Relief through digital modalities: evaluating text-based chatbots and voicebots in service failures Open Access

Chatbots are disembodied online AI-driven conversational agents that provide assistance to individuals through text (Liu et al., 2023). Voicebots share most of these features, except that, assistance is voice-based (Klaus and Zaichkowsky, 2020). Although there are obvious technological differences between chatbots and voicebots, such as voicebots using speech recognition, they share many features, including AI and machine learning algorithms (or, now less frequently, rules-based algorithms), integration capabilities and data collection. Both technologies are widely used in various sectors, with common applications in customer service across retail service firms (Klaus and Zaichkowsky, 2020; Rapp et al., 2021), but also with applications in sectors such as healthcare (Sezgin et al., 2020; Singh et al., 2024), as well as emerging sectors such as education (Sharma and Singh, 2021; Tärning and Silvervarg, 2019). Factors like their instantaneous and permanent availability, scalability, intuitive interfaces and potential for a high return on investment are strong selling points (Huang and Dootson, 2022; McLean and Osei-Frimpong, 2019).

In customer settings, a copious body of research has emerged on chatbots and voicebots. Both chatbots and voicebots have been associated with positive customer outcomes such as positive word-of-mouth (Mishra et al., 2022; Yun and Park, 2022), positive feelings (Bilquise et al., 2022; Hernandez-Ortega and Ferreira, 2021) and loyalty (Jenneboer et al., 2022; Shahzad et al., 2024). However, these outcomes, not surprisingly, depend on the particular design of the bots (e.g. language style: Jenneboer et al., 2022). Also, some negative outcomes associated with the use of chatbots and voicebots loom large, for instance, privacy concerns (Agnihotri and Bhattacharya, 2024).

While both chatbots and voicebots have been widely studied individually, and to a much lesser extent comparatively, their relative impact in service failure settings remains largely unexplored (Rohit et al., 2024; Zierau et al., 2023). This oversight is notable given the growing use of both technologies in service contexts (Begum, 2024; Dinh and Park, 2023), the cost implications of adopting one over the other (Akanksha, 2024) and their distinct communicative features, making findings from one modality hard to generalize to the other (Huang and Sénécal, 2023).

In response, this paper seeks to advance the research streams on chatbots and voicebots by contrasting their effects on customers in a service failure setting. Following Smith et al. (1999) and Liu et al. (2023, p. 2), service failures are broadly characterized as “instances where the service provider’s service fails to meet the expectations or needs of the consumer.” This research focuses on service failures resulting from product performance issues, where the company is responsible for the failure (Pagiavlas, 2020; Wei and Shi, 2024). So-called functional failures, where the service failure occurs due to the use of a chatbot or voicebot interaction (i.e. where the bot malfunctions or provides incorrect information, Xing et al., 2022), are not a focus for this research. Rather, the chatbot or voicebot is being used by the consumer as a means by which to rectify a perceived service failure event.

Informed by diverse theoretical approaches and methods, research on chatbots and voicebots in service failure contexts has documented both positive and negative effects. However, their comparative effects in service failure settings remain largely unexplored (see Table 1). Only a few studies have directly compared these modalities (Schindler et al., 2024; Rohit et al., 2024; Rzepka et al., 2022; Zierau et al., 2023) and, to the extent of our knowledge, none have yet examined them in the context of postpurchase service failure.

Most of these studies instead focus on earlier stages of the customer journey, prepurchase and purchase. For instance, Rohit et al. (2024) showed that voicebots, as a richer medium, enhance consumers’ cognitive and affective engagement during prepurchase (e.g. information search) and purchase (e.g. booking) stages of the customer journey for hedonic products versus more utilitarian ones. Similarly, Schindler et al. (2024) found that voice (vs text) interactions promote an emotional (vs reason-based) focus, increasing satisfaction with hedonic (vs utilitarian) product choices, both at the prepurchase stage (e.g. hotel selection) and the purchase stage (e.g. food ordering). Rzepka et al. (2022) demonstrated that in goal-directed search tasks (e.g. finding a restaurant based on cuisine and price), consumers report higher satisfaction when using voicebots versus chatbots.

Regarding the postpurchase stage, Zierau et al. (2023) found that voicebots lead to more flow-like experiences than chatbots when filing insurance claims, resulting in better service evaluations, higher conversion rates and stronger retention. Collectively, these studies suggest that modality effects may vary across the customer journey (Grewal et al., 2022; Lemon and Verhoef, 2016) and research directly comparing these technologies in service failure scenarios remains scarce.

A critical component of service failure is the negative feelings customers experience in such situations (Gelbrich, 2010; Khamitov et al., 2020). Thus, in a service failure context, a fundamental question is how to address such feelings (Balaji et al., 2017; Lee, 2023; Tsarenko and Strizhakova, 2013). Even though research has emerged on bots’ use of intuitive strategies to do so (e.g. chatbots’ service recovery attempts that leverage apologies: Song et al., 2023), there is a lack of understanding as to the differential effects of either text or voice human–robot interactions on consumer feelings derived from these strategies. Here, we contend that perceived relief, defined as the easing and smoothing of tension and negative feelings (Kobel and Groeppel-Klein, 2021; Shurcliff, 1968), is more effectively evoked from interactions with chatbots, given the emotional regulation afforded to customers in interacting with these AI-agents. Our rationale for this prediction builds on the insights offered by emotion regulation theory, which focuses on the factors and strategies that can serve individuals control their emotional reactions (Gross, 1998). Importantly, individuals can engage in emotion regulation at an implicit or subconscious level, without making a deliberate attempt to manage their emotions (Koole et al., 2015).

