AI‐Mediated Communication in E‐Commerce: Implications for Customer Trust

Hennighausen, Christine; Yarza Navarro‐Schär, Vanessa Gabriela; Eller, Eric Article — Published Version AI‐Mediated Communication in E‐Commerce: Implications for Customer Trust International Journal of Consumer Studies Provided in Cooperation with: John Wiley & Sons Suggested Citation: Hennighausen, Christine; Yarza Navarro‐Schär, Vanessa Gabriela; Eller, Eric (2025) : AI‐Mediated Communication in E‐Commerce: Implications for Customer Trust, International Journal of Consumer Studies, ISSN 1470-6431, Wiley, Hoboken, NJ, Vol. 49, Iss. 5, https://doi.org/10.1111/ijcs.70111 This Version is available at: https://hdl.handle.net/10419/330201 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ 1 of 17 International Journal of Consumer Studies, 2025; 49:e70111 https://doi.org/10.1111/ijcs.70111 International Journal of Consumer Studies ORIGINAL ARTICLE OPEN ACCESS AIMediated Communication in ECommerce: Implications for Customer Trust ChristineHennighausen1 | VanessaGabrielaYarzaNavarro-Schär2 | EricEller1 1Business School, Technische Hochschule Ingolstadt, Ingolstadt, Germany | 2Center for Leadership and People Management, LudwigMaximilianUniversität München, München,Germany Correspondence: Christine Hennighausen ([email protected]) Received: 28 May 2024 | Revised: 15 May 2025 | Accepted: 4 August 2025 Funding: The open access publication of this work was supported by the Open Access Publication Fund of Technische Hochschule Ingolstadt (THI). Keywords: AImediated communication (AIMC)| artificial intelligence| CRM| customer trust| Ecommerce| generative AI| service criticality ABSTRACT Generative artificial intelligence (AI) technologies offer new potential for marketing and customer operations, such as automation and personalization of customer service. However, more must be understood about how AImediated communication (AIMC) affects customer trust. We conducted an online experiment to investigate the impact of AIMC on customer trust in an online retail context. We presented N = 294 participants with two email scenarios describing a product return context, labeled as written by either (a) a service employee, (b) a service employee assisted by AI, or (c) AI on behalf of the service employee. We further varied levels of service criticality to consider customers' perception of vulnerability. Our findings revealed higher customer trust ratings in the online retailer when the email communications were written by the service employee, compared to those written by the service employee assisted by AI. When analyzing the different components of trust, it was found that communications written by the service employee assisted by AI reduced perceptions of both the online retailer's benevolence and integrity, while communications written by AI on behalf of the employee led to lower perceived integrity of the online retailer. Surprisingly, service criticality did not affect trust ratings. We discuss the managerial implications of integrating generative AI into customer service in the context of the EU AI Act, which came into force on 1 August 2024. 1 | Introduction With the release of ChatGPT3 in November 2022 (Crawford et al. 2023), a new era has begun. Artificial intelligence (AI) chatbots can now generate answers that are languagewise correct, compelling, and hard to distinguish from humangenerated content (Mei etal.2024). Due to its ease of use and potential to increase efficiency, ChatGPT has been widely adopted in both private (Thormudsson 2023a) and business contexts (Thormudsson2023b). In marketing and customer relationship management (CRM; i.e., interactions and processes facilitated by a business to manage and nurture relationships with customers through the integration of people, processes, and technology; Chen and Popovich2003), the enormous business potential of generative AI (i.e., AI systems that can autonomously generate content, such as text or images, using advanced machine learning models trained on large datasets; Kaplan and Haenlein2019) is reflected in exceptionally high adoption rates (Dencheva 2023; Thormudsson 2023b), for instance, for automatizing customer communication in social media and email messaging (Schweidel etal.2023). Under the EU AI Act, which came into force on 1 August 2024, European companies are obliged to label AIgenerated content and make its use transparent for their customers (European Commission2024). This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2025 The Author(s). International Journal of Consumer Studies published by John Wiley & Sons Ltd. Brief abstracts of these findings have been submitted for presentation at the 53rd Congress of the German Psychological Society (DGPs)/15th Conference of the Austrian Psychological Society (ÖGP) 2024, 1619 September and at the 22nd Congress of Work and Organizational Psychology (EAWOP) 2025, 2124 May. 2 of 17 International Journal of Consumer Studies, 2025 Such regulations reflect a broader international discourse on AI governance, highlighting the need to adapt legal frameworks, including data security and protection, as well as strengthen ethical foundations in response to the challenges posed by AI (Al Najdawi etal.2024, 2025). This raises the question of how customers react toward the explicit use of AI and AImediated communication (AIMC) in marketing and CRM and whether transparency about AI usage enhances or undermines customer trust (Bakonyi2024; Sigala etal.2024). Previous studies showed that customers are often skeptical when interacting with AI. For instance, customer purchase rates dropped by nearly 80% when the identity of an AI chatbot was revealed at the beginning of a sales call compared to when the identity of the AI chatbot was not disclosed (Luo etal.2019). A recent study by Liu etal.