Impact of AI-Oriented Live-Streaming E-Commerce Service Failures on Consumer Disengagement—Empirical Evidence from China

Abstract

Despite the popularity of AI-oriented e-commerce live-streaming, the service failures that can result from real-time interaction and instant transactions have not been taken seriously. This study aims to assess the failure of AI-oriented live-streaming e-commerce services and help retailers identify various risks. Based on expectancy disconfirmation theory and a stressor–strain–outcome framework, this study identified a comprehensive framework including information, functional, system, interaction, and aesthetic failures. The structural equation modeling (SEM) method is used to further examine its effect on consumers’ discontinuance behavior. Further research reveals the mediating role of consumer disappointment and emotional exhaustion, as well as the moderating role of the live-streaming platform type. These results shed light on the negative influence of AI-oriented live-streaming e-commerce service failures and contribute to the literature on live-streaming commerce, service failure, and virtual streamers.

1. Introduction

The development of artificial intelligence and augmented reality technology in recent years has promoted a comprehensive reform of business activities. In the field of live-streaming e-commerce, a new model based on AI technology to replace human streamers for live-streaming is becoming a trend and has been widely tried by enterprises. This model is called AI-oriented live-streaming [1]. For example, well-known companies such as the sports brand NIKE and the British fashion brand Boohoo have created their own AI streamers to achieve 24 h uninterrupted live-streaming at a lower cost, providing consumers with a more interesting and immersive shop** experience, and avoiding the potential risks of the personal behavior or scandals of human streamers [2]^. According to a research report on the Chinese virtual streamer industry in 2022, the size of the industrial market driven by virtual streamers is CNY 186.61 billion, which is expected to reach CNY 640.27 billion in 2025, showing a strong growth trend [3].

However, the rapid expansion of AI-oriented live-streaming e-commerce has led to diverse issues and uncertainties [4]. SONY, for example, launched a virtual streamer called Sony-N to demonstrate and promote products, but consumers complained of obvious service failures, such as invalid responses, wrong information, and a lack of empathy. After many live-streaming businesses adopted AI streamers, sales decreased rapidly, and in some cases the number of viewers was less than 10 [5]. According to the China Consumers Association (2023), the number of lawsuits related to AI-oriented live-streaming in China witnessed a 44% increase in 2023. Therefore, it is crucial to understand and avoid AI-oriented live-streaming e-commerce service failures for mobile commerce development.

Research on AI-oriented live-streaming commerce is still in its infancy. The existing related studies mainly involve discussions on the key technologies of AI [6] and the delivery process [7], or qualitative reviews and the prospects of the development process and applications of AI streamers [8]. In general, the recent literature only emphasizes the positive results of AI-oriented live-streaming e-commerce, such as attracting consumers’ engagement [5,9], bringing social presence and enjoyment to an audience [2,10], increasing viewers’ purchase intentions [1,11,12], and promoting parasocial interaction intentions or the urge to buy impulsively [13]. However, there is no literature examining the “dark side” of AI-oriented live-streaming e-commerce. Hu and Ma’s [12] research pointed out that virtual streamers lack true emotional expression and interpersonal communication skills compared with human streamers. So, it is crucial to recognize AI-oriented live-streaming e-commerce service failure as a new phenomenon and explore its potential negative consequences for the sustainable development of the live-streaming industry.

On the basis of the above research gaps and expectancy disconfirmation theory (EDT), this study attempts to clarify the connotations of AI-oriented live-streaming e-commerce service failures, and explore their impacts on consumers’ negative co** behaviors and psychological mechanisms. We identified five dimensions of AI-oriented live-streaming e-commerce service failure from the perspective of consumers’ inconsistent expectations, including information, functional, system, interaction, and aesthetic failures. Furthermore, this study discusses the impact of this failure on consumer disappointment, emotional exhaustion, and live-streaming discontinuance behavior based on the stressor–stress–outcome model (SSO). Finally, we investigated how different live-streaming platform types moderate consumer responses to AI-oriented live-streaming e-commerce service failures. Our research significantly enriches the literature in several key areas. Firstly, this study contributes to service marketing by pioneering an exploration of the connotations and dimensions of AI-oriented live-streaming e-commerce service failure, introducing e-commerce service failure into the context of AI live-streaming. Secondly, this study expands the literature of consumer engagement by analyzing the impact of AI-oriented live-streaming e-commerce service failure on consumers’ non-sustainable participation behavior. By strategically considering the matching of live-streaming platform types with consumer expectations, live-streaming merchants can improve the effectiveness and accuracy of AI streamer services, thus better attracting consumer participation and purchase. Overall, this study provides practical implications for AI-oriented live-streaming practitioners to identify the causes and effects of service failures, thereby guiding them to improve their business models and service strategies.

2. Literature Review and Theoretical Background

2.1. AI-Oriented Live-Streaming E-Commerce and Service Failure

AI streamers (that is, computer-produced artificial intelligence characters) integrate advanced perceptual intelligence and cognitive intelligence technologies, covering multi-modal fusion algorithms such as vision, speech, and natural language processing, and can show realistic expressions and actions [5]. In addition to broadcasting display products and explaining preferential mechanisms, AI streamers can also interact with consumers in real time, including greeting, guiding purchases, answering questions, and so on [11]. Virtual streamers can continuously broadcast and maintain a perfect brand image, possessing low operational costs, all-weather operation, high plasticity, a stable state, and other advantages, playing a vital role in attracting consumer attention and improving engagement [13,14]. At present, AI-oriented live-streaming has been widely used in practice, such as in the entertainment industry, advertising marketing, customer service, and medical health, but the impact of AI live-streaming in e-commerce services remains largely unexplored [15].

