Every innovation team has the same foundation models and prompt libraries. The results are wildly uneven. A working paper invited by the International Journal of Research in Marketing, by researchers at Harvard, Wharton, Northwestern and Columbia, argues that the reason is not the model but the human bottlenecks buried inside the innovation process. Across four stages — ideation, screening, consumer insight and market learning — AI’s default effect, absent deliberate redesign, is to deepen the bottleneck rather than shrink it.
In ideation, the model gravitates toward the statistically typical response, and once a person reads it they fixate and stop producing the odder ideas they might have had. Because most models draw on similar training data, independent teams converge on the same handful of design directions: individual productivity rises, collective diversity falls. Prompting the model to revise toward bolder territory works. Telling the fixated human to think more broadly does not.
Screening panels have always been biased against novel ideas, and AI-generated pitches are fluent by design, so fluency gets mistaken for quality. In a field experiment, evaluators given an LLM recommendation plus a written rationale became more likely to rubber-stamp rejections without getting better at separating correct AI judgments from incorrect ones, and the effect was strongest on borderline cases. A black-box recommendation with no rationale improved decisions.
Ask a model to role-play a thousand consumer segments reacting to a price change and it reads a higher price as premium positioning rather than as the same product costing more, producing a flat or upward-sloping demand curve. Digital twins do better, but make the rational choice more often than people do, understating the sunk costs and switching frictions that decide whether anything new gets adopted.
Post-launch feedback is the one stage where the bottleneck shrinks: AI can cluster and summarize reviews, tickets and social chatter at a volume no analyst team could match. What a frequency-weighted summary buries is the emerging use case or the segment quietly failing to adopt. And a fluent, confident synthesis is easy to cite selectively by a team with a stake in last year’s decision.
A broad business question looms: as AI absorbs ideation, screening and consumer simulation, junior employees stop getting the reps that build judgment. As that spreads across the innovation function, an “automation trap” emerges: every tool tests fine in isolation but the holistic feedback loop drifts from the market. “Everything gets faster. Nothing stays grounded.”
Read more at Harvard Business Review.