(2022) showed that individuals mistrust the sender of an email dealing with a product inquiry when the email has been written with the help of AI. Research on the “WordofMachine” effect suggests that customers' preferences for AI or human recommendations depend on whether the decision context is utilitarian or hedonic. Customers perceive AI as more competent than humans for utilitarian tasks but less so for hedonic tasks (Longoni and Cian2020). This highlights that AIdriven recommendations in utilitarian settings may foster trust, while they could erode trust in hedonic scenarios. Further evidence suggests that customers tend to mistrust AI, especially insituations of high service criticality, that is, when the outcome of a service interaction is significant to the customer (Chen etal. 2022; Xu, Shieh, etal.2020). High service criticality often leads to increased customer vulnerability, as the outcomes of service interactions can significantly impact customers' wellbeing or satisfaction (Dehghanpouri etal.2020). Mozafari etal.(2021) found that if service criticality is high, disclosing a chatbot's identity as such negatively affected customer trust in the conversational partner, while customer trust was not impaired when service criticality was low. This aligns with findings from Logg etal.(2019), who discovered that despite initial concerns about algorithmic decisions, individuals often exhibit algorithm appreciation, meaning they trust algorithmic judgments more than human ones in certain domains, especially when accuracy is emphasized. Since the widespread use of generative AI is quite a new phenomenon, however, research has just begun to investigate customers' reactions toward AI and AIMC in marketing and CRM. To make a significant and novel contribution to the literature on AIMC and customer trust, our study aims to delve deeper into the understanding of customers' perceptions of AIMC in an ecommerce situation, specifically focusing on customer trust, using a realistic CRM scenario. We, therefore, conducted an experimental study varying the sender of an email in a product return context from no AI agency to complete AI agency. We described the email communication with the customer as either written by (a) a service employee, (b) a service employee assisted by AI, or (c) an AI on behalf of the service employee. We further modified the service criticality of the product return scenario by describing either a highprice product urgently needed or a lowprice product not urgently demanded. As highlighted by recent AI research, managing AIdriven communications in a way that maintains consumer autonomy and trust is critical for avoiding negative reactions (Spais et al. 2023). By examining the trust components benevolence, integrity, and competence, we gain a deeper understanding of how AIMC affects customer trust and thus make both theoretical contributions (i.e., by extending the recent literature on how different degrees of AIMC impacts customer trust, specifically by disentangling trust in its components; Akbar etal.2024) and practical contributions (e.g., for developing marketing strategies to incorporate AIMC in CRM effectively). Our research thus provides a novel and significant contribution to the current literature on AIMC and customer trust, which is of high contemporary relevance, particularly in light of current regulations emphasizing transparency in the use of AI in marketing (European Commission2024). 2 | Theoretical Background With the new possibilities of generative AI creating texts, pictures, and audiovisual material, the question arises about how AIMC affects interpersonal communication and might change the way conversational partners perceive each other. In ecommerce, eliciting and protecting trust, i.e., “the willingness of a party to be vulnerable to the actions of another party based on the expectation that the other will perform a particular action important to the trustor, irrespective of the ability to monitor or control that other party” (Mayer etal.1995; 712), is a principal goal, as customer trust is essential for a company's success (Isaeva etal.2020). However, the rapid proliferation of generative AI along the customer journey (Dencik etal.2023) is faced with the recent observation that customers often mistrust AIgenerated communication (Prakash etal.2023). Customers may find themselves in a vulnerable position while interacting with AI due to their personal circumstances or the complexities of a particular situation (Mogaji etal.2020). The perception and use of AI could potentially exacerbate this vulnerability. As noted by Longoni and Cian(2020) AIMC is more readily trusted when customers view the context as utilitarian, where