In conventional software development, the boundary between user experience (UX) design and engineering is well defined: designers create specifications based on end-user needs, then engineers build to those specifications. However, AI application design poses a challenge to this "separation of concerns." Emerging Human-AI guidelines show that human-centered design extends beyond the user interface and into the design of AI sub-components and training data, thus 'puncturing' this separation. In this talk, I will share insights about collaboration challenges at the AI-UX boundary and discuss approaches to operationalize the vision for human-centered AI. Based on studies with industry practitioners, I will describe how "leaky" abstractions afford collaboration across expertise boundaries and discuss the critical role of end-user data in generating both AI and UX design specifications. Finally, I will present an approach for prototyping AI-powered interfaces for diverse users and use contexts by directly incorporating end-user data and machine learning models within UX workflows.