The railway network managed by SNCF spans 30,000 km of track, with components that are highly diverse in terms of their materials, technology, and age, requiring maintenance that occupies nearly one-third of the infrastructure manager's activities. Since maintenance is inseparable from surveillance, multiple means of monitoring assets have been deployed. The wave of digitization that took place about fifteen years ago has resulted in a data history that today allows us to consider advanced applications based on this feedback. 

Although AI is omnipresent in the company (seat distribution, chatbots, etc.), this presentation will focus on cases of AI application in the fields of predictive maintenance and automatic defect detection. We will endeavor to evaluate the contribution of these methods to maintenance professions, as well as the current limitations encountered for these applications, particularly in an organization managing assets that are sometimes centuries old, where evolutions are heavily constrained by security aspects and where the expertise of agents often still surpasses our best models.