Specific emotion regulation mechanisms such as passive regulation through time, often referred to as “affective fading” (Skowronski et al., 2014), and attentional deployment (Ferri and Hajcak, 2015; Gross, 2015) can explain why text-based (versus voice-based) interactions can be more conducive to perceived relief. Text is more asynchronous than voice, offering individuals additional time before responding: “Written conversational turns are […] more asynchronous […], so there is less of a rush to produce content” (Berger et al., 2022, p. 391). Voice, by contrast, is faster and more immediate: “speech production is faster than writing because it does not involve the physical process of writing” (Rzepka et al., 2022, p. 842), with “greater synchronicity due to fast-paced and more parallel processing with low revisability” (Zierau et al., 2023, p. 827). Empirical evidence confirms this gap: speaking occurs at roughly 160 wpm versus 50–55 wpm for typing (Ruan et al., 2018). This extra time matters because negative emotions such as anger tend to decrease almost immediately after onset and can even fade naturally within about 90 s if not retriggered (Kulper, 2016; Qiu et al., 2009; Skowronski et al., 2014; Taylor, 2021). Thus, the slower, more deliberate pace of text provides a natural buffer that facilitates affective fading.

In addition, writing requires more cognitive effort and deliberation than speaking: “Writing is more deliberative” (Berger et al., 2022, p. 392); it is “relatively objective, structured, and detached” (Schindler et al., 2024, p. 636); and “producing text involves a process of consciously applying intentionally learnt rules, while humans intuitively build and process speech” (Rzepka et al., 2022, p. 843). This attentional deployment, i.e. redirecting cognitive resources to composing a message, can temporarily shift focus away from the emotional trigger, reducing arousal (Naragon-Gainey et al., 2017). Consistent with these regulation mechanisms, Berger et al. (2022) showed that written word of mouth contains fewer emotionally charged words than spoken word of mouth, while Schindler et al. (2024) found that speaking about choices fosters an emotion focus, whereas writing fosters a reason focus. Across these mechanisms, the common outcome is the downregulation of negative affect and tension, which we conceptualize as perceived relief (Kobel and Groeppel-Klein, 2021).

It is further posited that these enhanced feelings of relief result in enhanced levels of customer service review, i.e. an overall assessment of the service provider that takes into account dimensions such as satisfaction and service quality (Crolic et al., 2021). The ability to alleviate negative emotions plays a crucial role in shaping customer evaluations following a service failure (Nguyen and McColl-Kennedy, 2003). Research demonstrates that strategies aimed at reducing emotional distress lead to better customer outcomes, including increased brand loyalty and trust (Reeck and Onuklu, 2022). In service recovery contexts, mitigating negative emotion has been found to significantly improve customer perceptions of the service provider, as customers who experience emotional relief are less likely to hold negative attitudes toward the company (Lee, 2023). Given that perceived relief reflects the easing of negative emotions in response to a stressful event, it follows that greater perceived relief should translate into more favorable customer reviews.

Based on the above discussion, this research hypothesizes that, in a service failure setting, chatbots (versus voicebots) will be associated with increased customer review evaluations, an effect that will be driven by increased perceived relief (i.e. the easing and smoothing of tension and negative feelings). Formally:

This work additionally advances the chatbot and voicebot fields by examining service failure severity as a moderator of this mediation effect.

Not all service failures are perceived equally. Service failures can be differentiated along several dimensions, such as responsibility (Leo and Huh, 2020), frequency and timing (Sivakumar et al., 2014) and severity (Leo and Huh, 2020; Weun et al., 2004). In this work, one of the foci is the severity of the failure, defined as the perceived criticality or intensity of the service problem (Weun et al., 2004), whereby more severe failures result in more intense negative feelings, which lead to decreased customer satisfaction (Smith et al., 1999). Service failure severity has been frequently used as a moderator in service research (de Mesquita et al., 2023; Fouroudi et al., 2020) and also operationalized dichotomously, as done here (e.g. Fox et al., 2018; Hess et al., 2003). In the context of robots, for instance, Shams et al. (2024) established that humor and informal language can be successfully used by text-based chatbots in service recovery attempts as long as the severity of a failure is low, an effect that is reversed for high-severity failures. Furthermore, Lv et al. (2021) found that voicebots using a cute voice increase customers’ tolerance of service failures, but again, only when the severity of the failure is low.

Building on the preceding section, we argue that the effect of modality (text vs voice) on perceived relief is contingent on the severity of the service failure. Emotion regulation theory (Gross, 1998, 2015) suggests that individuals deploy regulation strategies based on situational appraisals, such as urgency or threat (Matthews et al., 2021). When service failures are minor, customers can retain both the motivation and cognitive capacity to regulate their emotions, as suggested before (Ferri and Hajcak, 2015; Skowronski et al., 2014).

