Deployment of AI-Assisted Interventions: Capacity Constraints and Noisy Compliance
This work addresses how AI tools used for healthcare or education interventions fail when service capacity is limited and compliance with algorithmic suggestions is inconsistent. It shows that setting intervention thresholds based only on prediction accuracy leads to poor performance because it ignores the trade-off between underutilization (wasting capacity) and cannibalization (crowding out high-value requests).