Recent achievements of artificial intelligence (AI) have caused organizations to increasingly bring AI capabilities into their core business processes. Such AI-supported business processes often result in human+AI centaurs, which consist of an AI system, which performs most of the execution, and humans, who monitor this execution and occasionally provide additional inputs and overrides. Using sales data from Walmart, we conduct an online study to investigate if human supervision can improve upon state-of-the-art AI forecasts. Furthermore, we analyze the perceptions and behavioral intentions of the human participants over time. We find that human interventions consistently lead to less accurate forecasts and that participants initially underestimate the AI system’s accuracy and overestimate their own potential to improve upon the AI forecasts. However, perceptions quickly shift over the course of the study, causing the participants to perceive the AI system increasingly favorably, which also leads to behavioral changes and better overall system performance.