How the communication style of chatbots influences consumers’ satisfaction, trust, and engagement in the context of service failure

The study addresses how chatbot communication style (social-oriented vs. task-oriented) influences consumer outcomes following service failures. In service contexts, AI chatbots often face consumer aversion when failures occur, producing negative emotions and harsher evaluations compared to human agents. Existing research emphasizes positive effects of anthropomorphic cues in successful service encounters, but less is known about failure contexts. Drawing on mind perception theory (warmth and competence) and expectancy violations, the authors investigate whether social-oriented communication can mitigate negative reactions by enhancing perceived warmth, and whether expectancy violations condition these effects. The study proposes hypotheses: H1, social-oriented style increases interaction satisfaction, trust, and patronage intention; H2, perceived warmth (for social style) and perceived competence (for task style) mediate effects on outcomes; H3, expectancy violations moderate these indirect effects, strengthening warmth-based mediation for social style at high violation levels and competence-based mediation for task style at low violation levels.

The paper situates the work within human–computer interaction (HCI) and service research, noting that chatbots can evoke social responses akin to human interactions (CASA paradigm), and that automated social presence influences user attitudes and intentions. Prior studies show mixed outcomes: anthropomorphic cues can enhance user experience in positive contexts, but in failures, users often evaluate services and firms more negatively when chatbots are involved. Research has differentiated chatbot communication styles into task-oriented (formal, goal-focused) and social-oriented (relational, empathetic). Warmth and competence are highlighted as key mind-perception dimensions shaping consumer judgments. Expectancy violations literature indicates that unmet expectations (often inflated for AI) exacerbate negative reactions; anthropomorphism and language use can shape these expectations. The review consolidates evidence that social dialog can increase trust and engagement, while failures trigger algorithmic aversion, motivating the current focus on communication style as a means to mitigate negative evaluations during service failures. A table of prior studies summarizes contexts, methods, and findings across tourism, hospitality, health, finance, and online retail settings.

Design and participants: A between-subjects experiment manipulated chatbot communication style (task-oriented vs. social-oriented) in a service failure scenario. N = 141 participants (59.6% female; mean age = 31.28 years) from Sojump (China) were randomly assigned to one of the two conditions. A pre-test (N = 70) validated the communication style manipulations. Stimuli: Screenshots of conversations with a virtual sports brand’s customer service chatbot depicted either task-oriented (formal, goal-focused dialog limited to guidance and information) or social-oriented (informal, relational dialog including small talk, empathy, and positive sentiments) styles. A robot avatar was used as a visual cue. Service failure was operationalized via an online shopping scenario involving delivery delays, product quality issues (broken sole), and complications with returns/exchanges. The chatbot failed to understand needs or provide useful suggestions; outcomes were intentionally ambiguous. Procedure: Participants first reported pre-interaction efficacy expectations for the chatbot. They then read the failure scenario and viewed one of the style-specific chat screenshots. Post-interaction evaluations and demographics followed. Measures: All items used 5-point Likert scales (1 = strongly disagree, 5 = strongly agree). Communication style checks used adapted scales from Van Dolen et al. (2007). Perceived warmth (4 items: caring, friendly, warm, sociable; α = 0.829) and perceived competence (4 items: intelligent, energetic, organized, motivated; α = 0.751) from Judd et al. (2005). Interaction satisfaction from Joosten et al. (2016) (α = 0.913). Trust (ability, benevolence, overall trust; 7 items; α = 0.887) from Bhattacherjee (2002) and Mozafari et al. (2021). Patronage intention (3 items; α = 0.865) from Keeling et al. (2010). Expectancy violations were computed as the difference between pre-interaction expectations (α = 0.712) and post-interaction assessments (α = 0.798), with higher scores indicating greater violations (Crolic et al., 2022). Scenario realism was assessed with three items (α = 0.717). Analysis: Manipulation checks used t-tests. Main effects were tested with one-way ANOVA. Mediation analyses employed PROCESS Model 4 (5,000 bootstrap samples). Moderated mediation tested expectancy violations as a moderator using PROCESS Model 7 (Hayes, 2017).

