As AI systems proliferate across various institutions and organizations, questions of stakeholder participation and algorithmic transparency grow in importance and urgency. Yet very few in-depth empirical studies of real-world AI design and governance processes exist that we can draw lessons from for informing the future state of practice. In this dissertation, I present an in-depth qualitative analysis of the design processes and transparency practices related to Technology-Assisted Review, or TAR: an AI-driven workflow that has been in use in the U.S. civil justice system for over a decade. Through extensive interviews with computer scientists, lawyers, and judges, as well as archival analysis of government research and U.S. civil court documents, I uncover an unprecedented model for AI participatory design previously unrecognized in the literature. I also uncover a cautionary tale regarding what is likely to occur when we fail to treat AI governance as a design problem on the same order as AI system architecture or workflow. Leveraging these insights, I propose a move beyond current formulations of “human-centered design” focused on individual preferences, beliefs, and values, toward an institution-centered approach. This new approach tackles AI design as an integrative task anchored to a deep analysis of the governing norms, precedents, and structures constituting the institutions into which AI systems are to be embedded.