Using Artificial Intelligence To Reduce Food Waste

In this study, we estimate the reduction in food waste that arises from the deployment of a system that digitally records instances of food items discarded in a commercial kitchen. We also shed light on the mechanisms that drive this impact. In a quasi-experimental setting, where the system was deployed in approximately 900 kitchens in a staggered manner, we estimate the impact using the synthetic difference-in-differences method. We find that three months after adoption, kitchens generate 29% lower food waste, on average, than they would have in the absence of the system— without any corresponding reductions in sales. Utilizing a long-short-term-memory fully- convolutional-network classifier, we document that these reductions are accompanied by a 23% decrease in demand chasing, a known bias in human inventory management. Upgrading to a system that uses computer vision to automate waste classification leads to a further 30% reduction in food waste generated by the kitchen a year after the upgrade. This further reduction is due to the accurate recording of infrequent but very high-impact instances of food wasted that employees avoid entering manually. We also observe substantial effect heterogeneity. Smaller kitchens and those with buffet service (vs. table service) experience almost double the reduction in food waste from the adoption of the system and also from the computer vision upgrade. Low and high-demand- variability sites have higher reductions from adoption than those with medium-demand-variability (42% vs 25%). The impacts of the upgrade are not detectably different with different demand variability.