Human-Centric Perishable Inventory Management with AI-Assistance

Yu Nu, Meng Qi, Karan Girotra, Elena Belavina

Keywords: Perishable inventory management, nonparametric estimation, behavioral operations, human-AI interaction, machine learning

TL;DR: We develop two AI decision-making assistants that support decision-making in human-centric perishable inventory systems.

Abstract: Up to 20\% of food purchased by restaurants is wasted before reaching customers, with commercial leftovers being the major contributor due to poor production planning. This study introduces two AI decision-making assistants that support perishable inventory management differently in human-centric commercial kitchens. We address the periodic review inventory control problem for perishable goods with a fixed shelf life and demand censoring in an offline, data-driven setting. One AI assistant leverages a data-driven prescriptive solution to the multi-period inventory control problem, directly telling how much a human decision maker should replenish for the upcoming season given past sales data. We justify this prescriptive assistant with associated performance guarantees. Building on the data-driven prescriptive solution, the second AI assistant is enhanced in terms of detecting potential human decision-making biases in managing perishable inventory. Using machine learning models, it identifies from past user behavior whether a human decision maker is biased in their inventory decision making, and (if so) what human bias likely accounts for. Through an online experiment with Prolific workers, we further investigate how human users react to the deployment of different forms of AI assistance and uncover factors influencing their effectiveness. Results show that both types of AI assistants, whether providing data-driven prescriptive solutions or bias detection, improve perishable inventory management performance for human decision-makers. Additionally, integrating bias detection with prescriptive solutions could foster greater human adherence to algorithmic recommendations.

Submission Number: 123