Current management literature conceptualizes "data collection" — or the organizational process of gathering inputs to train and validate transformational technologies — as a bounded process involving technical work, limited interaction between stakeholders, and finite time. In our 16-month ethnographic study of a healthcare organization developing artificial intelligence (AI) systems, the project team initially approached data collection for AI as literature predicts. We found, however, the project team engaging in a process of data creation involving expansive interactions across different occupations, spanning many organizational practices, and involving diverse stakeholders. Our findings uncovered four consequential, but overlooked, components of the expansive data creation process: what is the phenomenon for which an AI model should be built, what is considered data about the phenomenon, which data can be collected, and which data are ultimately recorded. As a result, this paper's central insight is that rather than conceptualizing data for AI in organizations as a raw, independent, objective resource which is collected through a bounded process, our study highlights how data is contextual, subjective, and dependent, and is actively created through an expansive, iterative approach within organizations. We discuss the theoretical and empirical implications of our results for organizations pursuit of transformational technologies. [ABSTRACT FROM AUTHOR]