However, in high-severity situations, customers are likely to deprioritize emotional regulation in favor of swift problem resolution (Gross, 2015). When urgency is high, individuals are less inclined to invest cognitive resources in regulating emotions, especially when those emotions may functionally support goal attainment, in this case, problem resolution (Tamir and Ford, 2009; Glikson et al., 2019). As cognitive demands rise as a result of a more severe failure, the regulatory benefits of chatbots, which require effortful input (Tindle and Longstaff, 2015), become less accessible. Voice-based interactions, while less conducive to relief under low-severity conditions, due to their lower cognitive demands, may become more intuitive when mental resources are depleted (Sweller et al., 2011; Hawthorne et al., 2019). Consequently, the relief advantage of chatbots is expected to diminish as failure severity increases.

In essence, when customers face severe service failures, emotional intensity is so high and resolution so urgent that neither the slower, self-paced nature of text nor the lower-effort fluidity of voice holds a meaningful edge; what matters is resolving the issue. Thus, we predict:

Figure 1 demonstrates the conceptual model.

To empirically investigate our research questions and test the proposed hypotheses, we used a multistudy research design comprising one field study and two controlled experiments. Study 1 serves as an initial exploratory field study conducted in a real-world service failure setting, aimed at establishing the baseline effect of AI bot modality on customer evaluations. This naturalistic study provides ecological validity and grounds the research in actual consumer behavior. More importantly, the initial field study was particularly important in establishing the real-world relevance of the phenomenon, namely, that differences in customer responses to AI bot modalities are not only theoretically meaningful but also practically prevalent for online service firms. As many field studies, it provides preliminary evidence for the effect of modality on important customer outcomes. Study 2 is an online experiment designed to test the underlying psychological mechanism (i.e. perceived relief) that mediates the relationship between AI bot modality and customer review. Finally, Study 3 examines the moderating role of service failure severity. Together, these studies provide a comprehensive and theory-driven examination of why and when chatbot (vs voicebot) interactions enhance customer outcomes in service failure contexts.

To explore whether the effect of AI bot modality (chatbot vs voicebot) is observable in real-world service settings, Study 1 provides preliminary, ecologically valid evidence for RQ1 and H1. Specifically, it examines whether customers interacting with a chatbot report higher satisfaction ratings than those interacting with a voicebot following a postpurchase service experience. As a naturalistic field setting, Study 1 necessarily involved channel differences (Web vs phone, as indicated below) and customer self-selection, which may confound modality effects. Rather than providing definitive causal evidence, the purpose of Study 1 is to illustrate that meaningful differences can emerge in practice, thereby motivating the controlled laboratory experiments in Studies 2 and 3.

We conducted a field experiment in collaboration with a service firm within the retail IT sector. The service firm operates as a local computer dealer located in the capital city of an eastern province in China and provides both in-store and online retail services (preregistered: Link to AI bot in service failure (field experiment)Link to the cited article.). As part of their postsales customer service infrastructure, the company uses two AI-driven service agents: a text-based chatbot embedded in their online service portal and a voicebot integrated into their telephone hotline system for any product complaint. These tools are used to assist customers with troubleshooting product issues, obtaining service information and resolving complaints outside of normal business hours. The aim of the field experiment was to examine how different AI bot modalities (i.e. chatbot vs voicebot) affect customer reviews in a service failure context. As such, this real-world setting allowed us to observe how bot modality influences customer evaluations under ecologically valid conditions. The experiment was embedded within the IT service provider’s existing postsales service framework and ran for one month. Customers who encountered computer-related issues could visit the company’s service website and interact with an AI-driven chatbot (text-based interaction) by clicking the “Online Customer Service” option displayed on the webpage. Alternatively, they could also call the customer service hotline (the number was provided beneath the “Online Customer Service” option) and interact with an AI-based voicebot (voice-based interaction).

Participation was naturalistic and nonintrusive such that customers were not informed that an experiment was being conducted and self-selected into one of the two AI bot modalities based on their preferred channel of service. After the interaction with the chatbot and the voicebot, customers were prompted to provide a satisfaction rating on a five-star scale (1 = very dissatisfied, 5 = very satisfied), specifically reflecting their assessment of the AI agent’s ability to respond to their service request and manage the recovery interaction, by selecting the stars on the webpage or clicking on a number on phone’s dial keypad (see  Appendix). These customer ratings served as the primary dependent variable.

Customer responses were automatically recorded through the company’s website and phone system, depending on the modality used. The field experiment lasted one month, and a total of 138 customer ratings (61 from voicebot and 77 from chatbot) were received. As this was a naturalistic field setting, no participants were excluded, and no outlier removal was conducted. The data reflects real customer behavior in an ecologically valid service context.

