Splitting warmth and efficiency between chatbot and human agent lifted satisfaction in 800-person experiment

Mean satisfaction on a 7-point scale was 4.80 for social chatbot with task-focused human, 4.52 for task chatbot with social human, 3.64 for social-social and 3.69 for task-task. Participants in mixed-style conditions also rated the agents' collaboration and empathy higher, and these ratings partly explained the satisfaction gain.

Study 2 (1,200 participants) looked only at mixed-style teams. A warm-first sequence raised perceived collaboration more in post-purchase scenarios, and empathy mattered more to satisfaction when customers themselves asked for a human. All participants responded to scripted dialogues rather than real service encounters.

Study 1 used a 2 × 2 between-subjects design in which participants imagined themselves as customers contacting an e-commerce platform. The chatbot replied first and a human agent then took over. Social styles used greetings like "Hello there!" and empathic statements, while task styles were strictly transactional. Participants came from the Credamo panel, were paid RMB 2, and answered 7-point scales for collaboration, empathy and satisfaction.

Bootstrapped mediation analysis showed both perceived collaboration (indirect effect 0.246) and perceived empathy (0.190) carried part of the style-mismatch effect on satisfaction, with a direct effect remaining. Mean collaboration scores were 5.57 and 5.02 in the mixed conditions versus 4.30 and 4.45 in the same-style ones.

Study 2 manipulated communication order and service stage; transfer initiation was decided by participants who chose yes or no to a transfer prompt. Quotas of 150 per cell were applied, and respondents were paid RMB 3. The warm-first order outperformed task-first, with the gap largest after purchase and when customers requested the handoff.

The authors suggest service teams could structure chats so warm engagement comes before task resolution. They note that the sequencing effect reflects how warmth and competence cues are configured across agents, not sequencing acting as a separate mechanism.