When Should a Medical Agent Speak? A Decision-Theoretic Framework for Continuous Ambient Clinical AI

Continuous clinical AI systems excel at detecting abnormalities and predicting deterioration, but lack a principled answer to a more fundamental question: given what the agent currently believes about the patient, is it more valuable to act now or to keep observing? We argue this intervention timing problem is a distinct, underspecified layer in clinical AI—irreducible to anomaly detection or risk prediction. We formalise it as a POMDP in which WAIT is a first-class, cost-bearing action alongside MEASURE and ESCALATE. The resulting framework, CAMA (Continuous Ambient Medical Agency), evaluates actions by expected downstream value, explicitly trading off clinical benefit, information gain, delay risk, and intervention burden. We position CAMA against existing paradigms, state five falsifiable hypotheses, and outline an experimental agenda for testing whether this decision layer yields a more favourable timeliness–burden Pareto frontier than detection- and prediction-based baselines.