AI's competence shines, rather than hedonic, where customers may doubt AI's emotional intelligence and nuanced decisionmaking. While AI is merely a tool with no inherent harmful attributes (Tuffley2019), its usage in customer service is often segmented based on task complexity. Customers seem to prefer AI for less challenging issues but lean toward human assistance when tasks become more intricate (Xu, Chen, etal.2020). The preference is driven by customers' belief in the AI's capacity to solve problems (Prakash etal.2023). However, this can potentially amplify the customer's vulnerability if AI fails to solve complex issues, escalating the perceived risk (Xu, Chen, etal.2020). Furthermore, AI's growing role in marketing has raised concerns about its potential to undermine customer autonomy. The ability to automate decisions and recommendations without human oversight can sometimes lead to perceptions of control loss, further eroding trust in AI (Spais etal.2023). In healthcare contexts, for example, customers are often resistant to AI recommendations due to perceived uniqueness neglect—the belief that AI systems cannot account for their individual circumstances (Longoni etal.2019). These findings reinforce the importance of trust in AIMC in scenarios where the perception of personal uniqueness is crucial. AIMC can disappoint customers' expectations of companies and evoke fears of being deceived or manipulated (Jakesch etal.2019). In line with his, customers can perceive texts written with the aid of AI or by AI as less trustworthy 3 of 17 (Hancock etal.2020). Taking this a step further, AI's perceived control over the customer, or lack thereof, can also influence a customer's trust in AI agents (Yang etal.2022). The authors discovered that, while anthropomorphizing AI agents can enhance customers' acceptance in cases where the perceived control is high—it could conversely decrease their acceptance in low control situations—adding yet another source of vulnerability. Jakesch etal.(2019) investigated how using AI to create online profiles of Airbnb hosts influenced trust perceptions of these hosts. Their findings suggest that when profile descriptions of the Airbnb hosts were presented as either all humanwritten or all AIgenerated, participants gave almost identical trust ratings to the profiles matching the prerated trust level of the profiles. In a setting where participants suspected host profiles to be AIgenerated or labeled as such, however, participants rated these as less trustworthy. The authors refer to this observation as Replicant Effect and explain their findings with the Hyperpersonal Model of CMC (Walther and Whitty2021), proposing that overattributions based on the participants' suspicion that the profile was AIgenerated or by the label indicating that a profile was AIgenerated accounts for the lower trust ratings. Liu etal.(2022) explored AIMC in email communications, varying the degree of AIMC and the interpersonal emphasis of the email communication. They found that participants rated their conversational partner lowest on trustworthiness when AI completely mediated the email communication, while trust perceptions were higher when the conversational partner was aided by AI or used no AI in his email message. Interestingly, this effect occurred independent of the personal emphasis of the communication. Drawing on the described findings, we expect similar results for an ecommerce context and hypothesize: H1. Customer trust in an ecom retailer is higher when an online retailer's email is labeled as being written by a service employee than when the email is labeled as being written by a service employee assisted by AI or by AI on behalf of the service employee. Trust is a multidimensional construct comprising at least three elements: benevolence, integrity, and competence (McKnight and Chervany2001). Benevolence refers to the degree of positive orientation of a trustworthy party toward the welfare of the trust beyond selfish gain (Mayer etal.1995; McKnight and Chervany 2001). Integrity reveals itself in the perception that the trustworthy party adheres to recognized principles and values that align with the trustor's expectations (McKnight and Chervany 2001; Mozafari et al. 2021). Finally, competence relates to the ability of the trustworthy party to perform certain activities satisfactorily (Mozafari etal.2021; Rofiq and Mula2009). Recent research elucidates that AI systems often appear highly competent but less benevolent and less integral than their human counterparts, which poses an influence on trust formation (Li and Bitterly2024; Novozhilova etal.2024). In some contexts, AI systems are even perceived as more competent than humans, which means that people are more likely to follow the recommendations of AI than of other people (Spais etal.2023; Zhang etal.2023). The level of agency accorded to AI can both enhance and diminish trust (Vanneste and Puranam 2021). This prompts the need for a precise consideration of trust components when adopting AI systems in several societal settings (Novozhilova et al. 2024). Thus, acknowledging and understanding customer vulnerability in the blend of human and AI customer service interaction is crucial to ensure the successful management of consumer trust in the realm of AIMC and consequently in the success of ecommerce businesses (Ameen etal.2021). Mozafari etal.