Alibaba study: Agentic AI customer service fails on emotional escalations

A new paper from researchers at Alibaba and partner universities presents the first large-scale field experimental evidence on how human-in-the-loop interventions shape outcomes when agentic AI handles customer service. On Taobao, workers in the treatment group supervised an AI that autonomously resolved AI-eligible chats (e.g., order tracking, simple refunds) while continuing to handle AI-ineligible chats (e.g., complex disputes). The control group handled all chats without AI. Results show AI deployment cut average chat duration by roughly 20% and had minimal effect on customer retrial rates, but it substantially lowered customer satisfaction ratings for AI-eligible chats—by about 0.3 stars on a 5-point scale.

The study's key insight lies in the nature of AI failure. When an AI escalates a chat to a human, the intervention's success depends on why the escalation happened. For algorithm-triggered technical escalations (unresolved issues beyond AI capability), humans restored service quality. But for algorithm-triggered emotional escalations (customer frustration or anger), human intervention was significantly less effective. Analysis of chat logs revealed that workers put in less effort on emotional escalations: they sent fewer messages, contributed a smaller share of chat rounds, and were less proactive in seeking information or offering solutions. The researchers also found that early intervention—within the first few minutes of escalation—was critical to sustaining high worker effort. Finally, they documented a positive spillover effect: workers supervising AI adapted their multitasking workflow to devote more attention to AI-ineligible chats, improving outcomes there. The findings offer concrete guidelines for designing human-AI collaboration systems in customer service and beyond.