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

Abstract

Are the inputs used by your AI tool correct and up to date? In this paper, we show that the answer to this question: (i) is frequently a “no” in real business contexts, and (ii) has significant implications on the performance of AI tools. In the context of algorithmic labor scheduling, we propose, identify, and study a problem relating to inaccurate employee availability records, which are used by an AI tool to assign employees to shifts that are necessary to meet required service levels. We study this problem using granular data covering multiple retail chains, which contain more than 74 million shifts that are scheduled for more than 290,000 employees in more than 5,900 brick-and-mortar store locations. In our data, we find that employee availability records are often set incorrectly. Specifically, we find that employees who are no longer available to work are scheduled to work by the AI tool, and employees who are available to work have no existing availabilities. We find evidence that such input inaccuracies directly affect the number of overrides as managers rectify these errors, but also have a spillover effect on shifts that are not subject to input inaccuracies. Ultimately, we find that input inaccuracies take up significant managerial time and have a negative effect on the quality of work schedules, which may lead to a decrease in store performance. Overall, our findings suggest that poor AI input quality management could be one explanation behind the well-documented human distrust of algorithms and the lack of observed business gains in their use.