This paper presents a ten-week longitudinal field study of KUDO, an LLM-based documentation support system deployed in a residential nursing home. While prior research on AI-supported documentation has primarily focused on technical feasibility and supervised evaluations, we investigate what happens when such a system is introduced into everyday care work without continuous researcher presence. Although the system generated linguistically adequate reports that were positively evaluated by participating care workers, sustained adoption remained low. Using an ethnographic approach, we analyze non-use and limited integration as sociotechnical phenomena rather than as simple usability failures. We show how misalignments between managerial expectations of regular reporting and care workers' event-driven documentation practices, compounded by infrastructural constraints, privacy concerns, and organizational trust dynamics, undermined meaningful adoption. The system addressed documentation needs of managers more than the care workers' situated priorities, thereby adding work rather than reducing it. We argue that integrating AI into care documentation requires rethinking documentation regimes and accountability infrastructures, rather than layering new technical solutions onto existing systems. By positioning non-use as an analytical lens, this study contributes to CSCW discussions on infrastructure, resistance, and the limits of AI solutionism in complex care environments.
Kletter et al. (Thu,) studied this question.