Pathologists have historically evaluated glass slides of excised tissue under a microscope to assess cancer and other disease. Over the last few years, the field has faced an inflection point where glass slides can be digitized, creating images greater than 10 gigapixels in size which are navigated like Google Maps. This enables pathologists to transform their clinical workflow into the digital domain. This has, in turn, begun to produce massive data sets on the order of Petabytes from which to draw biological insights and new clinical applications with the aid of machine learning. This talk begins with a brief perspective on the historical challenges of high-throughput slide digitization and explores the promise of AI in clinical practice. I will then focus on the challenges that we face using image-based AI in high-stakes medical applications, which include generalizability, data drift, and implementation, and suggest engineering and computing approaches to solve them.