Consensus-building is a key issue in democratic processes. Discussion and deliberation can be expected to make a fair consensus, but there are concerns that everyone’s agreement takes time (efficiency) and is easily influenced by human relationships (rationality). Many technologies, such as visualizing discussion structure and providing a consensusable choice, have been proposed to make discussions efficient and rational by providing logically appropriate information to participants. These are a type of discussion facilitation technologies. However, discussion management is also an important function of facilitation, such as encouraging the proper people to talk at the proper timing. We propose an AI-facilitation method for managing the progress of a discussion, which encourages an appropriate participant to talk by virtually evaluating future discussion states. The future discussion states are estimated by simulating the discussion among participants using Large Language Models (LLMs) with the participants’ personas and talk histories. The effectiveness of the proposed method was experimentally evaluated using the facilitation scene for a discussion on a regional depopulation issue where there are conflicting opinions among the participants. Two types of participant state representation, preference-based and embedding-based, were evaluated. From the results, we show that the proposed AI is efficient to make a consensus on a given fair compromise choice in 11% through discussions. We also discuss the difficulties in the use of current LLMs as human subjects.