The Impact of Input Inaccuracy on Leveraging AI Tools: Evidence from Algorithmic Labor Scheduling

Problem Definition: Considerable academic and practitioner attention is placed on the value of ex-post interactions (i.e., overrides) in the human-AI interface. In contrast, relatively little attention has been paid to ex-ante human-AI interactions (e.g., the selection and maintenance of algorithmic inputs). To what extent do these latter interactions affect the performance of AI tools and the quality of subsequent ex-post interactions? Methodology/Results: We focus on AI-driven labor scheduling where store managers play a critical ex-ante role in maintaining employee availability records that are essential to the performance of the AI tool. These inputs tell the AI tool when employees can work, how long they can work, along with other shift preferences. Our empirical setting comprises five independent retail chains utilizing the same commercial AI scheduling tool, providing visibility into over 99 million shifts across more than 6,200 U.S. retail locations. We identify a major “garbage-in, garbage-out” phenomenon where inaccurate employee availability records lead to: (1) direct effects where managers must manually correct erroneous schedules generated from bad inputs, (2) negative spillover effects on employees not subject to input errors, (3) increased schedule instability as managers attempt to correct error-prone schedules, (4) substantial time burden as managers spend hours interacting with what should be a fully automated system, and (5) greater misalignment between labor supply and forecasted demand. Managerial Implications: While the visible nature of ex-post interactions (i.e., overrides) often draws academic and practitioner attention, our findings highlight the critical importance of managing ex-ante interactions through systematic input verification, clear operational protocols, and early problem detection.