A New Behavioral Experimental Paradigm: AI Agents, Risk Preferences, and the Bullwhip Effect in Inventory Management

As large language models (LLMs) increasingly participate in decision-making tasks, a pressing question emerges: do these systems exhibit consistent strategic behavior, and can they be guided to reflect specific behavioral traits such as risk sensitivity? In this paper, we address this question within the complex operational setting of inventory management. We conceptualize risk preferences not merely as labels, but as strategy-modulating cognitive lenses that fundamentally alter how LLMs perceive and react to environmental uncertainty. We conduct extensive multi-round experiments using LLMs with varied risk profiles: original (default), risk-averse (low), risk-neutral (balanced), and risk-seeking (high). Through rigorous statistical analyses of their behavior, our findings demonstrate that LLMs can rapidly achieve rational, stable behavior, effectively replicating and extending results from human experiments. Notably, we reveal that risk preferences significantly shape inventory strategies: risk-averse agents tend to maximize inventory, while original, risk-neutral, and risk-seeking types show no significant difference in their focus on maximizing individual profit. Furthermore, variance amplification for risk-averse agents is substantially reduced under information sharing, underscoring the potential to steer LLM decision-making. Our study advances behavioral operations research by providing robust empirical evidence on the controllability and underlying mechanisms of risk-driven AI behavior in complex tasks.