To correctly formalise requirements expressed in natural language, ambiguities must first be identified and then fixed. This paper focuses on behavioural requirements (i.e. requirements related to dynamic aspects and phenomena). Its first objective is to show, based on a practical, public case study, that the disambiguation process cannot be fully automated: even though natural language processing (NLP) tools and machine learning might help in the identification of ambiguities, fixing them often requires a deep, application-specific understanding of the reasons of being of the system of interest, of the characteristics of its environment, of which trade-offs between conflicting objectives are acceptable, and of what is achievable and what is not; it may also require arduous negotiations between stakeholders.