(2021) found that the loss of trust in chatbots after disclosure insituations of high service criticality was due to lower perceptions of the chatbot's competence and benevolence but not based on lower perceptions of the chatbot's integrity. Moreover, when a customer's service request could not be resolved, chatbot disclosure mitigated the adverse customer reactions so that customer trust was even increased (Mozafari etal.2021). In particular, customers reacted positively to chatbot disclosure in a failure situation due to firmer beliefs in the chatbot's integrity and benevolence, while competence beliefs were not affected. AIpowered services in shopping have been shown to improve the overall customer experience by increasing trust, convenience, personalization, and relationship engagement while reducing perceived sacrifices (Ameen etal.2021). However, this also depends on how competently the AI system performs its tasks—AI customer service has been found to be favored for tasks with low service criticality, while human customer service is favored for tasks with high service criticality, suggesting that perceived problemsolving ability plays an important role in mediating customers' usage intentions (Xu, Shieh, etal.2020). Heider's(1958) attribution theory posits that individuals make sense of their surroundings by attributing cause and effect to people's behaviors and events. In doing so, individuals are tempted to relate the behavior of others to internal factors such as personality traits or abilities (Heider1958). Recent research showed that conversational partners using AI in interpersonal relationships were not considered to put the same effort into a friendship as they would by personally writing the message, which, in turn, leads to decreased relationship satisfaction and uncertainty for the conversational partner (B. Liu etal.2023). Accordingly, we argue that in CRM contexts, customers may perceive online retailers as less benevolent and integral if a service employee, representing the retailer, partly or completely uses AI to engage with the customer, as the use of AI may imply diminished effort from the service employee in communication compared to manual drafting (Prentice and Nguyen2020). We thus hypothesize: H2. Customers' perceptions of benevolence and integrity are higher when an online retailer's email is labeled as being written by a service employee than when the email is labeled as being written by a service employee assisted by AI or by AI on behalf of the service employee. Since trust enables individuals to accept perceived vulnerability (Mayer etal.1995), customer trust seems to be especially relevant in customer situations of high service criticality, that is, when the outcome of a service situation particularly matters to customers (Crisafulli and Singh2017). Mozafari etal.(2021) showed that in an AIChatbot environment, service criticality moderated the trust–AI relationship across two separate experiments. Insituations of high service criticality, participants trusted less in a conversational partner when this partner disclosed herself as a chatbot at the end of the conversation compared to a nondisclosure scenario. However, chatbot disclosure did not affect customer ratings in a situation of low service criticality but in a situation of high service criticality. In contrast, when the chatbot could not solve the customer's service request, 4 of 17 International Journal of Consumer Studies, 2025 chatbot disclosure enhanced customer trust. It mitigated the effects of the negative service outcome as compared to when the chatbot was not disclosed. In line with these findings, we expect service criticality to moderate the relationship between AI condition and customer trust, especially insituations where service employee communication is partially or fully mediated by AI. More specifically, we suggest that insituations of high service criticality, the different amounts of AI in a communication will lead to larger differences in customer trust than those of low service criticality. We thus hypothesize: H3. Service criticality moderates the relationship between AIMC and customer trust in an online retailer, such that the difference in customer trust between AIMC conditions is more pronounced when service criticality is high compared to when service criticality is low. 3 | Materials & Methods 3.1 | Design and Procedure We conducted an online experiment to investigate how different degrees of AIMC affected customer trust in an online retailer and how much this relationship differs between high and low service criticality scenarios. The experiment followed a 3 (degree of AIMC: email communication written by service employee vs. written by service employee assisted by AI vs. written by AI on behalf of the service employee) ×2 (service criticality: high vs. low) betweensubjects design. To reduce demand effects and to conceal the purpose of our study, the consent form briefed participants that the study aimed to explore their perceptions and reactions to email communication from an online retailer. Participants were assigned randomly to an experimental condition and presented