AI-oriented live-streaming e-commerce service failure stems from live-streaming service failure, in which a live service provider fails to meet viewers’ expectations of a product, or its live-streaming service behavior is rated unsatisfactory by viewers [4]. Stock et al. (2018) [16] found that compared with humanoid service robots, consumers have higher satisfaction and pleasure with the service behaviors of real employees. Luo et al.’s (2019) [17] research shows that disclosing a chatbot’s identity before the chatbot talks to a consumer will make the consumer less willing to buy, because the consumer believes that the AI is less empathic than a real person. Some scholars also pointed out that virtual streamers only repeat instructions driven by artificial intelligence, lack flexibility in the real-time interactions present in more typical live-streaming, can not be trusted by consumers, and will reduce consumers’ interest in purchasing [9,18]. However, no studies have explored the characteristics and effects of AI-oriented live-streaming e-commerce service failure in depth. Hence, this research pioneers the integration of live-streaming e-commerce service failures within the AI domain, providing significant insights for a more discerning assessment of live-streaming e-commerce practices [19].

2.2. Stressor–Strain–Outcome Model

The SSO framework was proposed by Koeske and Koeske (1993) [20], delineating a sequential connection between stress-inducing factors (stressor), resultant negative emotions (strain), and subsequent adverse outcomes (outcome) [21]. The SSO framework has been employed in elucidating the overuse of social media [22], worker technostress [23], and e-education failure [24]. For instance, excessive social media usage correlates positively with social media fatigue, resulting in heightened anxiety and behaviors aimed at avoiding information [25,26]. The pressure from recommendation algorithms can induce feelings of overload and psychological resistance in individuals, leading to disruptive usage behaviors [27]. The rapid expansion of live-streaming e-commerce has spurred an increasing amount of research into its adverse social consequences [4]. The SSO framework has been utilized to explore how AI assistant service failures affect consumers negatively, encompassing consumer dissatisfaction, cognitive overload, and intentions to discontinue usage [28]. AI-oriented live-streaming e-commerce service failures result in consumers forfeiting money and time, and enduring emotional stress (stressors). This breaches the expectations that consumers hold when engaging in live broadcasts, leading to their disappointment and emotional distress (strain). These factors may be regarded as precursors that influence consumer discontinuance behaviors (outcome). Therefore, the SSO model delivers a more precise framework for analyzing the effects of AI-oriented live-streaming e-commerce service failure, particularly from the viewpoint of consumer stress.

2.3. Expectancy Disconfirmation Theory

According to expectancy disconfirmation theory (EDT), when a product or service fails to meet consumers’ initial expectations, it results in negative expectancy inconsistency [29], potentially leading to dissatisfaction [30], disappointment [31], and other negative emotions that persist and influence subsequent consumer behaviors [32]. The previous literature on service marketing mainly investigated the psychological comparison mechanisms of consumer expectations from the perspectives of processes, outcomes, costs, and emotional expectations [33]. Specifically, process expectation inconsistency entails challenges for consumers in navigating and evaluating transactional processes, stemming from issues such as irrelevant, untimely, inaccurate, and incomplete information [34]. Inconsistent outcome expectations, such as product design defects and delivery delays, often mean that a product or service does not meet the functional needs of consumers [35]. Inconsistent cost expectations are related to technical system failures that require consumers to expend excess time and cognitive effort such as common system delays in live-streaming [36]. Inconsistent emotional expectations mean that consumers do not experience positive emotional experiences such as pleasure and identity in the process of shop** [37]. The past literature on e-commerce service failure predominantly classifies service failures based on process, outcome, and cost dimensions. For instance, Yadav et al. (2023) [38] believed that the functional, information, and system failures of online services in virtual communities will lead to consumers’ brand conversion and public complaints, which are not conducive to promoting consumer participation. Lee and Pan (2023) [39] proposed that system function overload, information overload, technical uncertainty, and privacy disclosure risks in face recognition payment services led to users’ perceived pressure, subsequent resistance, and negative word-of-mouth. Nevertheless, these classifications mentioned above overlook the real-time social interactions of live-streaming and the emotional needs of consumers. Liu et al. (2023) [40] found that a lack of professionalism and attractiveness of anchors will lead to the total destruction of the value. Heo et al. (2020) [41] further proposed that the emotional value of live-streaming business mainly comes from the interactive ability and entertainment of streamers. From the perspective of aesthetic theory, Tong et al. (2022) [42] believe that the aesthetic design and visual appeal of the live-streaming business interface are conducive to improving consumers’ sense of enjoyment. Therefore, it is necessary to emphasize the important roles of streamers’ social interaction and the visual aesthetics of live-streaming space in consumers’ emotional expectations [43,44].

3. Hypotheses Development

3.1. AI-Oriented Live-Streaming E-Commerce Service Failure, Disappointment, and Emotional Exhaustion

AI-oriented live-streaming information failure pertains to virtual streamers providing information that fails to assist consumers in fulfilling transactional activities or the desired behaviors, including inaccuracies, incompleteness, irrelevance, or untimeliness [45]. For example, virtual streamers may pass on inaccurate information because they do not understand, causing consumers to misunderstand products or services. In addition, AI streamers may ignore the actual needs of consumers and instead focus on promoting a product or service, resulting in a poor user experience. Hua et al. (2023) [46] indicated that a lack of personalized explanatory content in virtual live-streaming may make consumers feel unattended to or disinterested, increasing the possibility of distraction and fatigue. If too much information is passed to consumers in a short period of time, exceeding their cognitive and processing capabilities, this results in information overload [47]. Song et al. (2022) [48] found that consumers are more accepting of human service agents than AI chatbots because consumers feel that the information quality is better. As a result, AI-oriented live-streaming information failure signifies that merchants are unable to fulfill consumers’ requirements for product or service information delivery, which goes against consumers’ process expectations and ultimately leads to consumer dissatisfaction. Thus:

H1. 

