoming valuable: using an AI response to expose the missing side of a problem.

A product researcher might ask what would make an idea understandable at the shelf. A commercial specialist might ask which material or engineering constraint would make a promise impossible. A manager might ask what evidence would change the recommendation. These are not requests for prettier prose. They are attempts to force a problem out of the groove created by one person’s training.

That is different from accepting the first plausible answer. Silicon Canals recently examined why critical thinking still matters when AI adoption moves quickly. The new point here is complementary: judgement is not only a final safety check. It determines which alternative perspective you ask the system to supply in the first place.

Assistance becomes learning only when the person stays engaged

A second, newer experiment makes this distinction clearer. A 2026 NBER working paper randomly gave 1,174 adults aged 25 to 45 access to a generative AI assistant while they completed a workplace-style business problem. Participants then completed a follow-up module without AI.

AI improved performance across education groups and reduced the initial performance gap between the higher- and lower-education groups by roughly three quarters during the assisted task. When the tool was removed, previously assisted participants did not perform worse than the control group, and the lower-education group retained part of its gain. However, much of the education gap returned.

The most useful detail concerned effort. Heavy use of the assistant improved performance on the AI-supported task even when a participant invested little effort. Better performance after the AI disappeared was associated with intensive AI use combined with sustained personal effort.

That is an important warning against confusing rented performance with acquired capability. A polished answer can be useful in the moment, but it does not necessarily leave the user more capable the next morning. If someone wants AI to support a career rather than merely complete a queue, the learning has to happen while the tool is present.

A practical definition of success is beginning to emerge

These studies do not provide a universal formula. One involved product innovation inside a single large company. The other used an online business exercise rather than observing careers over time. Both tested particular systems on particular tasks. AI tools are also changing faster than most workplace research can be completed.

Still, they support a grounded working definition. Successful AI use means expanding the range of perspectives available to you, interrogating the output with domain knowledge and remaining mentally involved enough to learn from the exchange.

In practice, that can mean asking the system to take the view of a function absent from the room, then checking its claims against a reliable source or colleague. It can mean writing your own diagnosis before requesting alternatives, so you can see where the machine changes your reasoning. It can also mean reviewing the final work without AI and explaining why each important decision survived.

None of those habits requires treating AI as an oracle or an enemy. They treat it as an unusually fast source of options whose value depends on the person framing, testing and absorbing them.

If AI makes adequate output cheaper, then output alone becomes a weaker signal of ability. The stronger signal may be the quality of the question, the range of the reasoning and the ability to stand behind the final decision. That is a demanding standard, but it is also a more human one.