We conducted an independent sample t-test to examine whether and how AI bot modality affects consumer ratings. As expected, consumers who interacted with chatbot in the Web page (M_Chatbot = 4.12, SD_Chatbot = 0.628) rated the AI services significantly higher than those who chatted with AI via the hotline (M_Voicebot = 3.03, SD_Voicebot = 0.795; p < 0.001, Cohen’s d = 1.54). Thus, this finding offers preliminary field-based evidence for H1, demonstrating that bot modality differences manifest in real customer evaluations. Given the ecological nature of this setting and the potential for confounds (e.g. channel and severity), the result is best interpreted as suggestive evidence that motivates further controlled studies. Importantly, this study provides a strong rationale for systematically testing the underlying mechanisms (e.g. perceived relief) and boundary conditions (e.g. failure severity and outcome expectations) in Studies 2 and 3.

Building on the findings of Study 1, which confirmed that chatbot interactions yield more favorable customer evaluations, Study 2 addresses RQ2 and empirically tests H2 by investigating the underlying psychological mechanism of perceived relief that may explain this effect.

We recruited 160 participants (48% females, M_age = 33, SD = 13) located in the USA from Prolific. Participants were randomly assigned to a 2× AI-driven bot (chat vs voice) modality between-subjects design.

Participants were provided with a service scenario where they were required to imagine that they had bought a camera specifically for an upcoming trip. Upon receiving the camera, they excitedly open it and snap a few pictures, only to realize they are disappointingly blurry. They suspect the lens might be cracked or misaligned. They search online for guidance on how to exchange the camera, only to find that the estimated delivery date for a replacement falls after the departure date for their trip, defeating the whole purpose of the purchase. Frustration sets in, and they contemplate reaching out to the camera company. Unfortunately, it is outside of their business hours, so they opt to explore their online customer service center instead. In the scenario with the AI chatbot, a window pops up, and there is a conversation with the chatbot about how to fix the issue. In the scenario of a voicebot, the participants discover the phone number for the company’s Automated Customer Service Center, they make the call, and listen to the conversation with a voicebot. The full scenario text and AI bot interaction transcripts are provided in the  Appendix.

All measures were adapted from established scales. For perceived relief (Kobel and Groeppel-Klein, 2021), participants were asked to indicate the extent to which they agree or disagree on a five-point scale with the following statements (α = 0.92): “The bot relaxed the situation,” “I perceived the bot as liberating,” “The bot lightened the atmosphere,” “The bot defused the situation” and “The bot enhanced the mood.” Customer review was measured by asking the respondents to evaluate the service provided by the bot: overall satisfaction, customer service, helpfulness and problem resolution (Crolic et al., 2021), measured on a five-point scale (1 = 1 star, 5 = 5 stars).

We conducted Confirmatory Factor Analyses (CFAs) to ensure the robustness of measurement instruments. The data were evaluated for convergent and discriminant validity (see  Appendix). The values of Cronbach’s α, Composite Reliability (CR) and Average Variance Extracted (AVE) were all greater than the cut-off of 0.7, 0.7 and 0.5, respectively (Hair et al., 2019). Thus, we confirm that our measures met the reliability and convergent validity criteria. As depicted in the correlation matrix (see  Appendix), the square root values of AVE (on the diagonals) are greater than the interconstruct correlational values (on the left of the diagonals), thereby indicating that discriminant validity was verified (Fornell and Larcker, 1981). To address potential common method bias, we compared the measurement model (i.e. perceived relief and customer review) against a single-factor model combining both perceived relief and customer review together and used the Unmeasured Latent Method Construct (ULMC) approach by incorporating an unmeasured latent variable connected to all observed items on the basis of the measurement model (Podsakoff et al., 2012). Across all studies, the fit indices for the measurement model (χ ²/df = 1.801, CFI = 0.980, TLI = 0.973, SRMR = 0.033) were significantly better than those for the single-factor model (χ ²/df = 4.743, CFI = 0.902, TLI = 0.874, SRMR = 0.054) and were not significantly different from those for the measurement model incorporating an unmeasured latent variable (χ ²/df = 1.478, CFI = 0.988, TLI = 0.984, SRMR = 0.050). These results suggest that common method bias did not pose a significant issue in any of the studies. Overall, these results suggest that we could proceed with hypotheses testing.

We conducted a separate pretest with 101 Prolific participants based in the USA, who were randomly assigned to one of two AI bot modality conditions (chatbot vs voicebot) in a between-subjects design. Participants evaluated the scenario’s believability and clarity on a five-point scale (1 = not at all, 5 = very much). They also rated the extent to which they would feel frustrated when encountering the described issue (1 = not at all, 5 = very much).

As expected, we found no significant differences between conditions; participants in both the chatbot and voicebot groups perceived the scenario as similarly believable [M_chat = 3.73, M_voice = 3.58, t(99) = 0.75, p = 0.456] and clear [M_chat = 4.37, M_voice = 4.26, t(99) = 0.74, p = 0.460]. In addition, reported frustration levels were comparable across conditions [M_chat = 4.29, M_voice = 4.40, t(99) = 0.63, p = 0.530]. These results suggest that believability, clarity and frustration were unlikely to introduce potential confounds.

An independent sample t-test revealed that participants in the chatbot condition (M = 3.83) reported higher levels of customer review than those in the voicebot condition [M = 3.53, t(158) = 2.15, p = 0.033], further supporting H1.