AI-oriented live-streaming information failure is positively associated with consumer disappointment.

Functional failures in AI-oriented live-streaming include inaccurate demand identification, unsmooth transaction processes, and imperfect after-sales service. The previous literature believed that the convenience and practicability of live-streaming e-commerce can promote consumers’ live-streaming purchase intentions [49]. The excessive concentration of live-streaming e-commerce product types makes it difficult to meet the diversified product demand preferences of consumers [50]. Yang et al. (2022) [51] found that problems such as delays or damages in logistics distribution would affect users’ shop** experience. Chen et al. (2022) [52] also suggested that poor product quality directly leads to a high return rate. Gao et al. (2024) [1] argued that AI streamers have difficulty meeting consumers’ individual needs due to absurd responses, vague purposes, or poor availability, so consumers generally have a more negative attitude toward AI than human streamers. A study by Zheng et al. (2023) [53] similarly pointed out that human streamers may be more effective at meeting the functional needs of consumers because they have a higher proficiency in providing personalized services. Therefore, AI-oriented live-streaming e-commerce may have potential problems such as insufficient personalized recommendations, complex transaction processes, and imperfect after-sales service, making it difficult for consumers to obtain functional value, which goes against their outcome expectations and leads to disappointment. Thus:

H2. 

AI-oriented live-streaming functional failure is positively associated with consumer disappointment.

AI-oriented live-streaming system failure refers to a variety of problems or adverse situations in a live broadcast system, resulting in the system not operating normally or achieving the expected effect; for example, network delays, stuttering, a lack of navigability, live broadcast interruptions, and private data leakages [54]. An unstable network connection will lead to blurred, intermittent, or even interrupted live broadcasts, affecting the user’s viewing experience [38]. Hua et al. (2023) [46] showed that unreasonable system designs or unfriendly user interfaces would lead to user operation difficulties, reduce user satisfaction, and even affect user retention. In addition, if the data in the broadcast room are improperly managed or there are security loopholes, it may lead to the loss or disclosure of user data, damaging user privacy and reducing user trust [55]. The AI-oriented live-streaming e-commerce usually involves complex artificial intelligence technologies such as speech recognition, natural language processing, and image recognition. These technologies have high requirements for the stability and performance of the system, and once the system fails or delays, the AI streamer may not work normally, thus affecting the broadcast [56]. Hence, AI-oriented live-streaming system failures increase consumers’ time cost and cognitive load, and violate their cost expectations. Thus:

H3. 

AI-oriented live-streaming system failure is positively associated with consumer disappointment.

AI-oriented live-streaming interaction failure is mainly manifested in the lack of attractiveness, professionalism and credibility of virtual anchors [57]. As a service contact person and brand endorsement, anchors need to have both product recommendation and social interaction capabilities. Therefore, interaction failure occurs when the virtual streamer has poor communication skills [58], cannot effectively explain the product [59], and lacks crisis handling ability [60]. Ma et al. (2022) [61] proposed that the lack of professionalism of streamers directly led to consumers’ hesitation in purchasing. In addition, it is difficult to provide an immersive virtual shop** experience if the streamer lacks enthusiasm and empathy and is unable to respond quickly to the audience’s needs and feedback [53]. AIs are often perceived as mechanical and aloof, and consumers feel nervous, unfriendly, and impersonal when interacting with them, which reduces users’ goodwill toward virtual streamers [62]. The study of Gao et al. (2023) [1] provides evidence that compared with human streamers, AI streamers lack perceived intimacy and responsiveness, thus weakening consumers’ purchase intention. Therefore, interaction failures in AI-oriented live-streaming violate consumers’ emotional expectations and further lead to disappointment. Thus:

H4. 

AI-oriented live-streaming interaction failure is positively associated with consumer disappointment.

AI-oriented live-streaming aesthetic failure means that the visual effects, interaction design, virtual streamer image design, or interface layout of the live-streaming environment are not attractive and may even cause consumers to be unhappy. Recent research on virtual influencer marketing emphasizes that sensory cues can create immersive and engaging experiences for consumers, especially in real-time interactive virtual live environments [63]. Visually appealing anthropomorphic designs (such as a human appearance or a female voice) are more likely to elicit social responses and behavioral changes in humans [64]. Cyr et al. (2007) [65] and Liu and Guo (2017) [66] suggested that the design aesthetics of e-commerce were shown to affect consumers’ perceived usefulness, perceived ease of use, and perceived entertainment, which in turn affects consumers’ loyalty to e-commerce. Therefore, if AI live-streaming layouts are chaotic, their operation is cumbersome, or their color collocations are not harmonious, etc., this makes it difficult for users to quickly find the required functions, damaging the user’s experience. In addition, if a virtual streamer’s image design is not attractive or personalized or does not match the target user group, it will not be able to attract users’ attention, reducing the retention rate of the broadcast room. Hence, aesthetic failures in AI-oriented live-streaming could reduce consumers’ emotional experience and lead to disappointment. Thus:

H5. 

AI-oriented live-streaming aesthetic failure is positively associated with consumer disappointment.

