AI on My Shoulder: Supporting Emotional Labor in Front-Office Roles with an LLM-based Empathetic Coworker

Research Background and Issues

Issues and Challenges

• Customer Service Representatives (CSRs) frequently interact with customers who use inappropriate language, posing a threat to their mental health.
• Uncivil customer behavior can lead to emotional exhaustion, detachment, and low sense of achievement for CSRs, while at the organizational level, it may result in high turnover rates and negative work attitudes.
• The customer service industry has long lacked innovative approaches to alleviate the emotional burden on CSRs.

Importance of the Issue

• CSRs are the frontline of interaction between organizations and customers, and their mental health and job satisfaction directly impact service quality and customer experience.
• The negative effects of uncivil customer behavior and emotional labor on CSRs have become widespread, necessitating urgent intervention measures to alleviate this stress.

Research Motivation and Related Work

• Existing social support mechanisms (e.g., colleague support) have proven helpful in reducing stress but are less effective in remote work environments.
• The rise of artificial intelligence, particularly generative language models (LLMs), offers new opportunities for scalable emotional support. However, the HCI field lacks research on emotional labor scenarios.
• This study aims to explore the applicability of AI assistants in emotional labor by integrating LLMs as tools to support CSRs in emotional regulation during tasks.

Solution

Method Overview

• This study developed an LLM-based AI assistant called Care-Pilot, which helps CSRs regulate emotions through emotional reframing and cognitive adjustment.
• The primary function of Care-Pilot is to generate empathetic and supportive messages to help CSRs cope with negative emotions caused by uncivil customer behavior.

Innovations

• Utilizing LLMs for situational regulation represents a novel intervention approach for emotional labor. Care-Pilot embeds emotional support into real-time task workflows rather than relying on traditional post-training or support methods.
• A systematic strategy was proposed to guide LLMs in generating emotionally supportive messages through Chain-of-Thought reasoning, demonstrating overall higher sincerity, actionability, and adaptability compared to human colleague support.

Implementation Steps and Techniques

1. Data Preparation
   • Real customer complaint data was used to generate realistic multi-turn dialogues simulating uncivil customer scenarios.
   • Structured classification was employed to categorize customer complaints into five types: service quality, product issues, pricing, policies, and solutions.
2. Empo-Reframe Module
   • LLMs were used to extract situational information from dialogues, infer CSRs' potential negative thoughts, and reframe them into supportive, new cognitive frameworks.
   • Chain-of-Thought prompting was applied to sequentially generate problem contexts, negative thoughts, and reframed coping strategies.
3. System Deployment
   • Care-Pilot was integrated into CSRs' daily workflows, providing emotional tags for chat histories, specific operational guidance (Info-Guide), and emotional reframing support.

Research Outcomes

Specific Results

• Supportive messages generated by Care-Pilot outperformed human-generated messages across multiple dimensions, including authenticity, empathy, warmth, actionability, and relevance.
• CSRs reported that Care-Pilot helped them avoid negative thinking, refocus attention, and assist in humanizing customer interactions. Additionally, Care-Pilot indirectly reduced the burden of seeking colleague support.

Experimental or Evaluation Results

• Technical Evaluation
  • Care-Pilot's messages were more adaptable to specific scenarios, exhibiting higher emotional expressiveness and adaptability compared to human-generated messages.
  • Care-Pilot's messages were more detailed (length increased by 67%) but required higher reading comprehension skills (readability index increased by 36%).
  • In empathy detection tasks, Care-Pilot's messages demonstrated superior empathy expression compared to other LLM benchmark models.
• User Evaluation
  • CSRs noted that Care-Pilot's emotional reframing function provided "angel-on-the-shoulder" support, helping them handle uncivil customer behavior.
  • Care-Pilot excelled in assisting CSRs in emotional adjustment, avoiding emotional traps, and refocusing on problem-solving.

Comparative Advantages Over Existing Solutions

• Care-Pilot significantly outperformed zero-shot prompting and other LLMs across multiple dimensions of emotional support, such as empathy and actionability.
• In practical CSR operations, Care-Pilot provided immediate emotional intervention support without interrupting tasks to seek colleague assistance.

Limitations and Future Research Directions

1. Limitations
   • Current evaluations rely on synthetic data and simulated scenarios, requiring further validation in real-world contexts and long-term deployments.
   • Care-Pilot's language generation may lead to information fatigue after repeated use, potentially causing diminishing marginal utility in high-efficiency task environments.
   • Lacks the social connectivity built on shared experiences that human colleagues provide.
2. Future Directions
   • Introduce long-term memory (LTM) capabilities to help Care-Pilot learn CSRs' historical interaction contexts for more personalized message outputs.
   • Utilize real-time emotion and stress sensing technologies to provide targeted emotional support at appropriate times.
   • Explore integrated testing based on real CSR logs to enhance scenario complexity and usability.

Overall, this study demonstrates the potential of LLMs in supporting employee mental health through innovative AI-mediated emotional labor design. It also provides practical and theoretical insights for the development of emotional AI and the redesign of social norms in the future.