with an introductory text and a screenshot of an email exchange with an online retailer. They were instructed to put themselves in the customer's shoes, visualizing the outlined scenario. Next, participants were asked to evaluate their trust in the online retailer. Trust ratings served as dependent variables. Participants were then asked to indicate their perceptions of service criticality and realism of the scenario as a manipulation and validity check, respectively. On the following pages of the questionnaire, data on the participants' dispositional trust and attitude toward AI were collected as control variables, as well as demographic information (gender, age, highest level of education, current vocational status, field of study/profession). Before the end of the questionnaire, participants were asked to indicate how the email communication presented had been written as an attention check. The attention check additionally ensured that the respondents understood whether the email was written by a service employee, an AI, or both. Finally, participants were fully debriefed about the study's objectives and thanked for participating. 3.2 | Participants Participants were recruited via a university distribution list and various social media platforms. A total of N = 494 participants completed the study. However, N = 200 had to be excluded as they failed the attention check at the end of the questionnaire by not correctly identifying the experimental condition they were assigned to (see Section3.4 for the attention check item). The final sample comprised N = 294 Germanspeaking participants, of whom 60.20% identified themselves as female, 37.76% as male, and 2.04% as diverse. The average age of participants was M = 27.14 years (SD = 12.08). Most participants were highly educated, holding a university entrance certificate (63.27%), a bachelor's degree/prediploma (11.56%), a vocational training/ apprenticeship (9.52%), or a master's degree/diploma (7.48%). Most of them indicated that they were currently enrolled in a study program (61.22%), followed by employees (28.23%) and freelancers (3.06%). 3.3 | Materials For our experiment, we crafted six different email communications. To vary the degree of AIMC, we used an approach similar to Liu etal.(2022). Our manipulation of service criticality followed a common scenario technique (see below; Mozafari etal.2021; Ostrom and Iacobucci1995; Webster and Sundaram 1998). The designed stimulus material showed an email communication involving a defective product in the context of a product return scenario from an online retailer and was designed by the research team based on the study design by Liu etal.(2022). The product return scenario was selected for its relevance and realism, mirroring realworld customer experiences (see also Huang and Dootson2022). Given its prominence in online shopping, we chose to situate our scenario within the customer electronics sector (Lohmeier 2024; Statista Market Insights2024). The six designed email scenarios illustrated a service employee's response to a customer's return request. Each email acknowledged the customer's request, expressed apologies for the inconvenience due to the faulty item, and assured prompt refund arrangement by the customer service. The email concluded with a request for the customer to follow a provided link to complete the return process and submit personal data for the return process (see AppendixA for the stimulus material). To manipulate the degree of AIMC, after the regards of the service employee, a sentence in asterisks was included indicating the sender of the email. For the condition in which the service employee wrote the email (complete human agency condition), we added the sentence “This email was personally written by the service employee” for the condition in which the service employee used AI to write the communication (shared human/AI agency condition) we added the sentence “This email was written by the service employee using an intelligent autocomplete system (artificial intelligence).”, and for the condition in which the AI wrote the email on behalf of the service employee (complete AI agency condition), we included the sentence “This email was written by an advanced AI (artificial intelligence) system on behalf of the service employee.” To manipulate service criticality, participants were asked to read a short scenario before reading the email communication. For the high service criticality condition, participants were asked to imagine having recently purchased a new and costly laptop from an online retailer. Upon its first use, they discovered that the computer was defective in a particularly distressing 5 of 17 situation of urgent need for the laptop to complete a crucial student assignment or to prepare a vital job presentation. In the low service criticality condition, participants were asked to envision having recently acquired a lowpriced USB charging cable on sale, which they found defective upon first use. However, this situation was framed as less critical, given the nonurgent nature of the need for the cable, as they had other charging cables available for use. Except for the defective product type and the sentence indicating the degree of AIMC, the content and wording across the emails remained identical. To improve realism, each email was presented as a screenshot from Gmail, a widely used email client (Petrosyan2024; Rabe2024). As a name for the service employee, Alex Schneider was chosen as both the given name and the surname are among the most common names in Germany (Wikipedia2024; Rüdebusch n.d.). For a pretest of the stimulus material, see the Supporting Information. 