A further consequence of AI-oriented live-streaming e-commerce services failure is emotional exhaustion. Emotional exhaustion refers to a state of fatigue caused by an individual’s overuse of their mental and emotional resources. Emotional exhaustion is usually caused by negative service experiences, work pressure, and intimate relationship problems, and is accompanied by emotional fatigue, increased stress, and decreased emotional regulation ability [67]. The research of Fu et al. (2020) [68] shows that information, system feature, and social overloads lead to users’ social media exhaustion. Soto et al. (2014) [69] pointed out that poor service experience would lead to consumers’ emotional exhaustion, which would lead to anxiety and anger. Emotional exhaustion is consumers’ subjective perception of the failure of live-streaming e-commerce services. Consumers participating in live-streaming e-commerce may need to spend a long time in waiting, product selection, placing orders, and other processes [70]. The occurrence of service failure makes consumers feel that their time and energy have been wasted, and this irreparable loss will lead to consumers’ disappointment, which will further lead to their emotional fatigue and exhaustion. Thus:

H6. 

Consumer disappointment is positively associated with emotional exhaustion.

3.2. Emotional Exhaustion and Discontinuance Behavior

According to the SSO framework, the disappointment and dissatisfaction caused by live-streaming e-commerce service failure will stimulate consumers to take negative measures [28]. Previous studies have shown that the emotional exhaustion caused by social media overload will reduce users’ engagement, or even cause them to withdraw from social media [71]. Baklouti and Boukamcha (2023) [72] found that online banking services that do not meet consumer expectations would trigger consumer resistance. Lee and Pan (2023) [39] pointed out that functional failures and inadequate system responses put technical pressure on users and thus reduce the frequency of system use. It can be seen that failures of information, transaction function, system fluency, social interaction, and aesthetic design in virtual live-streaming violate the expectations of consumers, causing them to lose and fatigue (emotional exhaustion). At this time, emotional exhaustion makes consumers unable to focus on live interaction and product information evaluation, and they tend to consume via other e-commerce platforms or physical stores. Thus:

H7. 

Emotional exhaustion is positively associated with live-streaming discontinuance behavior.

3.3. The Moderating Effect of Live-Streaming Platform Type

Commercial livestreaming platforms aim at product promotion, and users’ participation motivation is mainly product purchasing [73]. Consumers expect accurate and reliable information, high-quality products, easy-to-use shop** interfaces, and timely and effective after-sales service support [74]. These are closely related to the information, functions, systems and other services of live-streaming businesses [75]. Therefore, the information, functions, and system failures of AI-oriented live-streaming commerce will disappoint commercial live-streaming platform consumers, because their expectations are not consistent with the actual service experience. On the other hand, social e-commerce platforms are good at social interaction and cultivating user relationships [76]. Consumers pay more attention to the interactivity level of anchors and the pleasant atmosphere of live broadcasting [77]. Therefore, when there are interactive and aesthetic failures such as anchor indifference and color mismatches on social live broadcasting platforms, users will feel neglected and unable to enjoy the expected emotional experience. Thus:

H8. 

The type of live-streaming platform used moderates the impact of AI-oriented live-streaming service failures on customer disappointment.

H8-a. 

For commercial live-streaming platforms, information, function, and system failures have a greater impact on customer disappointment.

H8-b. 

For social live-streaming platforms, interaction and aesthetic failures have a greater impact on customer disappointment.

Therefore, the conceptual model of this study is displayed in Figure 1.

4. Research Methodology

4.1. Sample and Data Collection

This study employs a purposive sampling approach for data collection in order to mitigate common method bias and ensure data robustness [78]. To ensure the representativeness of the research sample, we disseminated electronic questionnaires through Credamo (www.credamo.com) to live-streaming consumers who have experienced AI-oriented e-commerce service failures in active cities such as Guangzhou, Shenzhen, Hangzhou, Chengdu, Bei**g, and others within the live e-commerce industry. In addition, the authors collected questionnaires from participants in ten Chinese universities who had experienced AI livestreaming e-commerce service failures.

The live-streaming video materials used in this study were all recorded in real live-streaming e-commerce environments. The experimental live-streaming interface mainly included a virtual streamer who introduced the product information; the product display area; and the interactive text display area, while attempting to exclude the influence of the name of the live room and the ID as well as the number of viewers.

Three key filter questions were posed to confirm participant eligibility: (1) How frequently do you engage with AI-oriented e-commerce live-streaming? (2) Which brands are commonly endorsed by streamers in your viewing experience? (3) Have you encountered any failures in AI-oriented live-streaming e-commerce services? Next, following the procedure for starting research scenarios on e-commerce service failures [4], we first presented a short scenario: “Suppose you are a viewer of an AI-oriented live-stream. Imagine watching a virtual streamer explain a product and having a service experience that makes you feel unsatisfied.” Finally, the participants were asked to complete thirty-five items related to the study’s constructs on a seven-point Likert scale and five questions about their socio-demographic information. The survey was conducted from 4 October to 29 December 2023. From an initial pool of 690 questionnaires, 588 valid samples were obtained after removing 102 invalid samples due to reasons such as lack of live shop** experience, misinterpretation of questions, and excessive missing values. This resulted in an effective response rate of 85.22%. A statistical power analysis using G*Power 3.1 software confirmed that a sample size of 588 was sufficient for a one-tailed test with a power (1 − $\beta$) = 0.95, at a significance level of α = 0.05.

Table 1. Respondents’ profiles in empirical tests (N = 588).

