Due to the increase in the number of elderly people living alone, the lack of opportunities for daily conversation among the elderly has become a problem. Dialogue robots that provide opportunities for conversation in place of humans are attracting attention. In this study, we improved system efficiency by semi-automating the creation of conversation texts and the recording of conversation content using large language models (LLMs). Through a social experiment targeting elderly people, we demonstrated the effectiveness of the proposed method and evaluated the continuity of conversations. The experiment demonstrated that the quality of the maintained dialogue was equivalent to or better than that of human-generated dialogue, and by leveraging past dialogue, the time required for dialogue generation was reduced to less than 30 s, confirming the effectiveness of LLM-based efficiency improvements. Regarding the continuity of dialogue in the experiment, 25% of the participants consistently responded to the virtual robot’s questions.