Learning to Advise Humans in High-Stakes Settings