Demographic Profile	Categories	Frequency	Percentage
Gender	Male	233	39.63%
	Female	355	60.37%
Age (years)	18–25	269	45.75%
	26–35	211	35.88%
	36–45	66	11.22%
	>45	42	7.14%
Monthly Income (CNY)	<5000	214	36.39%
	5000–10,000	165	28.06%
	10,001–20,000	103	17.52%
	20,001–30,000	77	13.10%
	>30,000	29	4.93%
Education	≤Middle school degree	28	4.76%
	Bachelor	396	67.35%
	≥Master	164	27.89%
Live-streaming Shop** Experience	<3 times	12	2.04%
	3–10 times	299	50.85%
	>10 times	277	47.11%

presents a summary of the sociodemographic characteristics of the eligible sample. The respondents included 39.63% males and 60.37% females, grouped into four age brackets: 18–25 (45.75%), 26–35 (35.88%), 36–45 (11.22%), and over 45 (7.14%). Self-reported income indicated that one-third of participants earned less than CNY 5000 (36.39%), followed by CNY 5000 to 10,000 (28.06%). More than half (67.35%) had a bachelor’s degree and 27.89% had a master’s degree or above. More than half (50.85%) had 3-10 live-streaming shop** experiences, and 47.11% had more than 10 live-streaming shop** experiences, indicating the high suitability of this survey sample. Demographic factors were used as control variables for data analysis. The results showed that the influences of gender ($\beta$= −0.002, p = 0.57), age ($\beta$= 0.061, p = 0.79), income ($\beta$ = 0.024, p = 0.83), education ($\beta$= −0.033, p = 0.98), and live-streaming shop** experience ($\beta$= 0.022, p = 0.84) on consumers’ discontinuous participation intention were not significant.

4.2. Questionnaire and Instruments

This study primarily employed established scales, adapting items to suit an AI-oriented live-streaming context. In order to reduce translation bias, the English questionnaire underwent translation into Chinese, followed by the back-translation of the Chinese version into English. In terms of the live-streaming platform types, e-commerce live-streaming platforms emphasize product promotion and purchase, while social live-streaming platforms emphasize social interaction, creativity, and entertainment. A total of 35 items were used to measure the eight latent constructs (see

Table 2. Scale items of constructs with their sources.

Variables	Items	Sources
Information Failure	AI streamer supplies me with misleading information.	[54,79]
	AI streamer supplies me inconsistent information.
	AI streamers supplies me irrelevant information.
	AI streamer supplies me untimely information.
	AI streamer supplies me incomprehensive information.
Functional Failure	AI-oriented live-streaming fails to accommodate my needs and preferences for specific content.	[54,79]
	AI-oriented live-streaming lacks the ability to help me compare content of interest.
	AI-oriented live-streaming fails to aid me in searching for product information.
	AI-oriented live-streaming is unable to help me place orders for desired products.
	AI-oriented live-streaming lacks the capability to offer a flawless post-purchase guarantee.
System Failure	Accessing the AI-oriented live-streaming I desire is proving to be challenging.	[54,79]
	I find myself constrained when utilizing the AI-oriented live-streaming e-commerce service.
	I need substantial effort to utilize the AI-oriented live-streaming e-commerce service.
	I need additional time to utilize the AI-oriented live-streaming e-commerce service.
	I feel unsafe while engaging in the AI-oriented live-streaming e-commerce service.
Interaction Failure	The AI streamer lacked enthusiasm in their interaction with me.	[80]
	The AI streamer did not offer enough opportunities for questions and answers.
	The AI streamer failed to grasp my needs and provide personalized attention.
	If I pose questions, the AI streamer cannot respond promptly.
	My questions often elicit responses from the AI streamer that are not closely related to my inquiry.
Aesthetic Failure	The background of live-streaming room and AI streamer does not match.	[81]
	The image quality of the AI-oriented live-streaming interface is fuzzy.
	AI streamer has poor personal image design.
	The AI-oriented live-streaming broadcast room is too cluttered.
	The arrangement of the AI-oriented live-streaming room is not visually appealing.
Disappointment	The AI-oriented live-streaming e-commerce failed to uphold the promise it made to me	[82]
	The AI-oriented live-streaming e-commerce disappointed me when I needed it the most.
	The AI-oriented live-streaming e-commerce failed to offer the support I required.
Emotional Exhaustion	I feel tired from AI-oriented live-streaming shop**.	[67,68]
	Engaging AI-oriented live-streaming is a strain for me.
	I feel burned out from AI-oriented live-streaming.
Discontinuance Behavior	I have stopped using AI-oriented live-streaming.	[68]
	I don’t plan to stay much longer in this AI-oriented live-streaming room.
	I might transition to another live-streaming platform that provides superior service.
	I often contemplate transitioning to another live-streaming platform for shop**.

): information failure (IF) [54,79], functional failure (FF) [54,79], system failure (SF) [54,79], interaction failure (IF) [80], aesthetic failure (AF) [81], disappointment (DP) [82], emotional exhaustion (EE) [67,68], and discontinuance behavior (DB) [68].

We utilized a 7-point Likert scale to measure the dimensions of all multi-item constructs, with ratings ranging from “1”, indicating strong disagreement, to “7”, representing strong agreement. Furthermore, following the pilot testing of the instruments involving 56 scholars, industry practitioners, and experienced live-streaming e-commerce consumers, the adequacy of measurements was ensured, with Cronbach’s alpha falling within the prescribed range of 0.733 to 0.967. We used the statistical software Smart PLS 4.0 for PLS-SEM.

4.3. Common Method Variance

In order to control the common method bias, a double-blind experimental design was adopted in this study, and confidentiality measures were adopted for subjects to reduce their bias in answering questions [83]. Additionally, we employed Harman’s single-factor analysis to verify the absence of common method bias in our data. The analysis revealed that a single item explains only 23.81% of the total variance, which is below the 50% baseline. Consequently, we inferred that our data are free from such bias.