We proposed that perceived relief mediates the impact of AI bot modality on customer review. An independent sample t-test revealed that participants in the chatbot condition (M = 3.47) reported higher levels of perceived relief than those in the voicebot condition [M = 3.14, t(158) = 2.18, p = 0.031]. A mediation analysis was performed (Hayes, 2017, Model 4). This analysis examined the indirect effects of AI bot modality types as a categorial variable (voicebot = 0, chatbot = 1), on customer review via perceived relief. AI bot modality types had a significant impact on perceived relief. Perceived relief was a significant predictor of customer review. The mediation effect was significant (b = 0.24, SE = 0.11, 95% CI: 0.03–0.45). See Figure 2 and Table 2. Thus, H2 was supported.

While Study 2 establishes the mediating role of perceived relief, it does so in a generalized service failure context. To further refine the conditions under which this mediation holds, Study 3 addresses RQ3 and empirically tests H3 by examining whether the severity of the service failure moderates the indirect effect of bot modality on customer evaluations via perceived relief.

We recruited 320 participants (50% females, M_age = 43, SD = 14) located in the USA from Prolific. Participants were randomly assigned to a 2 AI bot modality (chat vs voicebot) × 2 service failure severity (high vs low) between-subjects design.

In the high-severity scenario, participants were required to imagine that they had recently purchased a brand-new refrigerator, and after just a few days of use, it had completely stopped working. Their food was at risk of spoiling, and they were facing an urgent situation. They decided to reach out to the refrigerator company for assistance. In the low-severity scenario, participants were required to imagine that they had recently bought a new refrigerator, but they have been experiencing intermittent problems with its cooling system. Sometimes, the temperature seemed fine, but on other occasions, they noticed that their food was not as cold as it should be. It was not an immediate emergency, but a concern, and they wanted to address it with the refrigerator company. Unfortunately, it was outside of their business hours, so they opted to explore their online customer service center. Participants either saw a conversation with a chatbot or listened to a conversation with a voicebot. The full scenario description and AI bot interaction transcripts are provided in the  Appendix.

Perceived relief and customer review were measured with validated scales as adopted in Study 2. In a manner parallel to Study 2, we conducted the CFA tests and the results showed that the computed values for Cronbach’s α, CR and AVE significantly surpassed the criteria of 0.7, 0.7 and 0.5, respectively. The discriminant validity was also achieved, as the square root values of AVE were greater than the interconstruct correlations (see  Appendix). We also performed a comparison among three distinct models: the two-factor measurement model, the single-factor model and the three-factor model, introducing an unmeasured latent variable (ULMC). The fit indices of the two-factor measurement model (χ²/df = 3.630, CFI = 0.976, TLI = 0.967, SRMR = 0.028) exhibited marked improvement when compared to the single-factor model (χ²/df = 24.424, CFI = 0.780, TLI = 0.707, SRMR = 0.095), while the model incorporating an unmeasured latent variable (χ²/df = 2.904, CFI = 0.984, TLI = 0.976, SRMR = 0.026) yielded similar fit indices. Thus, common method bias was not a significant concern.

We conducted a separate pretest with 160 Prolific participants based in the USA, who were randomly assigned to a 2 AI bot modality (chat vs voicebot) × 2 service failure severity (high vs low) between-subjects design. As in Study 2, participants evaluated the scenario’s believability and clarity. They also rated the extent to which they would feel frustrated when encountering the described issue, as well as the perceived severity of the issue. All items were measured on a five-point scale (1 = not at all, 5 = very much).

We conducted 2 × 2 ANOVAs and found no significant differences across conditions in how participants perceived the scenario’s believability and clarity. Specifically, we found nonsignificant main effects of AI bot modality [believability: F(1, 156) = 1.62, p = 0.204; clarity: F(1, 156) = 0.72, p = 0.398], service failure severity [believability: F(1, 156) = 0.78, p = 0.379; clarity: F(1, 156) = 0.37, p = 0.546] and interaction effects [believability: F(1, 156) = 2.16, p = 0.143; clarity: F(1, 156) = 0.72, p = 0.398].

However, we found significant main effects of service failure severity on frustration and perceived severity. Specifically, participants in the high-severity condition reported higher levels of frustration [M_high = 4.50, M_low = 3.52, F(1, 156) = 37.11, p < 0.001] and perceived the issue as more severe [M_high = 4.11, M_low = 3.28, F(1, 156) = 35.48, p < 0.001] than those in the low-severity condition. However, the main effects of AI bot modality [frustration: F(1, 156) = 0.10, p = 0.755; perceived severity: F(1, 156) = 2.28, p = 0.133] as well as the interaction effects were nonsignificant [frustration: F(1, 156) < 0.001, p = 1.00; perceived severity: F(1, 156) = 0.96, p = 0.330].

