Truace
Subscribe
68/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Two student experiments find chatbot fixes satisfy more when avatar look matches the reply

Study title: The Effects of Chatbot Recovery Strategies and Anthropomorphic Image Type on Service Recovery Satisfaction: The Mediating Roles of Perceived Warmth and Competence

Evidence type
Controlled or validated comparison
Sample
n = 512
Sources
1 sourcePeer-reviewed
Published
3 min read
Truvace record
TRV-2026-2657cite or check it

Study title: The Effects of Chatbot Recovery Strategies and Anthropomorphic Image Type on Service Recovery Satisfaction: The Mediating Roles of Perceived Warmth and Competence

Two offline experiments with university students tested what a customer service chatbot should say, and look like, after it fails. In Study 1 (162 students), those who saw a chatbot with a competent-looking avatar and an informational recovery scored satisfaction 6.479 on a 7-point scale, against 4.370 when the same avatar gave an emotional reply. With a warm avatar, the emotional reply won, 6.096 to 4.864.

Read the study abstract

The widespread adoption of chatbots enables companies to provide customer service at low cost and with high efficiency. However, the inherent limitations of chatbots often lead to service failures, highlighting the necessity of effective recovery strategies. Nevertheless, the optimal approaches for chatbots to implement self-recovery strategies remain unclear. Drawing on the Computers As Social Actors (CASA) paradigm and the Stereotype Content Model (SCM), this study investigates the interactive effects of recov…

The gain

In two scenario experiments with 162 and 350 Chinese university students, a chatbot whose avatar matched its recovery style (competent look with an informational fix) scored 6.479 on satisfaction versus 4.370 for the mismatched emotional reply.

In brief

Students judged a flight-booking failure in which the chatbot hit an error and then completed the booking. The match between wording and avatar worked partly by changing how warm or capable the chatbot seemed. Study 2 (350 students) added time pressure: under high pressure the competent look plus informational reply did best, and under low pressure the warm look plus emotional reply did best.

The work used imagined scenarios and student participants from two Chinese universities, so it shows how people judged described interactions rather than real customer complaints. The authors present the results as guidance for designing chatbot recovery scripts and avatars.

Main points

  1. Study 1 used 162 valid students in a 2x2 design: emotional versus informational recovery, and warm versus competent avatar.
  2. With a warm avatar, the emotional recovery scored 6.096 on satisfaction against 4.864 for the informational one.
  3. With a competent avatar, the informational recovery scored 6.479 against 4.370 for the emotional one.
  4. Perceived warmth and perceived competence each partly explained the effect of a matching recovery and avatar, with bootstrap tests that excluded zero.
  5. Perceived competence also mediated the warm-avatar plus emotional-reply effect, which the authors had not predicted.
  6. Study 2 had 350 valid students and added low versus high time pressure; the three-way interaction was significant (F(1, 338) = 4.527, p = 0.034).
  7. Under high time pressure the competent avatar plus informational reply gave higher satisfaction; under low pressure the warm avatar plus emotional reply did.

The full story

The scenario had a customer booking a flight from Harbin to Tianjin through a chatbot. After about 20 seconds the chatbot returned an unknown error. In the emotional version it apologized, expressed empathy and urged a retry; in the informational version it listed possible causes and solutions. Both then completed the booking on the second try. Avatars differed by a bright smile and casual dress (warm) or a slight smile and formal dress (competent).

Study 1 ran in March 2026 at a behavioral laboratory at a university in Northeast China. It recruited 169 undergraduates, screened for prior chatbot use, and dropped four who failed an attention check, leaving 162 (63 men, 99 women). A two-way ANCOVA found no main effect of avatar type, a significant main effect of recovery strategy, and a large interaction (F(1, 154) = 124.384, p < 0.001).

Study 2 ran in April and May 2026 at a university in eastern China, with 356 recruited and 350 valid participants (162 men, 188 women). High time pressure was a scenario with six hours before a 14:00 meeting and a flight to book at once. Low pressure was two days of lead time. Manipulation checks for avatar, strategy and time pressure all passed, and participants rated the scenarios as realistic and immersive.

The authors say the findings give firms practical guidance for designing chatbot recovery scripts, such as pairing informational fixes with a professional-looking avatar when customers are in a hurry. Participants were paid 3 RMB each and responded anonymously, and tests for common method bias did not show a serious problem.

Sources

1 on file
  1. Behavioral Sciencesdoi:10.3390/bs16091564
    Peer-reviewed2026-09-03

Talk it over

Four questions to start onWho gets this benefit, and who was left out of the study or story?

Start with a question

Two student experiments find chatbot fixes satisfy more when avatar look matches the replyhttps://truvace.com/item/the-effects-of-chatbot-recovery-strategies-and-anthropomorphic-image-type-on-ser

  1. Who gets this benefit, and who was left out of the study or story?

    The gain In two scenario experiments with 162 and 350 Chinese university students, a chatbot whose avatar matched its recovery style (competent look with an informational fix) scored 6.479 on satisfaction versus 4.370 for the mismatched emotional reply.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a controlled comparison with a sample of 512, from Behavioral Sciences.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Business