Estimating the Economic Impact of ‘Humanizing’ Customer Service Chatbots
by
Tuesday, April 30, 2019
12:30 PM – 2:00 PM
Speakman Hall Suite 200
Abstract
We consider the economic impacts of ‘humanising’ AI-enabled autonomous customer service agents (chat-bots). Implementing a field experiment in collaboration with a dual channel clothing retailer based in the United States, we automate a used clothing buy-back process, such that individuals engage with the retailer’s autonomous chatbot to describe the used clothes they wish to sell, obtain a price offer, and (if they accept the offer) print a shipping label to finalize the transaction. We causally estimate the impact on transaction conversion and price sensitivity from randomly exposing consumers to (1) exogenous variation in price offers, in tandem with (2) exogenously varied levels of chatbot anthropomorphism, operationalized by incorporating a random draw from a set of three anthropomorphic features: humor, communication delays and social presence. We provide evidence of a non-linear relationship, consistent with the ‘Uncanny Valley’ effect documented in the HCI-literature. That is, we show that while introducing either a small (1 treatment) or large (3 treatments) degree of anthropomorphism increases conversion rates substantially (on the order of 10% in the latter case), introducing only a moderate level (2 treatments) is counterproductive. Moreover, we show that a large degree of anthropomorphism (3 treatments) causally increases consumers’ price sensitivity. We argue that this latter effect occurs because, as a chatbot becomes more human-like, consumers shift from a price-taking mindset into a fairness evaluation or negotiating mindset. We discuss the implications for the implementation of AI-enabled autonomous agents in human-facing job roles, and customer service settings in particular.