Conversational AI Agents: The Effect of Process and Outcome Variation on Anthropomorphism and Trust

ABSTRACT Organisations increasingly deploy conversational AI agents (CAs) in agentic roles where behavioural variations are inevitable. Prior work often conflates two distinct forms of variation: outcome variation (where success fluctuates) and process variation (where the path to completion varies). This study investigates how these two types of variation jointly impact user anthropomorphism and trust, integrating the theory of anthropomorphism with the misconception‐of‐chance bias. Using a CA in an agentic role within a car parking simulation experiment, we manipulate process and outcome variation. Results indicate that outcome variation negatively impacts trust, but process variation shows no such direct effect. Interestingly, while neither variation alone increases anthropomorphism, their interaction does. We theorise that the co‐presence of process and outcome variation prevents falsification of anthropomorphic explanations for the CA's behaviour by creating non‐random‐appearing sequences that violate individuals' randomness expectations due to misconceptions of chance. This allows users to apply human‐like mental models, attributing agency and experience to the CA, making failures interpretable as exploratory behaviour. In contrast, when there is outcome variation without process variation (i.e., the CA fails to achieve the desired outcome but exhibits no process variation), users will tend to infer that the CA is mechanically defective. Our results further demonstrate that the joint effect of process and outcome variation indirectly increases trust through anthropomorphism as a mediator. These findings provide insights into the nuanced ways behavioural variation in CAs influence user trust, contributing to the literature on anthropomorphism, trust in AI, and the misconception of chance.