4.4. Assessment of Structural Model

First, the structural model was assessed by examining the significance of determination (R²), predictive relevance (Q²), and path coefficients by using Smart PLS software. A total of 5000 bootstrap** samples were performed without any alterations shown [84]. The coefficients of determination (R²) of disappointment, emotional exhaustion, and live-streaming discontinuance behavior were 0.552, 0.522, and 0.561, respectively, which exceeded the threshold of 0.50, indicating that the model had significant explanatory power [84]. The blindfolding procedure resulted in Q² values of 0.365 for disappointment, 0.248 for emotional exhaustion, and 0.289 for live-streaming discontinuance behavior, all greater than zero, indicating a favorable predictive relevance for the dependent variables.

5. Data Analysis and Results

5.1. Assessment of Measurement Model

A comprehensive statistical summary of all items is clearly outlined in

Table 3. Construct reliability and validity assessment.

Constructs	SFL	CR	AVE	CA
Information Failure	0.782	0.912	0.674	0.903
	0.836
	0.804
	0.885
	0.796
Functional Failure	0.807	0.927	0.719	0.915
	0.834
	0.799
	0.876
	0.918
System Failure	0.863	0.951	0.794	0.928
	0.925
	0.911
	0.874
	0.882
Interaction Failure	0.839	0.932	0.734	0.918
	0.754
	0.843
	0.915
	0.923
Aesthetic Failure	0.867	0.931	0.729	0.917
	0.854
	0.883
	0.829
	0.834
Disappointment	0.864	0.886	0.722	0.852
	0.858
	0.826
Emotional Exhaustion	0.924	0.946	0.854	0.925
	0.915
	0.934
Discontinuance Behavior	0.873	0.920	0.742	0.913
	0.895
	0.836
	0.841

Notes: SFL = Standardized factor loading; CR = composite reliability; AVE = average variance extracted; CA = Cronbach’s alpha.

. The factor loadings for each item ranged from 0.754 to 0.934, surpassing the desirable threshold of 0.7 [84]. The composite reliability (CR) values for all constructs range between 0.886 and 0.951, while Cronbach’s alpha (CA) estimates fall within the range of 0.852 to 0.928, surpassing the suggested cutoff of 0.8, indicating sufficient internal consistency [85]. Furthermore, all constructs’ average variance extracted (AVE) values exceed the acceptable threshold of 0.6. Moreover, KMO and Bartlett’s test results were between 0.784 and 0.859, which were greater than the cut-off score of 0.5. Lastly, the normality test reported that the skewness and kurtosis coefficients fell below the threshold scores of ±3 and ≥10, respectively. The variables exhibit robust performance due to the satisfactory convergent validity findings.

Cross-loadings, the Fornell–Larcker criterion, and the Heterotrait–Monotrait Ratio (HTMT) were tested for discriminant validity between constructs [78]. Firstly, as shown in

Table 4. Descriptive statistics and evidence of discriminant validity.

Construct	M	SD	VIF	IFF	FF	SF	ITF	AF	DP	EE	DB
IFF	5.136	0.658	1.056	0.821	0.435	0.524	0.538	0.567	0.561	0.551	0.435
FF	5.078	0.472	1.375	0.453	0.848	0.639	0.566	0.438	0.513	0.545	0.548
SF	5.114	0.819	2.034	0.518	0.423	0.891	0.485	0.647	0.642	0.428	0.519
ITF	4.968	0.752	1.461	0.604	0.352	0.614	0.857	0.519	0.523	0.643	0.643
AF	5.036	0.864	1.523	0.539	0.568	0.539	0.536	0.854	0.556	0.552	0.572
DP	5.842	0.739	1.376	0.425	0.497	0.523	0.345	0.369	0.924	0.484	0.568
EE	5.631	0.657	1.238	0.533	0.599	0.438	0.617	0.634	0.574	0.849	0.567
DB	5.175	0.682	1.036	0.469	0.578	0.471	0.534	0.556	0.526	0.604	0.862

Notes: IFF = Information failure; FF = functional failure; SF = system failure; ITF = interaction failure; AF = aesthetic failure; DP = disappointment; EE = emotional exhaustion; DB = discontinuance behavior. The bold diagonal elements signify the square root of the AVE; the Heterotrait–Monotrait Ratio is represented by underlined font.

, the cross-loadings in this study did not surpass the outer loadings of the indicators, indicating the establishment of discriminant validity. Secondly, the corresponding correlation coefficients are smaller than the square root of the AVE for each variable, meeting the Fornell–Larcker criterion [85]. Additionally, all Heterotrait–Monotrait ratios (HTMT) fall below the threshold of 0.9 [86]. Therefore, the findings indicate that the variables demonstrate adequate discriminant validity. Table 4 reveals that the maximum Variance Inflation Factor (VIF) value stands at 2.034, well below the threshold of 3, implying no multicollinearity among the constructs [87].

The PLS-SEM outcomes yielded the subsequent fit statistics: Model fits Chi-square = 898.469, df = 423, p = 0.000, χ²/df = 2.12 (<3), RMSEA = 0.069 (<0.08), CFI = 0.928 (>0.9), NFI = 0.945 (>0.9), TLI = 0.929 (>0.9), IFI = 0.938 (>0.9). These findings indicate a strong alignment between the model and the data.

5.2. Hypotheses Testing

This study employs PLS-SEM for hypotheses testing and path modeling analysis, utilizing Smart PLS software version 4.0 [88,89]. The results of the path diagram estimation and analysis obtained from the bootstrap** method are depicted in Figure 2 and

Table 5. Outcome of the structural examination.

