Scholars and practitioners have called for personalized and widely accessible professional development (PD) for teachers. Yet, a long-standing tension between customizing support and increasing access to such support has hindered the scale-up of high-quality PD for individual teachers. This study addresses this challenge by developing a computerized program for middle school mathematics teachers that provides frequent opportunities for teachers to interact with and obtain personalized and real-time feedback from a virtual facilitator based on natural language processing. Based on the data collected from 1727 middle school students in an experiment in which the teachers of these students were randomly assigned to the program or the business-as-usual condition (i.e., the control group), we found that the program had a statistically significant impact on students' mathematics performance. These results demonstrate the potential of incorporating an automated, interactive feedback tool supported by artificial intelligence to create effective, scalable teacher PD.