When Should Your Team Override AI? A Trust Calibration Routine
针对团队过度依赖或过早拒绝AI建议的问题，提出一个五阶段信任校准流程，帮助管理者通过推翻率和推翻准确率等指标，在具体实践中把握何时该信任AI、何时该推翻它。
Teams that work with artificial intelligence (AI) frequently get the balance wrong: they either accept AI recommendations uncritically (automation bias) or reject them prematurely (algorithm aversion). Both tendencies degrade decision quality, yet managers lack practical guidance on when to rely on AI and when to override it. Drawing on recent empirical evidence on human–AI team performance, we propose a five-stage trust calibration routine—comprising co-sensing, co-framing, co-deciding, action and feedback, and trust reconfiguration—that helps teams maintain appropriate reliance on AI. We identify override rates (the proportion of AI recommendations a team reverses) and override accuracy (how often those reversals are correct) as core indicators of calibrated trust. The routine synthesizes established insights from organizational behavior, information systems, and human factors research into an actionable sequence of concrete review practices and metrics that team leaders can implement. We illustrate its application through two worked examples in Additional Materials, which contrast trust calibration for an agentic AI system (credit scoring) and an anthropomorphic AI system (a customer service assistant), showing how the same five stages apply with different emphases depending on how the AI is presented.