Floods are the natural hazards that are causing the most deaths worldwide. Flood early warning systems are one of the most cost-efficient methods to reduce death rates, triggering decisions about the evacuation of exposed population. Although previous studies have investigated the effect of human behaviours on evacuation processes, studies analysing a combination of behaviours, flood onset and warning timing are limited. Our objective is to explore how changes on the aforementioned factors can affect casualties. This is done within a modelling framework that includes an agent-based model, a hydraulic model, and a traffic model, which is implemented for the case study of Orvieto (Italy). The results show that the number of casualties is most impacted by people’s behaviour. Besides, we found that a delay of 30 min in releasing the warning can boost the number of casualties up to six times. These results may help managers to propose effective emergency plans.