Customer review. Setting customer review as the dependent variable, a two-way ANOVA analysis was performed with independent variables including AI bot modality and service failure severity and their interaction. Results revealed a nonsignificant main effect of AI bot modality [F(1, 315) = 1.32, p = 0.252] and a significant main effect of service failure severity [F(1, 315) = 51.92, p < 0.001]. Moreover and as predicted, we also found a significant interaction effect [F(1, 315) = 4.23, p = 0.041]. Specifically, in the low-severity condition, participants in the chatbot condition (M = 4.22) reported higher levels of customer review than those in the voicebot condition [M = 3.89, F(1, 315) = 5.12, p = 0.024]. However, such differences were nonsignificant in the high-severity condition [M_chat = 3.25, M_voice = 3.35, F(1, 315) = 0.42, p = 0.520].

Perceived relief. Setting perceived relief as the dependent variable, a two-way ANOVA analysis was performed with independent variables including AI bot modality and service failure severity and their interaction. Results revealed a nonsignificant main effect of AI bot modality [F(1, 315) = 2.71, p = 0.101] and a significant main effect of service failure severity [F(1, 315) = 35.48, p < 0.001]. Moreover, and as predicted, we also found a significant interaction effect [F(1, 315) = 4.85, p = 0.028]. Specifically, in the low-severity condition, participants in the chatbot condition (M = 3.49) reported higher levels of perceived relief than those in the voicebot condition [M = 3.07, F(1, 315) = 7.38, p = 0.007]. However, such differences were nonsignificant in the high-severity condition [M_chat = 2.58, M_voice = 2.65, F(1, 315) = 0.16, p = 0.693].

Moderated mediation analysis. We propose that the AI bot modality types on customer reviews are mediated by perceived relief and moderated by service failure severity. A moderated mediation analysis (Hayes, 2017, Model 7) was conducted. AI bot modality types as a categorial variable (voicebot = 0, chatbot = 1) affect customer review, moderated by service failure severity (low severity = 0, high severity = 1), via perceived relief.

AI bot modality significantly predicted perceived relief, and in turn, perceived relief significantly predicted customer review. The index of moderated mediation was significant and negative (b = −0.33, SE = 0.15, 95% CI: −0.63 to −0.04), such that the indirect effect was significant in the low-severity condition (b = 0.29, SE = 0.10, 95% CI: 0.09–0.49) but not in the high-severity condition (b = −0.04, SE = 0.11, 95% CI: −0.27–0.18). See Table 3. H3 was supported.

This research extends existing theory on AI service interactions, specifically human–robot interactions, by demonstrating that chatbots and voicebots elicit different consumer responses in service failures, with chatbots (versus voicebots) significantly increasing perceived relief and, via this mechanism, having a positive influence on customer reviews. In addition, this research advances theoretical understanding by identifying a key moderator of this effect. Specifically, we show that the severity of the service failure critically determines the impact of bot modality on perceived relief and, ultimately, customer evaluations.

These findings make significant contributions to several research streams. First, we advance research on AI-driven customer service by providing one of the first empirical comparisons between bot modalities (i.e. chatbots and voicebots). While prior research has explored each modality separately (e.g. Agnihotri and Bhattacharya, 2024; Cai et al., 2024; Lv et al., 2021; Li et al., 2023), our study is one of the first to systematically examine their differential effects in service failures. While a few studies have compared these modalities (e.g. Rohit et al., 2024; Rzepka et al., 2022; Schindler et al., 2024), they did not examine their effects in failure scenarios, where emotional responses are heightened and customer expectations shift dramatically (Obeidat et al., 2017; Valentini et al., 2020).

Second, this research makes a novel contribution to the service failure literature by directly comparing the two prevailing AI bot modality interactions (i.e. chatbot and voicebot), an empirical contrast largely absent in prior service failure studies, which have typically examined these modalities in isolation (Agnihotri and Bhattacharya, 2024; Crolic et al., 2021; Huang and Sénécal, 2023; Li et al., 2023). In doing so, this study challenges prevailing assumptions in service failure research, which often suggest that more humanlike or richer media, such as voice-based interfaces, are inherently more effective in emotionally charged service contexts. While anthropomorphizing AI service agents has been widely regarded as an effective strategy (de Visser et al., 2016; Huang and Sénécal, 2023; So et al., 2024), our findings suggest that in certain circumstances, less humanlike interactions, such as text-based chatbots, can be more beneficial in mitigating negative emotional reactions (Crolic et al., 2021; Li et al., 2023). This finding contradicts the assumption that richer, more immersive media (such as voice; Rohit et al., 2024) are inherently superior in emotionally charged service contexts.

Third, this study extends emotion regulation theory by identifying bot modality as a factor that influences how consumers regulate emotions in service failures. Service failures inherently elicit negative emotions (Balaji et al., 2017), yet prior research has barely examined how different communication modalities facilitate or hinder emotion regulation during these events. While Berger et al. (2022) showed that consumers express more emotionality when speaking rather than writing in product reviews, this research extends that insight to service failure settings, demonstrating that chatbots (text) are more conducive to emotion regulation, as evidenced by increased perceived relief. Our findings are consistent with the idea that chat-based interactions facilitate mechanisms such as affective fading and attentional deployment (Ferri and Hajcak, 2015; Skowronski et al., 2014).

