<p>Sepsis is a life-threatening disease requiring immediate medical treatment in an intensive care unit. Sepsis patients require close monitoring, and the rapid and aggressive treatment increases the chances of surviving. An existing study introduces a Markov decision process model that determines the optimal amounts of intravenous fluids and vasopressors to reduce the mortality of sepsis patients. The treatment policy suggested by this model, namely the artificial intelligence (AI) policy, could be effective in the sense of the mortality. However, the AI policy does not try to manage the critical health states of the patients, such as septic shock directly causing organ dysfunction and physical and neurocognitive decline, while clinicians try to avoid such states during the treatment process. Therefore, we introduce a new measure, called the <i>critical state probability</i> (i.e., the probability of being in a critical health state), and propose a framework to manage both the mortality and the critical state probability. Our proposed framework successfully derived effective treatment policies, namely the risk-adjusted AI policies that reduce the mortality while keeping customized upper bounds of the critical state probability. In a case study, experimental results demonstrated the advantages of our framework compared to both clinicians’ and AI policies. Notably, when the tightest upper bound of the critical state probability was applied, our framework effectively reduced both the mortality and the critical state probability compared to the clinicians’ policy. Additionally, it significantly lowered the critical state probability while maintaining a similar mortality level when compared to the AI policy.</p>