Abstract: Many domains provide us with (largely untapped) detailed trace data on human decisions in complex, dynamic environments. These decisions are often made by experts with deep experience in the task at hand. For example, nearly every physician action is logged in electronic medical record data; every movement of a driver is recorded on a ride-sharing platform; even store managers’ decisions on daily pricing and inventory management are recorded. How can we leverage these large-scale logs of sequences of decisions and outcomes to improve operational performance? We propose that machine learning algorithms are well suited to identify “best practices” that can help supplement human decision-making.