3.4 | Measures Trust in the online retailer was measured with an established questionnaire assessing trust in ecommerce (McKnight etal.2002). The questionnaire measures trust on the components of benevolence (e.g., “I believe that the online retailer would act in my best interest.”; ɑ = 0.80), integrity (e.g., “The online retailer is truthful in its dealings with me.”; ɑ = 0.83), and competence (e.g., “The online retailer is competent and effective in providing electronic products.”; ɑ = 0.81). The internal consistency of the aggregate trust measure was ɑ = 0.90. In line with McKnight etal.(2002), trust was additionally assessed as a behavioral intention with items asking participants to indicate their willingness to engage in trustrelated behaviors, such as sharing sensitive personal information (Currall and Judge1995; McKnight etal.2002) with the online retailer in the context of this email communication. The three items we used were as follows: “I would click on the link in the email.”, “I would provide my contact details (name and address) on the linked website.”, and “I would provide my bank details (IBAN) on the linked website.” As previous work has produced mixed findings regarding the influence of general trust and attitude toward AI when investigating trust in AIMC and in AI (Chua etal.2023; Liu etal.2022), disposition to trust was additionally measured with the items of McKnight etal.(2002; e.g., “In general, people do care about the wellbeing of others.”, “In general, most folks keep their promises.”; ɑ = 0.89.). Attitude toward AI was assessed with the German version of the Attitude Towards Artificial Intelligence scale (ATAI scale; Sindermann et al. 2021). This short measure captures attitude toward AI with five items (e.g., “Artificial intelligence will benefit humankind.”, “Artificial intelligence will cause many job losses.”, ɑ = 0.81). As a manipulation and validity check, participants were asked to indicate perceived service criticality (“The fastest possible resolution of my service request is critical to me.”; Mozafari etal.2021), and realism (“This scenario is realistic.”) of the presented scenario, respectively. As an attention check, participants were asked to indicate how the presented email communication was written (“Who wrote the online retailer's email?”). The attention check further helped to ensure that respondents understood correctly whether the email had been written by a service employee, a service employee with the help of an AI or by an AI on behalf of a service employee depending on their respective experimental condition. Response options were identical to the wording used in the experimental manipulation (“This email was personally written by the service employee” vs. “This email was written by the service employee using an intelligent autocomplete system (artificial intelligence)” versus “This email was sent by an advanced AI (artificial intelligence) system on behalf of the service employee.”). All measures of the study were assessed using 7point Likerttype scales (1 = strongly disagree to 7 = strongly agree). 4 | Results 4.1 | Data Preparation The average percentage of missing data across all recorded variables was 1.80%. These missing data were replaced with median (continuous variables) and mode values (categorical variables). 4.2 | Statistical Analyses ANOVA and ANCOVA analyses were conducted with type III SS, as experimental conditions yielded unbalanced sample sizes (Hector etal.2010; Shaw and MitchellOlds1993; see Table1). To further examine significant main effects within ANCOVA, Tukey post hoc tests on the estimated marginal means (EMMs) of the different experimental conditions were calculated. Accordingly, for group comparisons, EMMs and SE are reported (H1–H3). Effect sizes are reported in line with Cohen(1998). 4.3 | Manipulation and Validity Checks Twoway ANOVAs were performed to evaluate the effects of AIMC condition, service criticality, and their interaction on perceived service criticality (manipulation check) and realism (validity check). For perceptions of service criticality, results yielded a significant main effect of service criticality, F(1, 288) = 6.53, p = 0.011, ηp2 = 0.02, while the main effect of AIMC condition and the interaction term were nonsignificant (ps ≥ 0.474). We thus further investigated perceptions of service criticality among the high and low service criticality conditions. An independent sample Welch's ttest revealed that in the high TABLE 1 | Sample sizes across experimental conditions. Service criticality AIMC Service employee Service employee assisted by AI AI on behalf of service employee High 25 55 74 Low 19 52 69 Note: Total N = 294. Abbreviations: AIMC, artificial intelligencemediated communication; AI, articifical intelligence. 