Opening the black box: How reasoning-enabled AI agents influence user perceptions and behavior in sustainable consumption

Recent advances in chain-of-thought (CoT) reasoning in large language models are expected to improve the transparency and coherence of AI-generated recommendations, yet evidence on how CoT affects users’ perceptions and downstream behavior in real decision contexts remains limited. To address this gap, we develop and evaluate a reasoning-enabled conversational agent in the domain of sustainable consumption, where uncertainty, value-laden trade-offs, and multi-criteria decisions make reasoning transparency especially salient. We employ a three-phase research design that combines model development with empirical evaluation. In Phase 1, we fine-tune a conversational agent using a multi-agent framework to produce high-quality CoT data, leveraging Group Relative Policy Optimization (GRPO) and Low-Rank Adaptation (LoRA). Phase 2 tests the agent in a between-subjects experiment (N = 417) comparing a CoT-enabled chatbot with a standard chatbot. Grounded in the Technology Acceptance Model, IS Success Model, Cognitive Load Theory, and Privacy-Calculus Theory, we examine how user-friendliness, usefulness, personalization, trust, transparency, cognitive load, inefficiency, and privacy concerns mediate CoT effects on perceived knowledge and behavioral intentions. Results show that CoT significantly increases perceived knowledge and behavioral intentions. Specifically, PLS-SEM supports our mediation model, demonstrating that CoT acts through promoters such as enhanced trust, perceived usefulness, and personalization, which outweigh inhibitors like cognitive load and privacy concerns. Phase 3 complements these findings through qualitative content analysis, indicating that CoT improves user experience by providing clearer, more helpful reasoning, whereas standard chatbots are often perceived as verbose, vague, or technically unreliable. Overall, the study provides empirical evidence on how reasoning-enabled conversational systems shape user perceptions and behavioral outcomes, offering actionable guidance for designing decision-support transparent AI in complex domains.