Product Service Systems (PSS) refer to integrated systems that combine tangible products and intangible services to jointly deliver specific functionalities and fulfill customer needs. Among these, business models such as rental, leasing, and sharing have been widely adopted across various product categories due to their high economic efficiency. Products managed under these systems typically achieve higher utilization rates, thereby contributing to reductions in environmental impact. Owing to these characteristics, PSS has attracted significant attention as a key enabler of the circular economy. Yet consumer acceptance of PSS remains difficult to predict because the usage mode is relatively novel, making service design a central challenge. In practice, firms increasingly rely on experiments, such as proof-of-concepts and survey-based choice tasks, and then use the results to simulate alternative designs; however, these activities are resource-intensive in time, money, and operations. Hence, there is value in methods that identify the most consequential design factors with minimal pretests. This study examines the extent to which large language models (LLMs) can externally reproduce human choices between purchase and sharing. We benchmark five general-purpose LLMs against a discrete-choice model estimated from real choice-based conjoint data in the refrigerator and laptop markets, using a hierarchical bayesian multinomial logit as the human baseline. Following prior work, we evaluate LLMs along two axes: (i) alignment of choice-probability distributions and (ii) sensitivity to key sharing attributes. Taken together, these evaluations test whether LLMs exhibit human-like responses to design factors and clarify how LLMs can be used upstream of fielding human surveys to inform experimental design for PSS.
Tsurusaki et al. (2026) studied this question.