AI · Full-text empirical research; three scenario-based between-subjects experiments
Is a Chatbot More Effective? Investigating the Effect of Service Recovery Agents and Consumer Loss on Consumer Forgiveness
Journal of Theoretical and Applied Electronic Commerce Research · 2026
Summary
Humans generally outperform chatbots for symbolic loss, where emotional support matters, while the difference diminishes for utilitarian loss. Recovery-agent effects depend on failure severity: for low-severity failures human recovery is more effective, whereas the paper reports different comparative effects under high severity. The results caution against treating chatbot recovery as universally superior. For a founder, use this as a question to test in your own context, not a universal playbook.
For your marketing
Possible founder application
Treat this as evidence for the specific customer behavior and adoption question studied. Compare the setting and sample to your company before changing a growth decision; use a local experiment where the paper does not match your context.
Reported result
What the study found
Humans generally outperform chatbots for symbolic loss, where emotional support matters, while the difference diminishes for utilitarian loss. Recovery-agent effects depend on failure severity: for low-severity failures human recovery is more effective, whereas the paper reports different comparative effects under high severity. The results caution against treating chatbot recovery as universally superior.
Research question
What the authors investigated
How do recovery agent (human versus chatbot), consumer-loss type (utilitarian versus symbolic) and service-failure severity interact to influence consumer forgiveness?
Design
Three scenario experiments with factorial manipulations. Study 1 used a 2×2 design with a hotel check-in failure; Study 2 used a taxi-service failure; Study 3 tested severity moderation in another service scenario. Participants viewed human/chatbot recovery dialogue screenshots and rated responsiveness, emotional support and forgiveness.
Sample and setting
Chinese online participants recruited through Wenjuanxing. Study 1: 270 recruited, 30 excluded, 240 valid (mean age 24.89; 59.7% female). Study 2: 566 recruited, 46 excluded, 520 valid (mean age 23.72; 65% female). The full paper reports Study 3's sample and setting separately.
Scope and transfer
Limitations to keep in view
Scenario experiments limit external validity; low- and high-severity manipulations clustered in the mid-to-high range; samples were Chinese and domains were scenario-specific. The authors call for field tests, more finely separated severity levels, cross-cultural research and tests in high-risk domains such as medical and financial services.
Research sources
MDPI publisher full text supplies exact bibliographic record, DOI, authors, all three designs, participant exclusions/counts, results and a dedicated limitations section. It is a particularly current direct study of chatbot service recovery rather than a review or aggregator summary. Correction notices were not separately verified in this collection; consult the current publisher record before applying the findings.