Human-Computer Interactions in Demand Forecasting and Labor Scheduling Decisions

We investigate whether corporate officers should grant managers discretion to override AI-driven demand forecasts and labor scheduling tools. Analyzing five years of administrative data from a large grocery retailer using such an AI tool, encompassing over 500 stores, 100,000 employees, and 1.5 million store-date observations, we find that managers persistently make overrides, on average 3 weeks in advance of each focal work date, that reallocate labor away from the AI tool's forecasted demand. Fixed effects and instrumental variables regressions reveal that these overrides increase store labor productivity. Supporting the hypothesis that managers' private information about store demand drives these productivity gains, we show that overrides made by managers: (1) positively correlate with foot traffic, with increases (decreases) in labor on a given day following increases (decreases) in foot traffic, (2) have larger effects for customer-interaction-intensive categories like made-to-order foods and tobacco products, (3) increase basket sizes, (4) reduce self-checkout usage, and (5) become more effective with increasing managerial tenure, suggesting accrual of domain knowledge over time. However, overrides consume significant managerial time and lead to less consistent employee work schedules, potentially due to managers aggressively aligning labor with their private beliefs about demand, which may not fully account for its impact on employee welfare.