I helped build something kind of like this in college for a research team that was aimed at facilitating in-person arguments between 2+ people. It took speaking

I helped build something kind of like this in college for a research team that was aimed at facilitating in-person arguments between 2+ people. It took speaking tone, body language, and facial expressions as input to judge how productive the conversation was in terms of resolving an argument, and would offer spoken interjections to steer the conversation toward more productive / less aggressive paths. It was significantly more effective when participants went in blind and didn't know how it "knew" how they were feeling, but usually after 30-45 minutes of observation people would start to pick up on how it worked and often tried to manipulate it instead of actually trying to resolve the conflict. Our application seemed good in theory, but worked horribly in practice. Hopefully Microsoft can figure something out that is net-beneficial in some scenarios without going full-dystopian.