Call a meeting, and the ideas multiply. Open an AI, and the ideas multiply. But how much easier has it become to choose just one?
When filling a blank page takes less effort, the whole job can feel further along. That feeling is not wrong. Yet having options is different from being able to choose. A splendid restaurant menu does not mean dinner has been decided.
An experiment conducted at P&G by Fabrizio Dell’Acqua and colleagues brings this distinction into focus. It asks the broad question of whether AI can be a teammate. Look closely at the results, though, and the work of a team begins to separate into several different jobs.
Another specialty, at a table for one
The participants were professionals in research and development or commercial roles. A total of 791 people were randomly assigned to conditions for a one-day online product-innovation exercise: working alone or in a cross-functional pair, with AI or without it. Participants generated five ideas, chose one, and developed it. Experts who did not know the assigned condition evaluated the final proposals. [1]
People working alone with AI produced results comparable to pairs working without it. AI also made proposals more balanced: without AI, R&D professionals tended toward technical considerations and commercial professionals toward commercial ones.
This does not mean there is no longer any reason to consult the next department. The balance of a submitted proposal changed; its author did not become an expert in another field. Still, being able to extend an idea beyond your specialty before convening a meeting is no small opening.
We tend to bundle together the benefits of working in pairs. Filling gaps in knowledge. Spotting an omission. Offering encouragement when someone hesitates. Making a tedious task a little less lonely. Calling all of this “collaboration” is convenient, but it can hide exactly what has changed.
Making a good idea, recognizing a good idea
The part worth pausing over examines the five preliminary ideas. AI raised their average quality. But participants selected the idea that experts rated highest about 50% of the time in pairs without AI, compared with roughly 37% in the AI-enabled conditions. It is not a result that lets us say AI made selection better too. [1]
It would be equally premature to leap from here to “AI takes away judgment.” How much participants used AI during selection is not established, and the portfolios themselves differed across conditions. These were not judges being handed the same five ideas. Better candidates might also have been harder to distinguish.
Even so, a pattern emerges that final performance alone would conceal. With enough good candidates, a submission can improve even when its author misses the very best one. A better outcome and a better selection process are not the same fact.
Separating them makes the discussion of AI adoption more concrete. Alongside “How many people produced what score?” we can ask “What improved, at which stage?” If idea generation benefited, the next thing to improve may not be a system that generates still more ideas.
The unchosen ideas still have work to do
What follows is a reflection prompted by the research. When exploring a new business proposal, for example, giving creation and evaluation to the same conversational flow can blur the point at which judgment begins. As we polish a favorite idea, it starts to feel as though it must have been the best. There is no guarantee that judgment grows more reliable at the same pace as the prose grows smoother.
It may help to make a separate space, apart from generating ideas, to say what the choice is for. Prospects of selling? Speed of testing? Knowledge gained if it fails? Change the criteria, and the same ideas change places. The experiment’s expert assessments had their own particular yardstick. At work, someone must take responsibility for the yardstick too.
A brief reason for rejecting each idea might also make judgment more visible. Not “because it was bad,” but “because, under these conditions, this was missing.” If the conditions change, the reason may become useful. When only the chosen proposal gets a handsome file and everything else sinks to the bottom of a chat, some of that newfound abundance ends there.
Of course, this was one company, one day, and remote pairs whose members were mostly unfamiliar with one another. It did not measure the tacit knowledge of long-standing teams, responsibility for bringing a real product to market, or long-term learning. Whether AI can replace a colleague is not a question headcount alone can answer.
When we gain another partner for generating ideas, are we quietly treating the work of choosing as finished too?