What changed
Most workplace AI studies ask whether a tool makes one person faster. This one asked whether a language model can reproduce some of the benefits of an actual teammate. Researchers worked with Procter & Gamble on a full-day product-development workshop using real business-unit problems, not toy prompts. The 791 participants were randomly assigned to one of four conditions: one person without AI, a two-person team without AI, one person with AI, or a two-person team with AI.
The outputs were not graded by the participants themselves. Each final solution was scored by blinded evaluators with business and technology backgrounds on a 1 to 10 quality scale, then standardized against the solo no-AI group. Teams without AI improved quality by about 0.24 standard deviations over the solo baseline. Individuals with AI improved by about 0.37 standard deviations, and teams with AI by about 0.39. Across 550 submitted solutions, that meant a solo worker with AI performed at roughly the same quality level as a two-person team without it.
The more interesting result was not only raw quality. Without AI, research-and-development staff leaned toward technical proposals and commercial staff leaned toward market-oriented ones. With AI, both groups produced more balanced ideas. The paper also reports that participants working with AI experienced more positive and fewer negative emotional shifts than people working alone, suggesting the system partly reproduced the motivational side of collaboration as well as the performance side.
What this could change for you
If your job involves early-stage planning, ideation, or turning a rough problem into a structured proposal, the practical lesson is not "replace the team." It is that AI can help one person get farther before the first handoff. A marketer can pressure-test a more technical idea. A technical worker can draft something that is easier for a commercial or operations partner to evaluate. That can reduce some of the waiting and translation work that usually slows projects down.
The findings also suggest a better default for managers than measuring AI by time saved alone. In this study, AI seemed most useful at broadening the starting set of viable ideas and helping people work outside their home lane. That is a good reason to use it earlier in a workflow, when the job is to widen the option set, surface tradeoffs, and prepare something another human can challenge.
The paper is also a warning against overclaiming. The authors found that AI improved idea generation more than idea selection. Human judgment still mattered when choosing which concept deserved to move forward. In practice, that means AI may be well suited to helping create the menu of options, while a team still needs to decide what is strategically sound, realistic, and worth funding.
What it does not prove
This was one company, one kind of work, and one AI setup. Participants were Procter & Gamble professionals working on early-stage product innovation with GPT-4 delivered through Microsoft Azure after a short training session. The study does not show that every organization, model, or workflow would get the same lift.
The main outcome was evaluator-rated solution quality, not customer sales, launch success, or long-term business value. Some of the best ideas entered P&G's real innovation pipeline, but the paper did not track which ones later worked in the market. A stronger workshop draft is not the same thing as a better shipped product.
Finally, the experiment does not prove that AI replaces teams in general. The strongest evidence is narrower: on this kind of creative, cross-functional drafting task, AI helped solo workers produce outputs comparable to two-person teams and helped specialists broaden their proposals. The paper itself says human evaluative judgment retained value.
The bottom line
AI did not make teamwork obsolete, but it did help solo workers act more like they had already crossed the first functional boundary. In this P&G field experiment, a language-model teammate lifted idea quality, reduced siloed thinking, and improved the emotional experience of the task. The most credible use is not final decision-making. It is helping one person arrive at a better first serious draft before the human debate starts.
Primary research
The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork
Organization Science · 2026 · DOI 10.1287/orsc.2025.20702
