While artificial intelligence (AI) is often seen as a panacea for automating various tasks, AI-based software introduces new challenges to organizations. It often necessitates expensive hardware rented from cloud providers and requires extensive data collection from existing infrastructures while ensuring compliance with data protection regulations. Drawing from a two-year-long participant-observer ethnographic study conducted within a federal administration department in Switzerland, this article delves into the challenges experienced throughout the development and deployment of a chatbot designed to address citizens’ queries. I examine the project’s evolution, from its ideation as an intelligent chatbot entity by innovation leaders, its implementation in a vastly reduced scope by an intrapreneurial team, to subsequent efforts to assemble infrastructure for running AI computations. The article argues that the main challenges of applying artificial intelligence within public administration do not lie primarily in the complexities of machine learning algorithms or programming intricacies. Rather, as I show, the operationalisation of A necessitates extensive infrastructural work (Star et Ruhleder, 1996), characterized by the unique hardware and software requirements of AI systems, and by the specific organizational context of public administration. This infrastructural work, often overshadowed by portrayals of AI initiatives as purely technical and radical innovations, is crucial yet frequently overlooked. While existing research in the field of infrastructure studies has focused on the invisible work of collecting, preparing, and interpreting data (Denis, 2018 ; Dagiral et Peerbaye, 2012), this article broadens the analytical lens to include the distinctive infrastructural work required by the deployment of AI projects. These tasks encompass not only data-related activities but also extend to regulatory compliance and computing resource management. We describe three kinds of infrastructural work that emerge in the implementation of AI in public administration : integrating (e.g. operationalizing a prototype that interfaces with legacy systems working across organizational boundaries or setting up manual data collection processes), resourcing (e.g. procuring AI-specific hardware and software), and aligning (e.g. making an AI application compatible with public administration regulations or with organizational structures). We illustrate how these types of infrastructural work correspond to specific kinds of translations (Callon, 1986) that have distinct characteristics. Those characteristics imply that rather than innovating “on” infrastructures (Grisot, Hanseth et Thorseng, 2014), successful implementation of AI in public administration requires thinking relationally by innovating with existing infrastructures, considering hidden work and actors, material computing resources, regulations, and embedded sociotechnical arrangements. We conclude that the successful implementation of AI projects in public administration is hindered by the current framing of AI in public discourse as a disruptive and purely technical innovation and would benefit from a more holistic and critical approach, based on actual practices in the field.