When AI advice backfires: Field experimental evidence on personalization, intrusiveness, and purchase behavior

As conversational artificial intelligence (AI) systems increasingly serve as frontline interfaces in retail, understanding how specific AI recommendation message styles shape consumer behavior has become a central concern. This study examines two alternative AI recommendation response styles - personalized and humorous message framing - and identifies the psychological mechanisms through which these styles influence purchasing decisions. Drawing on persuasion theory and the heuristic-systematic model, we propose a parallel dual-pathway framework in which response style affects purchase behavior through perceived helpfulness attribution and perceived intrusiveness, with shopping goal urgency and customer-retailer relationship strength acting as key boundary conditions. We employ a multi-method research design consisting of an exploratory qualitative study (46 semi-structured interviews) and a randomized field experiment conducted in a U.S. retail context (N = 409), in which purchase behavior is measured using secondary transaction data. The results show that personalized AI communication significantly increases the likelihood of purchase relative to humorous communication. This effect is driven by stronger attributions of helping intent, but is partially offset by heightened perceptions of intrusiveness, revealing personalized message framing as a double-edged strategy. Further analyses indicate that urgency amplifies both pathways, whereas stronger customer-retailer relationships enhance helpfulness attribution and attenuate intrusiveness concerns. Robustness checks confirm the stability of the findings across alternative specifications and outcome operationalizations. Overall, this study advances theory on human-AI interaction by linking AI recommendation message framing to observed consumer behavior and clarifying when personalization enhances or undermines behavioral effectiveness.