Artificial intelligence decision systems are increasingly deployed in safety-, policy-, and human-sensitive settings where actions must satisfy feasibility constraints under incomplete information. Existing post-deployment monitoring approaches can detect observable failures, distribution shifts, or performance degradation, but they cannot, by themselves, determine whether feasibility can be guaranteed when safety-relevant latent states remain indistinguishable at decision time. This paper develops a formal framework for reliable deployment and monitoring of AI decision systems under fixed observation structures. We model deployment through latent states, observations, observation-consistent state sets, state-wise feasibility constraints, and observation-based policies. The framework characterizes when feasibility-guaranteed deployment is structurally possible, when it requires intervention, and when it is impossible without modifying the information structure. We prove that every task falls into one of three regimes: deployable and automatically measurable systems, non-automatically deployable but remediable systems, and hard non-deployable systems. We further introduce an operator-assisted review and rollback mechanism for remediable cases and show that additional data or monitoring alone is insufficient unless such measurements refine feasibility-relevant latent-state ambiguity. Empirical examples and a digital health case study illustrate how the framework supports practical deployment assessment, monitoring design, and human-in-the-loop safeguards for AI systems operating under partial observability.
Zhu et al. (Thu,) studied this question.