Human-AI Interaction in Product Recommendation

We study the strategic interaction between a user and an AI agent in product recommendation. The user conveys preferences through a costly, noisy message, incurring cognitive communication cost, while the agent interprets this signal to form a posterior belief and provides a set of recommendations that balances the diversity in the recommendation with the search cost incurred by the user to evaluate the recommendations. The objective is to optimally trade off the utility the user derives from the best recommendation against communication and search costs, modeled through a rational inattention framework. Our main contribution is a formal characterization of the optimal interaction strategy, derived through a high-dimensional approximation that yields the near-optimal trade-off between utility and costs. We show how the optimal strategy depends on feature-space dimension and cost parameters, and we identify distinct regimes where the balance between communication and search shifts sharply. A key insight from our result is that, even in high dimensional setting, the optimal strategy cannot rely only on communication or search in isolation. Instead, both mechanism must be jointly optimized to achieve high utility.