Hypotheses	$\mathit{\beta }$	f2	R2	Q2	p	Remarks
DP			0.552	0.365
H1:IFF→DP	0.562	0.196			***	Accepted
H2:FF→DP	0.438	0.165			**	Accepted
H3:SF→DP	0.641	0.292			***	Accepted
H4:ITF→DP	0.443	0.174			**	Accepted
H5:AF→DP	0.375	0.167			**	Accepted
EE			0.523	0.248
H6:DP→EE	0.339	0.154			**	Accepted
DB			0.561	0.289
H7:EE→DB	0.572	0.208			***	Accepted

Notes:β = Standardized path coefficients; f² = effect size of path; R² = coefficients of determination; Q² = Stone–Geisser’s Q²; *** p < 0.001, ** p < 0.01.

. Specifically, information failure ($\beta$_IFF→DP = 0.562, p <0.001), functional failure ($\beta$_FF→DP = 0.438, p <0.01), system failure ($\beta$_SF→DP = 0.641, p <0.001), interaction failure ($\beta$_ITF→DP = 0.443, p <0.01), and aesthetic failure ($\beta$_AF→DP = 0.375, p <0.01) all support H1–H5. Furthermore, our findings also indicate a positive influence of consumer disappointment on emotional exhaustion ($\beta$_DP→EE = 0.339, p <0.01), confirming H6. As for the dependent variables, consumer emotional exhaustion showed a positive impact on livestreaming discontinuance behavior ($\beta$_EE→DB = 0.572, p <0.001), supporting H7. Importantly, the effect sizes (f²) for hypotheses H1, H2, H3, H4, H5, H6, and H7 surpass the 0.02 threshold, signifying significant effects.

5.3. Moderating Effect Analysis

In order to distinguish the impact of AI-oriented live-streaming e-commerce service failures on consumer disappointment, we divided the total sample into two parts: social live-streaming platforms (N = 296) and commercial live-streaming platforms (N = 292). The multi-group analysis (MGA) function of Smart PLS was utilized in this study to compare the differences in path coefficients among multiple groups.

The data in

Table 6. Path coefficient comparison between social and commercial live-streaming platforms.

Path	Coefficient		Difference	T-Value
	Social	Commercial
IFF→DP	0.217 **	0.526 **	−0.309 **	−3.861
FF→DP	0.285 **	0.585 *	−0.3 **	−3.113
SF→DP	0.263 **	0.447 **	−0.184 *	−1.998
ITF→DP	0.426 **	0.237 **	0.189 *	2.086
AF→DP	0.384 **	0.126 **	0.258 **	3.061

Note: ** p < 0.01, * p < 0.05.

show the path coefficient and the significance level of the structural model constructed between two group samples. There is a difference between the two groups of samples. The path coefficients between interaction failure ($\beta$_ITF = 0.426, p < 0.01), aesthetic failure ($\beta$_AF = 0.384, p < 0.01), and emotional exhaustion are significantly larger in the social live-streaming platform sample than in the commercial live-streaming platform sample. In contrast, the path coefficients between information failure ($\beta$_IFF = 0.526, p < 0.01), functional failure ($\beta$_FF = 0.585, p < 0.05), system failure ($\beta$_SF = 0.447, p < 0.01), and disappointment are significantly larger in the commercial live-streaming platform sample than in the social live-streaming platform sample.

6. Discussion and Conclusions

6.1. Conclusions

This study explores the impact of AI-oriented live-streaming e-commerce service failures on consumer discontinuance behavior based on an SSO framework and EDT. The results showed that consumer-perceived information failures such as inaccurate, untimely, and incomplete information led to consumer disappointment and emotional exhaustion. These findings are consistent with previous research on mobile payments [39] and social media scenarios [75]. This study extends previous research on information delivery by addressing AI-oriented live e-commerce. Second, functional failure is positively correlated with consumer disappointment. Previous studies have shown that the lack of functional utility value in virtual communities can lead to negative word-of-mouth for brands by consumers [38]. Our study further explored the impact of functional failure on consumer emotion. This finding underscores the practical value of AI-oriented live-streaming. Moreover, our findings indicate a positive relationship between system failure and consumer disappointment, aligning with prior studies highlighting the significance of system fluency in the adoption of AI personal assistants [28] and mobile banking [79]. In addition, interaction failures have a positive impact on consumer disappointment. Virtual streamers lack interactive ability, professionalism, and credibility, which are inconsistent with consumers’ emotional expectations. Past research has come to similar conclusions that voice AI services that lack empathy and responsiveness reduce user satisfaction and loyalty [90]. The empirical results show that aesthetic failure positively affects consumer disappointment. Previous studies on interactive virtual reality have reached a similar conclusion, i.e., that the color, font, image, layout design, and product display of a virtual shop** platform will affect the aesthetics of the platform [57]. The aesthetic element of virtual environments is particularly critical, as good visual design can often inspire a sense of pleasure, excitement, joy, and fun [91].

It was found that consumer disappointment has a positive effect on emotional exhaustion, which further leads to discontinuation behavior. In essence, disappointment and emotional exhaustion mediate the relationship between AI-oriented e-commerce live-streaming service failure and customer discontinuance behavior. Previous studies also hold that the severity of the negative impact of service failure depends on consumers’ perceived betrayal and dissatisfaction [82]. In this study, disappointment and emotional exhaustion are regarded as stress perception, which further broadens the research on the mechanism of service failure in mobile commerce.

Finally, the test results of the moderating effect of live-streaming platform type show that users on commercial live-streaming platforms experience higher disappointment when they experience information failure, function failure, and system failure. This is because they pay more attention to the practical value of AI-oriented livestreaming business. However, interactions and aesthetic failures on social live-streaming platforms inspire higher levels of disappointment because the participation motivation of these consumers is entertainment-seeking [92].

