When AI Advice Is Not Enough: Institutional Design with Behavioral Frictions and Scarce Verification

Organizations often deploy AI first and redesign workflows later. We model a clinician who records a preliminary assessment, observes a fixed AI risk score, and then treats, withholds treatment, or seeks verification. Verification may generate independent clinical evidence or may explain the AI output and reduce the burden of reversing the clinician's initial judgment. The model shows why AI value depends on governance. Flexible exposure weakly improves the clinician's perceived loss relative to the clinician-only benchmark because the clinician can ignore the score. Divergence frictions, however, shift treatment thresholds toward the reference action and make verification attractive as justification. Poorly chosen selective binding can then make the deployed AI worse than no AI for the organization. We characterize an optimal policy map: bind when enforcement lowers planner loss, assign scarce orthogonal verification by incremental benefit relative to the best non-orthogonal workflow, and choose explainability by comparing reduced operating losses.