The analysis of bone breakage has always been underrepresented in taphonomic studies. Analysts, thus, lose the opportunity to resolve an important part of the equifinality related to activities that hominins and different types of carnivores may produce. Recent studies have shown that the use of powerful machine learning (ML) algorithms allow the accurate classification of bone surface modifications (BSM). Here, we present an experimental study, applying these algorithms to the analysis of bone breakage patterns. This statistical methodology allows the correct classification of three different assemblages which have been generated anthropogenically and by the activity of carnivores (i.e., hyenas and wolves). ML algorithms applied to a multivariate set of properties of broken bone specimens yielded an accuracy of 95% and were higher in classifying agency without the need to include information from BSM. This paper proposes a methodological approach that opens the door to improve our understanding of referential frameworks regarding bone breakage and to determine agency in prehistoric bone breakage processes.