Working Papers (Under Review or Preparing for Submission)
Informal Ties and Sales Agent Turnover: Evidence from Randomized Dormitory Assignment
Salesperson turnover poses significant challenges for firms that rely on sales agents to maintain customer relationships and generate revenue. This paper offers the first causal evidence on how informal ties (e.g., shared hometown ties) among agents affect turnover. Partnering with a leading Chinese medical equipment company, we construct a unique data that tracks roughly 2,000 agents from hire to exit. New hires are randomly assigned to company dormitories, giving us exogenous variation in exposure to peers who share background traits such as hometown, dialect, or alma mater. Overall, we find that stronger informal ties reduce turnover. Agents initially placed to dorms with more same-background roommates remain with the company significantly longer. Yet these ties also carry a cost: when a strong-tie roommate leaves, the likelihood that others quit rises, creating a contagion effect. We show that the retention benefit stems from psychological support, not productivity gains due to peer learning. A back-of-the-envelope calculation suggests that giving each newcomer one additional same-background roommate could save recruitment and training costs by roughly 8.2% of annual profit. Our findings suggest that cultivating friendships among new hires help reduce turnover at little cost.
LLM Simulation Training: Field Evidence on Turnover and Performance in Healthcare B2B Sales
Sales agent training costs billions annually, yet the results are often disappointing. As a result, managers are continually seeking more effective sales training methods. Partnering with a leading medical device company, we conduct a five-month randomized field experiment with over 200 newly recruited sales agents. We randomly assign agents to LLM simulation training or traditional role-play training and find that LLM training reduces early-career turnover by 21% and enhances customer engagement. We then analyze millions of training interactions through text analysis and examine 160,000 following early-career real customer conversations through voice analysis. Combined with interviews, these analyses reveal when, why, and how LLM simulation proves effective in sales training. Our rough estimates indicate a return on investment of 900%, primarily driven by reduced turnover costs and improved performance metrics.
Optimizing the Use of LLM as a Co-Pilot in Sales Interactions
LLM has shown huge potential in sales service. From a workflow view, there are three questions: Should experts pre-train LLM? When to refine customer questions for LLM? How to modify LLM responses to customers? We answer them through field studies with a Fortune 500 EV company. Working with new energy vehicle sales agents over three months, we experimentally manipulate each workflow phase. We find that expert knowledge integration increases customer satisfaction by 18%, query enhancement shows diminishing returns beyond basic clarification, and selective human editing outperforms both full automation and complete manual revision. Our analysis of large-scale sales interactions reveals that strategic human-AI collaboration improves conversion rates by 23% while reducing response time by 40%, providing empirical guidance for LLM deployment in competitive automotive retail environments.
Work-in-Progress
The Economics of Sales Visits: Evidence from Healthcare B2B Price Negotiations
Face-to-face sales visits remain costly yet critical for B2B relationship building and price negotiations. Optimal visit strategy (considering distance, duration, timing, and agent characteristics) remains poorly understood. Using 20,000+ healthcare sales visit records in four years, we model the sales visit and its impacts.
Volume to Content: Multi-Stage Field Evidence on LLM Appraisal's Impact on Sales Agents
Sales agent effort is always difficult to measure even with huge investment. LLM now can auto assess sales agent interactions, yet its impacts remain unknown. To answer this, we conduct a three-stage field experiment with a leading sales center to separate new appraisal's awareness, learning, and adoption effects.
Inclusive Visual Marketing: Overcome Cognitive Biases for Vision-Impaired Consumers
Visual marketing often excludes consumers with visual impairments through biases like poor grouping, framing, and contrast. Working with a major eye hospital, we study how these biases affect patients with cataracts, glaucoma, macular degeneration, and retinopathy, and how to correct biases with deep learning methods.
Video Listening In: Extracting Purchase Intent in Physical Retail
Cameras are more common in retail and can record unprecedented consumer behaviors. However, most retailers waste this data without benefiting customers or themselves. Partnering with a shopping mall chain (60000 daily), we develop a video listening approach to extract purchase intent from trajectories and motions.
Beyond Silicon Participants: Probing True Consciousness in LLM Marketing Applications
We do not know if LLMs show true consciousness or just predict the next word. More researchers now use LLMs as silicon participants or silicon raters, raising concerns about this key question. We use computational probing and two-stage controls to separate consciousness from statistical prediction.