With the development of the web, a huge number of services available on the web have been published. These web services operate in several application domains like concurrent engineering, semantic web, system engineering or electronic commerce. Moreover, due to the ease of use of the web, the idea of composing these web services to build composite ones defining complex workflows arose. Even if several industrial standards providing specification and/or design XML-oriented languages for web services compositions description, like BPEL, CDL, OWL-S, BPMN, the activity of composing web services remains a syntactically based approach. Due to the lack of formal semantics of these languages, ambiguous interpretations remain possible and the validation of the compositions is left to the testing and deployment phases. From the business point of view, customers do not trust these services nor rely on them. As a consequence, building correct, safe and trustable web services compositions becomes a major challenge. It is well accepted that the use of formal methods for the development of infor-mation systems has increased the quality of such systems. Nowadays, such methods are set up not only for critical systems, but also for the development of various infor-mation systems. Their formal semantics and their associated proof system allow the system developer to establish relevant properties of the described information sys-tems. This talk addresses the formal development of models for services and their com-position using a refinement and proof based method, namely the Event B method. The particular case of web services and their composition is illustrated. We will focus on the benefits of the refinement operation and show how such formalization makes it possible to formalise and prove relevant properties related to composition and adaptation. Moreover, we will also show how implicit semantics carried out by the services can be handled by ontologies and their formalisation in such formal develop-ments. Indeed, once ontologies are formalised as additional domain theories beside the developed formal models, it becomes possible to formalise and prove other prop-erties related to semantic domain heterogeneity.