DIPLOMAT is an AI system built to handle one of the trickiest human skills: negotiating at work without torching the relationship.
Researchers trained DIPLOMAT using a new preference-learning method called Dialogue-Span-Aware Direct Preference Optimization, which flags the specific stretches of a conversation where tone and strategy matter most. The model learned from PROWESS, a dataset of multi-turn workplace negotiation dialogues that the researchers generated with a multi-agent framework rather than pulling from real transcripts. Each conversation was labeled for negotiation strategy, politeness level, and persuasive technique. In testing, DIPLOMAT beat baseline models on automatic and human evaluations for coherence, politeness, and persuasiveness.
Here is the catch: every dialogue DIPLOMAT trained and tested on came from AI agents role-playing a negotiation, not from actual coworkers hashing out a raise or a deadline. Outperforming other bots on a benchmark the researchers built is a narrower claim than working at your next performance review, and the paper does not evaluate DIPLOMAT against real workplace conversations.
It is a useful research step, but until someone tests this on an actual annoyed manager instead of another chatbot, treat polite and persuasive as a lab result, not a job reference.