Human perception of intelligent agent errors under different recovery strategies

Intelligent agents are increasingly embedded in everyday services and decision-making, yet their errors can undermine users’ trust, satisfaction, and willingness to continue interacting. Drawing on the Stereotype Content Model, this research conceptualizes post-error recovery as a two-dimensional process involving perceived warmth and competence. Across four scenario-based experiments, we examined how apology, offering alternatives, and self-repair shaped users’ evaluations of intelligent agents after execution and planning errors. The results showed that recovery strategies generally improved perceptions relative to unrecovered errors, but their effects varied across perceptual dimensions and error types. Warmth was more consistently restored than competence, whereas competence recovery was more context-dependent and less robust. Recovery strategies were generally more effective for execution errors than for planning errors. Self-repair showed relatively robust benefits across contexts, particularly for restoring warmth after planning errors, while offering alternatives also produced reliable benefits in several task-oriented conditions. Apology effects varied across scenarios. These findings identify error type as a boundary condition of recovery effectiveness and suggest that intelligent agents should select recovery responses based on the nature of the error and the damaged perceptual dimension.