On Worker Behavior in the AI-Enabled Service Triad

Abstract

AI algorithms are playing an increasingly important role in modern service provision. In this dissertation, I develop and explore a concept I term the “AI-enabled service triad,” to describe the interdependent relationships between human workers, human customers, and algorithms inside AI augmented service systems. Although each pairwise interaction in the triad is well-understood, the three-way interaction has remained relatively unstudied. With results from experiments, interviews, observations, and survey questions, I examine the triad across three contexts—personalized recommendation, K12 education, and lending—each characterized by distinct objectives.

In Chapter 2, I explore the triad in the context of a queueing system. Using a novel laboratory experiment, I study how system loads influence workers’ decisions to either default to or deliberate on algorithmic advice. Results from two studies reveal system load and algorithm quality jointly shape workers’ algorithm-use behavior, with downstream effects on service quality and customer throughput times. In Chapter 3, following the evolution of algorithmic decision-support from conventional AI to generative AI, I shift my focus to worker productivity in an entire, discretionary workflow. Through a longitudinal case study of 24 US public school teachers, I identify a new use case for generative AI associated with higher reported productivity—not only outsourcing effort for task outputs (e.g., material creation), but also guiding inputs to workflow planning. In Chapter 4, building on findings from the previous chapters, I consider broader implications of AI decision-support for another operationally relevant service outcome: fairness. Specifically, I use a laboratory experiment to study the effect of algorithm design on fairness and accuracy, as mediated by worker use of (fair) AI recommendations. I show patterns in users’ deviation and reliance result in less fairness than either humans or fair algorithms acting alone.

Together, my dissertation provides insights into the nuanced ways in which human behavior—beyond rate of algorithm use—mediates the intended or assumed benefits from AI—beyond accuracy. While it is motivated by service examples, the implications of the triad framework extend further, setting the stage for future work beyond service about new outcome measures, about dynamic behavior associated with learning over time, and about the decision subjects’ response to algorithm-augmented decision-making.