Human-machine teaming for intelligent demand planning

Abstract

The second machine age is reshaping the way we work, do business, and collaborate. Today collaboration is switching from just among humans to between humans and machines. Mundane and repetitive tasks will be done by machines automatically, while humans can develop insights and make wise decisions supported by data streaming from intelligent machines. If and how different human-machine teaming decision-making structures would influence the organization's performance is important to understand, so that human-machine teaming capabilities could contribute the most to business outcomes. By using the augmented inverse propensity weight estimator method, this research empirically analyzes the average treatment effects of three different human-machine decision-making structures: Full human to AI delegation, Hybrid AI-Human with adequate human intervention, and Hybrid AI-Human with all steps of demand planning overrides.

These three decision-making structures are defined as treatment groups, and the traditional manual demand-adjustment process is defined as the control group. Effects of switching human-machine teaming decisionmaking structures from one to another are also analyzed. The performance of each treatment and control group is measured by the long-term forecast accuracy, short-term forecast accuracy, and customer inventory level. The project is based on an IT collaboration project between a large fast-moving consumer goods company and one of its largest e-commerce customers. The project implemented an AI-enabled demand-adjustment process to incorporate the external e-commerce customer demand signals into existing demand-planning process. Demand planners engage in the demand-adjustment process via web-based interfaces, to apply human judgment-based decisions. All the stock keeping units are randomly assigned to treatment and control groups.

The results show that after the implementation of human-machine teaming decision-making structures, both demand-forecast accuracy and inventory level are strongly improved by at least 47%. Overall, the Hybrid AI-Human with adequate human intervention model is the optimal decision-making structures between human and machine, which improves the short-term forecast accuracy by 53%, long-term forecast accuracy by 64%, and inventory level by 70%. The Hybrid AI-Human with all steps of demand planning overrides model performed worse than the previous model, because of the heavy human overrides. Additionally, those AI enabled decisionmaking structures works better for low-turnover products than high-turnover ones.