How AI assistance shapes human performance, learning, and service delivery across organizational settings.
Working Papers
[1] Generative AI in Action: Field Experimental Evidence from Alibaba’s Customer Service Operations, with Yiwei Wang, Tianjun Feng, Lauren Xiaoyuan Lu, Yitong Wang, and Congyi Zhou [SSRN]
Minor revision at Management Science
First Prize, 2026 POMS College of Service Operations Best Student Paper Competition
Using a field experiment at Alibaba, we examine how generative AI assistance affects customer service agents' productivity and service quality. AI improves service speed and customer-rated quality on average, but its effects vary across workers: lower-performing agents benefit more, while top-performing agents experience declines in service quality, consistent with disruption to their established workflows.
Work in Progress
[2] How AI Agent Design Shapes Student Engagement: Large-Scale Textual Evidence from Higher Education, with Jun Li, Andrew Wu, Robert Bray, Sébastien Martin, Nils Van den Steen, ..., Rebecca Quintana, and Leslie Rupert Herrenkohl
Using large-scale textual data from student–AI interactions, we examine how the design of instructor-created AI agents relates to students' interactive and cognitive engagement. We investigate how different forms of AI assistance shape students' participation in the learning process.
[3] Instructor-designed AI Improves Learning in a Randomized Trial across 19 Universities, with Robert Bray, Jun Li, Sébastien Martin, Nils Van den Steen, Andrew Wu, ..., Wenchang Zhang, and Ruihao Zhu.
In a large, pre-registered randomized trial across 19 universities, we evaluate the effects of instructor-designed AI agents on student learning. Students using these agents perform better on assessments completed without AI assistance, while engagement and satisfaction remain statistically unchanged. The findings demonstrate that instructor-designed AI can support the development of students' knowledge and skills.
[4] AI for the Last Mile: Conversational AI and Public Service Delivery in Rural Communities, with M. S. Krishnan, Anjana Susarla, Andrew Wu, Krishnapriya V, and Pragya Singh
This ongoing project examines whether and how conversational AI can support frontline service providers in rural communities. We investigate the potential of AI assistance to improve public service delivery, extending the study of human–AI collaboration to settings where access to information and service resources may be limited.
How organizational rules shape worker behavior, service performance, and economic outcomes.
Working Papers
[1] Pricing the Gender Gap: Multidimensional Performance Pay in Alibaba’s Gender-Blind Gig Labor Market, with Qiuping Yu and Tianjun Feng
Using operational data from more than 32,000 gig customer-service workers, we examine how standardized compensation converts gender differences in service performance into earnings gaps. Men earn 2.4% more per hour and handle more chats, while women achieve higher customer satisfaction and quality-based compensation. We show how compensation weights determine the earnings implications of these performance differences and how accumulated experience helps explain the observed gap.
[2] The Experience Ratchet: Gender-Blind Work Allocation and the Making of the Gender Gap, with Qiuping Yu and Tianjun Feng
We examine how gender-blind work-allocation rules can generate persistent earnings differences among gig workers. By prioritizing workers with greater accumulated experience, the platform creates a feedback mechanism in which past work increases access to future opportunities, amplifying small differences in labor availability into larger gaps in experience and earnings. Our findings highlight the distinction between static and dynamic neutrality in algorithmic work allocation.
[3] Compensation Plan Design for Heterogeneous Agents in a Large Retail Platform’s Customer Service Center, with Tingliang Huang and Tianjun Feng
Using a principal-agent model, we study how to design compensation plans that balance service quantity and quality for heterogeneous agents. We show that the optimal contract can be reduced to a two-part tariff under certain conditions and propose a simpler contract linear in quantity and quality. This contract outperforms the platform's existing discrete piece-rate scheme, achieves near-optimal profit, and mitigates passive gaming.