Agentic AI systems are increasingly positioned as productivity infrastructure for cognitive work, yet they conflate scalable generation with productive completion, systematically transferring the burden of verification and operational reliability onto the human operator. This paper introduces the concept of Supervision Debt: the cumulative supervisory effort imposed on human operators by the asymmetric relationship between automated output generation and human-bounded validation.