Finally, this work adds to the literature exploring the construct of perceived consumer relief. In service failure settings, this construct has received limited attention (Kobel and Groeppel-Klein, 2021). Kobel and Groeppel-Klein (2021) investigated how humorous remarks lead to relief, demonstrating that content, i.e. what a service provider says, positively influences perceived relief. Here, it has been established that perceived relief is influenced not only by content but also by form, i.e. not what is said, by either the customer or service provider, but also how it is said, or the medium chosen for communication (text or voice).

By identifying why and when chatbots outperform voicebots in service failure settings, this research provides a novel understanding of how different AI modalities influence customer emotions and evaluations. It also lays the foundation for future work exploring the wider impact of AI communication modality across various service settings.

The findings of this study provide valuable insights for managers seeking to optimize customer service strategies, particularly in the context of service failures. The findings indicate that chatbots (text-based) lead to increased perceived relief and customer review compared to voicebots, especially in instances of nonsevere service failures. Therefore, managers should consider deploying chatbots as a response mechanism in these scenarios to improve customer outcomes.

Several factors might encourage the adoption of voicebots to the detriment of chatbots. First, excitement surrounding the adoption of voicebot technology due to its complexity (e.g. its use of natural language processing and speech recognition) might favor voicebots. The extensive media coverage of ChatGPT-o’s voice capabilities, which makes the LLM sound remarkably human, provides a clear example (Scharth, 2024). Second, there is a strong tendency (which, presumably, extends beyond the academic sphere) to assume that interactions with robots should be “humanized”, i.e. made to approximate, to the largest extent possible, the complexities of human-to-human exchanges (Crolic et al., 2021). Speech is an inherently richer form of communication than text, so voicebots once again look like a promising route to pursue in this regard (Luo et al., 2019; Rohit et al., 2024). The findings of this study suggest that this approach may not always yield the best results and that, indeed, in service failure settings (and, particularly, those involving low-severity failures), chatbots are the optimal strategy.

Based on the findings of this study, it would be a stretch to conclude that chatbots are invariably better. The dynamic nature of customer preferences and technological advancements necessitates continuous evaluation and adaptation of customer service strategies. Managers should regularly review the performance of both chatbots and voicebots, incorporating customer feedback and technological improvements to optimize their service delivery approaches. By understanding and applying these insights, managers can better navigate service failures, ultimately leading to improved customer outcomes.

This study has several limitations, and future research directions are suggested. First, it is important to note that while this research found that chatbots are preferable under certain circumstances, it cannot be concluded that they are invariably better. Future research should examine alternative moderators that might influence these preferences. For instance, process and outcome failures emphasize, respectively, the service delivery or its result (e.g. a product) (Fouroudi et al., 2020). The scenarios chosen in this study mainly refer to outcome failures (camera and refrigerator). It would be interesting to explore if process failures, where the problem lies in the way the service is delivered rather than the end result, negate or reverse the patterns found in this research.

Second, consistent with prior research on service failures (e.g. Fox et al., 2018; Shams et al., 2024), we used a dichotomous manipulation of failure severity. Future research could extend this approach by using a continuous measure to capture finer variations in how bot modality interacts with failure severity.

Third, while this study explored a diverse set of products, future research could further investigate the role of product value in shaping customer expectations and responses to AI service interactions. High-value products, particularly those in luxury markets, may have customers who expect personalized, 24-h human service availability (Kapferer and Bastien, 2012), which could reduce reliance on chatbots and voicebots for service interactions. However, as luxury brands begin integrating AI-driven service tools (Chung et al., 2020), understanding the boundary conditions under which such technologies influence the customer experience represents an important avenue for future research.

Fourth, one key limitation of this study is the lack of a pre- and postmeasurement of frustration, which could have provided deeper insights into how emotional responses evolve over time in reaction to service failures and AI-driven resolutions. While our findings demonstrate that frustration levels vary based on severity assessments, capturing frustration before and after the intervention would allow for a more precise assessment of AI-driven recovery strategies. Future research could adopt a longitudinal or real-time tracking approach to examine moment-to-moment changes in frustration and their subsequent impact on customer perceptions, satisfaction and loyalty.

In addition, while our randomized experimental design controls for individual differences, future studies could further explore the role of personal factors such as technology familiarity, AI aversion and service expectations, which may moderate consumers’ emotional and behavioral responses to AI interactions.

Finally, based on prior work (Crolic et al., 2021), this study used an overall measure of satisfaction as the dependent variable. Future research might examine similar effects with other dependent variables, such as customer loyalty, trust or perceived value. Investigating these additional dimensions could provide a more comprehensive understanding of how chatbots and voicebots influence various aspects of customer experience and behavior.

(a) The local dealer website and online chatbot

(b) Voicebot greeting transcript

Chinese (origin): 您好, 我是智能语音助手小智, 欢迎致电 XX电脑售后服务中心。我们致力于为您提供高效便捷的服务支持。在接下来的通话中, 我会根据您的语音描述, 识别您遇到的问题, 并为您提供相应的解决方案。请尽量用简洁明确的语言描述您的问题, 比如“电脑开不了机”、“屏幕黑屏”、“键盘失灵”等。感谢您的耐心, 现在请您开始描述您遇到的问题。

English (Translation): Hello, this is Smart Voice Assistant Xiao Zhi. Thank you for calling the XX Computer After-Sales Service Center. We’re here to provide you with fast and convenient support.

