The AI Voice Assortment Problem: Field Experimental Evidence from Voice-Based Chatbot Persuasion

Abstract

Voice-based AI systems increasingly allow firms to choose among many possible voices, yet research and practice often treat voice design as the selection of a generally preferred voice. This view overlooks a distinct design challenge: firms must match voice features with the social contexts in which AI agents interact with users. We conceptualize this challenge as the AI voice assortment problem. We examine it through five large-scale randomized field experiments with 123,946 outbound chatbot calls. Drawing on adaptive structuration theory, we theorize pragmatic voice features, including illocutionary acts and emotional tone, and sociophonetic features, including perceived age and gender, as technical structures whose effects depend on alignment with user, task, and environmental conditions. The results show no evidence of a universally optimal AI voice. Affirmative illocutionary acts outperform question and rhetorical formulations. Gentle and stern tones outperform coquetry tones. Perceived voice age has no overall effect, and voice gender matters only in conjunction with emotional tone. These effects further vary across customer gender, product price tier, city level, and weekday versus weekend contexts. This work shifts research attention from static voice optimization to contextual voice assortment and shows how users’ behavioral appropriation translates AI voice structures into persuasion.

Keywords: AI Chatbot, voice features, adaptive structuration theory, persuasion, pragmatics, sociophonetics