6.2. Theoretical Implications

This study has several contributions to the study of live-streaming commerce, expectation disconfirmation theory, and service failure. To our knowledge, few studies have explored the structural dimensions and impact outcomes of service failures in live-streaming commerce from the perspective of consumer expectations. Unlike previous studies on live-streaming commerce, which mainly focus on positive values such as purchase, participation, and reward [93], this paper focuses on the dark side of AI-oriented live-streaming, thereby broadening the research boundaries of virtual live-streaming [94]. Further, this study applies the expectancy disconfirmation theory to the field of live-streaming, and proposes the concept of AI-oriented live-streaming e-commerce service failure, including information, functional, system, interaction, and aesthetic failures. These findings enrich the theoretical connotation of AI-oriented live commerce and lay the foundation for subsequent research on the negative status of live commerce.

Secondly, this study also explores the boundary role of live-streaming platform type. Different from previous studies that focused on negative events on a single or same live-streaming platform type, this study conducted an empirical study on the differences in the impact of cross-platform AI-oriented live-streaming e-commerce service failures on consumer behavior. It provides a new perspective for understanding the important role of livestreaming consumers’ motivation for participating in live-streaming marketing by revealing the potential reasons for the impacts of service failures of different live-streaming platforms on consumers’ discontinuation behavior, and complements and expands the existing research on live-streaming commerce [95]. Although previous studies have examined a variety of negative cases of live-streaming platforms, such as emergencies, value co-destruction, and product failures [4], they have not explained the potential differences in the impacts of negative events across different scenarios. Our study identifies these potential differences and provides insights for future research.

Finally, this paper explores the impact of AI-oriented live-streaming e-commerce service failures on consumer disengagement, further enriching the theoretical literature on consumer engagement in virtual live-streaming commerce [96]. Consumer discontinuation behavior, as an extremely bad situation of consumer engagement, is an obstacle to the development of AI-oriented live-streaming business. However, few papers have tried to reveal the factors that hinder the continuous engagement of consumers in virtual live-streaming business [97,98,99]^. The research question of this paper considers the current development trends in live-streaming business, and understands the relationship between virtual live-streaming services and user engagement from the perspective of service failure, providing a new perspective for subsequent research on digital avatars and avatars in the field of live-streaming.

6.3. Managerial Implications

AI-oriented live-streaming e-commerce service failures include information, functional, system, interaction, and aesthetic failures, which provides important ideas for live marketing enterprises to prevent service failure. Live-streaming businesses should use big data analysis tools to dig deep into user data, understand user needs and preferences, and then decide live-streaming content and promotion strategies to ensure the accuracy of information dissemination. Secondly, system testing and troubleshooting should be carried out in advance to ensure that the live-streaming platform and technical equipment used are stable and reliable in order to avoid problems in the live-streaming process. Virtual anchors should provide real product information, reliable payment security, enhance customer trust, reduce purchase risks, and promote the occurrence of transactions. Live-streaming platforms should adopt more advanced artificial intelligence and machine learning technologies to improve the speech recognition and natural language processing capabilities of virtual streamers, so that they can interact with the audience more naturally and smoothly. This improves professionalism and interactivity, effectively reducing the likelihood of interaction failure. Live-streaming merchants should choose colors that fit their brand style and product characteristics, maintaining a consistent tone and color scheme to enhance visual appeal and brand recognition. They should ensure a well-organized layout of elements in the broadcast room, including product displays and introductions, for a clear and user-friendly interface.

In addition, various types of live-streaming platforms should implement distinct strategies for preventing and recovering from service failures. Social live-streaming platform merchants should provide rich, diverse, interesting, and in-depth content that meets audience interests and needs, can trigger resonance and discussion, and enhances user participation. These could include providing clear, high-quality sound effects, including the voice of the anchor and background music, to enhance auditory enjoyment and atmosphere; providing a variety of social functions, such as gift tips, fan interaction, virtual gifts, etc., to increase the interaction and sociability between users; for commercial live-streaming platform merchants, optimizing the technical architecture and performance of the live-streaming platform to ensure smooth video playback, reduce latency and loading time, and improve the viewing experience; providing detailed product information and explanation, including introductions of functions, advantages, uses, etc., to help consumers fully understand the product and make purchase decisions; and based on consumers’ browsing history and interests, recommending livestreaming content of related products or similar functions to provide personalized shop** advice and guidance.

6.4. Limitations and Future Research

This study has several limitations, and future studies can try to expand the research depth from the following aspects. First of all, in addition to the situational questionnaire method adopted in the current study, future studies can explore more vivid and intuitive forms, such as using video simulations or real market data, to eliminate unknown interfering factors and improve the external validity of the study. This method may be closer to a real live broadcast scene, making the experimental results more realistic and credible. At the same time, more variables and factors, such as user personal characteristics and viewing environment, can also be considered to further refine the experimental design and improve the internal validity and explainability of the study. Secondly, the sample of the current study is mainly focused on live-streaming audiences in China, especially students; future studies should consider audience groups in different cultural contexts more broadly. To ensure that the findings are representative and generalized, the sample size can be expanded and a wider range of people can be included, such as audiences of different ages, occupations, and geographical backgrounds. Finally, this study does not discuss the interactions between AI-oriented live-streaming e-commerce service failures and consumer characteristics, the visual designs of live-streaming business, service recovery strategies, and other factors, which are areas to be further studied. Future research can further broaden the research boundaries of AI-oriented live-streaming e-commerce service failure.