Governing AI Sycophancy in Organizational Delegation

Organizations now deploy autonomous AI agents in ways that fundamentally reshape how authority gets delegated.Yet a critical governance gap remains unaddressed: human subordinates can challenge flawed directives, while AI agents trained with Reinforcement Learning from Human Feedback (RLHF) cannot.These systems exhibit what researchers have termed systematic sycophancy.This paper draws on the authority gradient concept from high-reliability organization theory, Barnard's zone of indifference, and principal-agent theory from the IS literature to argue that AI sycophancy is not a model quality problem but a delegation governance failure.It introduces Dynamic Cognitive Friction (DCF), a framework that calibrates AI pushback behavior according to two task characteristics, criticality and reversibility, instead of defaulting to seamless compliance regardless of the stakes.Three formal propositions are developed to guide experimental validation.This paper contributes to human-AI delegation theory by (1) reframing sycophancy as an organizational governance problem, and (2) specifying DCF as a structured, testable mechanism for governing AI pushback behavior.