Forty years ago Benjamin Bloom showed that one-on-one tutoring moves a learner two standard deviations, but the resources required to scale one-on-one tutoring have historically been prohibitive. Generative AI offers a potential solution to Bloom's 2-sigma problem, but not without risks. Advanced AI technologies can provide preclinical students with on-demand practice and immediate feedback; however, the same tools can eliminate the productive failure that practice is for if they are not thoughtfully integrated into learning workflows. This Ignite talk draws on four tools built and studied with first- and second-year students at the University of Cincinnati, and on an international study of how medical education leaders decide which AI tools reach learners. Tangible examples will illustrate best practices for AI tool development, deployment, and continuous monitoring and evaluation, drawn from the medical education literature and from 2-Sigma Labs. By the end of this session, participants will be able to describe the design conditions under which AI feedback tools support learning, identify the faculty judgment each tool depends on, and name one change they can make in a course they teach.