During this call, I will listen to your voice input, identify the issue you’re experiencing and offer the most appropriate solution. Please try to describe your problem clearly and briefly, such as “the computer won’t turn on,” “the screen is black” or “the keyboard is not working.” Thank you for your patience. Now, please go ahead and describe the issue you’re experiencing.

Imagine you have bought a camera specifically for an upcoming trip. Upon receiving the camera, you excitedly open it and snap a few pictures, only to realize they are disappointingly blurry. You suspect the lens might be cracked or misaligned. You search online for guidance on how to exchange the camera, only to find that the estimated delivery date for a replacement falls after your departure date – defeating the whole purpose of your purchase. Frustration sets in, and you contemplate reaching out to the camera company. Unfortunately, it is outside of their business hours, so you opt to explore their online customer service center instead.

A chatbot window pops up, and the conversation between you and the chatbot is as follows. You discovered the phone number for the company’s Automated Customer Service Center. You made the call, and the conversation between you and a voicebot is below.

Jamie: Hey there, I’m Jamie, your Customer Service Assistant. Thanks for getting in touch with our Automated Customer Service Center. What can I help you with today?

Your response: Hi Jamie, I’m honestly so frustrated right now. I bought a camera from your company last week, and the pictures I took are all unbelievably blurry. I’ve got a trip coming up, and I really need this sorted out.

Jamie: I’m really sorry to hear that you’re going through this. Did you know about our Advanced Video Stabilization software? It’s designed to fix that exact issue caused by camera movement. You can download it from our website.

Your response: No, I had no clue about that software. This whole situation is just messing with my plans for the trip, and it’s driving me crazy.

Jamie: I totally understand your frustration, and I’m sorry for the inconvenience. Downloading the software might make a significant difference. Hope it helps salvage your photos before your trip.

Your response: Thanks for the info. I’ll give it a shot, though I’m seriously hoping it works. This trip means a lot to me. Thanks for your help. I appreciate it.

Jamie: No problem at all. Wishing you the best with your photos and your trip. If anything else comes up, don’t hesitate to reach out. Take care!

Imagine you recently purchased a brand-new refrigerator, and after just a few days of use, it has completely stopped working. Your food is at risk of spoiling, and you are facing an urgent situation. You decide to reach out to the refrigerator company for assistance.

Imagine you have recently bought a new refrigerator, but you have been experiencing intermittent problems with its cooling system. Sometimes, the temperature seems fine, but on other occasions, you notice that your food is not as cold as it should be. It is not an immediate emergency, but it is definitely a concern, and you want to address it with the refrigerator company.

Unfortunately, it is outside of their business hours, so you opt to explore their online customer service center instead.

A chatbot window pops up, and the conversation between you and the chatbot is as follows.

You discovered the phone number for the company’s Automated Customer Service Center. You made the call, and the conversation between you and a voicebot is below.

Jenny: Hey there, I’m Jenny, your Customer Service Assistant. Thanks for getting in touch with our Automated Customer Service Center. What can I help you with today?

Your response: Hi, I recently purchased one of your refrigerators, and it’s completely stopped working after just a few days. I’m really worried because all my food is inside, and it’s at risk of spoiling. Can you please help me with this urgent issue?

Jenny: I’m sorry to hear about the issue with your refrigerator. We understand the urgency. Let’s try to troubleshoot this together. Have you checked if it’s plugged in properly, and is the power outlet working?

Your response: Yes, I’ve checked that, and everything seems fine on that end. It’s just not cooling at all.

Jenny: Thank you for confirming that. It might be a more serious issue. Can you provide your purchase reference number, so we can verify the warranty status and will schedule a technician to come and inspect the refrigerator as soon as possible?

Your response: Sure, my reference number is 1368. Please send someone as soon as you can; I’m really concerned about my food.

Jenny: Thank you for the information. We’ll prioritize your case and get a technician out to you as soon as possible. We apologize for any inconvenience this has caused and appreciate your patience.

Jenny: Hey there, I’m Jenny, your Customer Service Assistant. Thanks for getting in touch with our Automated Customer Service Center. What can I help you with today?

Your response: Hello, I recently bought one of your refrigerators, and I’ve been having some issues with the cooling system. It’s not consistently keeping my food as cold as it should be. It’s not an emergency, but I’d like to address this issue.

Jenny: Thank you for reaching out. I’m sorry to hear about the cooling problem you’re experiencing. Let’s try to resolve this. Have you checked if the temperature settings on the refrigerator are correctly adjusted?

Your response: Yes, I’ve double-checked the settings, and they seem to be okay. It just seems to work fine sometimes and not so well at other times.

Jenny: I see. It could be a minor issue, but we want to make sure it’s resolved for you. Can you provide your purchase reference number, so we can verify the warranty status and schedule a technician to come and inspect the refrigerator?

Your response: Sure, my reference number is 1368. Please let me know when someone can come take a look.

Jenny: Thank you for providing the information. We’ll schedule a technician to investigate the issue and make the necessary repairs. We appreciate your patience while we work to resolve this for you.

Reliability and validity tests