Observing and recognizing precursors and near misses are essential parts of risk management. Many industries have implemented monitoring programs to that effect. The problem is to decide what to monitor, how to interpret the signal(s), and how to respond given the lead time. This article focuses on a Bayesian (probabilistic) approach to the design of warning systems, which permits addressing the trade‐off between the probabilities of false positives and false negatives. The results can then be used as input in a decision analysis to support system‐wide risk management. Probabilistic risk assessment (PRA) is used here as a fundamental tool that allows identifying and fixing the weakest parts of a system. It is shown how one can update the probability of failure of a system given signals that one of the components or subsystems has deteriorated, and how to use PRA to identify potential precursors. The example of a risk analysis for the tiles of the space shuttle is used to illustrate the identification of the failure modes and the weakest parts of a system, a warning that was ignored at the time of that study, years before the Columbia accident. In the dynamic case, it is shown how to optimize a warning threshold to minimize risks and losses. Finally, it is shown how human and organizational factors can be linked to the risk of system failure and the effectiveness of an organizational warning system.