Manipulation checks: Participants exposed to social-oriented screenshots rated higher on social orientation (Mtask = 2.993, Msocial = 4.044; t(139) = −9.267, p < 0.001). Those exposed to task-oriented screenshots rated higher on task orientation (Mtask = 3.914, Msocial = 3.514; t(139) = 3.171, p = 0.002). Scenario realism did not differ by condition (Mtask = 3.841, Msocial = 4.018; t(139) = −1.464, p > 0.05). Main effects (H1 supported): Social-oriented style increased outcomes relative to task-oriented: - Interaction satisfaction: Mtask = 2.881 (SD = 0.998) vs. Msocial = 3.802 (SD = 0.735); F(1,139) = 39.548, p < 0.001. - Trust: Mtask = 3.269 (SD = 0.868) vs. Msocial = 4.037 (SD = 0.494); F(1,139) = 42.768, p < 0.001. - Patronage intention: Mtask = 2.836 (SD = 1.042) vs. Msocial = 3.820 (SD = 0.796); F(1,139) = 40.139, p < 0.001. Total effects (unstandardized): Interaction satisfaction = 0.922 [0.632, 1.212]; Trust = 0.768 [0.536, 1.000]; Patronage intention = 0.984 [0.678, 1.291]. Mediation (H2): Warmth significantly mediated communication style effects on all outcomes (bootstrapped indirect effects; 95% CI excludes 0): - Interaction satisfaction: b = 0.448 [0.225, 0.709]. - Trust: b = 0.336 [0.160, 0.562]. - Patronage intention: b = 0.358 [0.161, 0.600]. The social style increased perceived warmth (β = 0.683, p < 0.001), which in turn improved interaction satisfaction, trust, and patronage intentions. Competence did not mediate any outcome (indirect effects’ CIs included 0): interaction satisfaction b = 0.037 [−0.025, 0.126]; trust b = 0.018 [−0.013, 0.074]; patronage b = 0.048 [−0.029, 0.184]. Thus, H2a supported; H2b not supported. Moderated mediation (H3): Expectancy violations moderated the warmth-based mediation for social (vs. task) style (indices; 95% CIs exclude 0): - Communication style → Warmth → Interaction satisfaction: 1.677 [1.126, 2.798]. - Communication style → Warmth → Trust: 1.265 [0.730, 2.230]. - Communication style → Warmth → Patronage intention: 1.352 [0.776, 2.343]. Expectancy violations did not moderate competence-based mediation (indices’ CIs include 0). Overall, social-oriented style is particularly effective at higher levels of expectancy violation, via increased perceived warmth.

The findings confirm that chatbot communication style substantially shapes consumer responses after service failures. Social-oriented style elevates interaction satisfaction, trust, and patronage intention relative to task-oriented style, addressing the core question of how to mitigate negative reactions to chatbot failures. The mechanism operates chiefly through perceived warmth—a key mind-perception dimension—rather than perceived competence. This indicates that when failures occur, relational and affective cues are more salient and effective than purely functional cues. Expectancy violations condition these effects: when violations are high, social-oriented communication enhances warmth more strongly, which in turn improves evaluations and intentions. These results underscore the importance of human-like, empathetic language in failure recovery for AI agents, advancing theory by integrating mind perception and expectancy violations into HCI and service recovery research. Practically, deploying socially oriented chatbots can buffer negative emotions and evaluations following failures, though competence cues alone may be insufficient in such contexts.

This study contributes by: (1) focusing on chatbot communication styles under service failure, demonstrating that social-oriented style improves satisfaction, trust, and patronage intentions; (2) identifying perceived warmth as the key mediating mechanism, with competence not mediating in this context; and (3) showing that expectancy violations moderate the warmth-based mediation, making social style particularly beneficial when violations are high. Managerially, firms should configure chatbots to use social-oriented communication during recovery to reduce negative emotions and improve evaluations, while recognizing that complex or emotionally charged cases may still require human agents. Future research should examine real-time interactions (e.g., Wizard-of-Oz), gradations of social orientation, chatbot-initiated interactions, broader demographic moderators, larger samples and multi-study designs, different service domains and emotional contexts, and the role of low-emotion or emotion-neutral bots in specific scenarios.

- Stimuli were static screenshots rather than real-time interactions, potentially limiting ecological validity. - Communication style was manipulated dichotomously (task vs. social), not along a continuum, which may obscure nuanced effects of degree of social orientation. - Focused on consumer-initiated interactions; did not examine chatbot-initiated contacts, which may affect autonomy perceptions and sincerity attributions. - Limited demographic variables (age, gender) collected; potential heterogeneity across user segments remains underexplored. - Modest sample size and single-experiment design; broader generalizability would benefit from larger, multi-study replications. - Specific context (failed online shopping) may limit external validity; other domains and emotional intensities may yield different patterns. - Current AI’s limited emotion capabilities may differentially affect outcomes; future work should consider low- or no-emotion chatbot designs and contexts. - Suggested use of Wizard-of-Oz approaches to enhance realism in future studies.