6 of 17 International Journal of Consumer Studies, 2025 service criticality condition, participants reported higher perceptions of criticality regarding the resolution of their service request (M = 6.45, SD = 0.96) than in the low service criticality condition (M = 6.16, SD = 1.03, t (284.82) = 2.50, p = 0.014, d = 0.29 [0.06; 0.52]). For perceived realism, there were no significant main effects or interaction (ps ≥ 0.439), suggesting that participants' perception of realism did not differ across the six experimental conditions. A followup onesample ttest against the midpoint of the scale (i.e., 4) indicated that participants perceived the product return scenarios and email communications as highly realistic (M = 5.49, SD = 1.30, t (293) = 19.64, p < 0.001, d = 1.15 [1.00, 1.29]). 4.4 | Effects of AIMC on Customer Trust (H1) To test H1, we conducted a oneway ANCOVA to determine whether customer trust ratings significantly differed between the three AIMC conditions. Mean customer trust ratings served as the dependent variable, while age, gender, disposition to trust, and the attitude toward AI were entered as covariates. In line with H1, the main effect of AIMC condition proved significant, F(2, 286) = 4.65, p = 0.010, ηp2 = 0.03. Regarding the covariates, disposition to trust was significantly related to customer trust ratings, F(1, 286) = 11.08, p < 0.001, ηp2 = 0.04, while other covariates were nonsignificant (ps ≥ 0.526). Tukey post hoc tests revealed that the covariate adjusted mean of customer trust ratings was significantly greater in the complete human agency condition (EMM = 5.16, SE = 0.14) than in the shared human/ AI agency condition (EMM = 4.66, SE = 0.09, t (291) = 3.03, p = 0.007, d = 0.54 [0.185, 0.90]), indicating a moderate effect. Group differences between the complete human agency condition and the complete AI agency condition (EMM = 4.84, SE = 0.08) as well as between the shared human/AI agency and the complete AI agency condition failed to reach significance (ps ≥ 0.118). Figure 1 illustrates the EMMs for customer trust ratings across AIMC conditions along with the SE, 95% CI, and significant group differences. As an additional test of H1, we conducted the same statistical analyses as reported above with the behavioral intention trust measures as dependent variables across AIMC conditions. The main effect of AIMC condition was nonsignificant for all three behavioral intention trust measures (ps ≥ 0.167). In terms of the covariates, age and disposition to trust were significantly related to the participants' willingness to click on the link in the online retailer's email (age: F(1, 286) = 6.12, p = 0.014, ηp2 = 0.02; disposition to trust: F(1, 286) = 6.30, p = 0.013, ηp2 = 0.02), to share their name and address with the online retailer (age: F(1, 286) = 8.46, p = 0.004, ηp2 = 0.02; disposition to trust: F(1, 286) = 6.26, p = 0.013, ηp2 = 0.02), and to share information on their bank account with the online retailer (age: F(1, 286) = 4.68, p = 0.031, ηp2 = 0.02; disposition to trust: F(1, 286) = 10.07, p = 0.002, ηp2 = 0.03). Other covariates were not significantly related to the behavioral trust measure (ps ≥ 0.053). See Figure2 for the EMMs and SE with the 95% CI of the behavioral intention trust measures across AIMC conditions. Given these findings, H1 proposing that customer trust in an ecommerce retailer is higher when the ecommerce retailer's email is labeled as being sent by a service employee compared to when the email is labeled as being sent by a service employee assisted by AI or by AI on behalf of the service employee could be partly accepted (see discussion for the interpretation of nonsignificant finding of the difference between the complete human and the complete AI agency conditions and the behavioral intention trust measures, see Table2 for a summary of the study results). FIGURE 1 | Customer trust ratings across AIMC conditions (H1). The float values above each bar represent the trust score (EMM); the vertical black lines indicate the 95% confidence intervals of the SE. Lines with asterisks indicate significant differences between groups based on adjusted pvalues (Tukey post hoc tests; *p < 0.05; **p < 0.01; ***p < 0.001, twotailed). 7 of 17 4.5 | Effects of AIMC on the Trust Components Benevolence, Integrity, Competence (H2) To test H2, the same statistical analyses as described for H1 were conducted with perceptions of the trust components benevolence, integrity, and competence as dependent variables. Analyses showed a significant main effect of AIMC condition on perception of the online retailer's benevolence, F(2, 286) = 4.59, p = 0.011, ηp2 = 0.03, integrity, F(2, 286) = 4.02, p = 0.019, ηp2 = 0.03, and competence, F(2, 286) = 3.16, p = 0.044, ηp2 = 0.02. In terms of the covariates, disposition to trust yielded significant relationships with all trust components (benevolence: F(1, 286) = 9.53, p = 0.002, ηp2 = 0.03; integrity: F(1, 286) = 8.09, p = 0.005, ηp2 = 0.03; competence: F(1, 286) = 7.70, p = 0.006, ηp2 = 0.03). Other covariates were nonsignificant (ps ≥ 0.153). Tukey post hoc group comparisons on the covariate adjusted means revealed that in the complete human agency condition, participants perceived the online retailer as more benevolent and integral than in the shared human/AI agency condition (benevolence: EMMcomplete human agency = 5.15, SE = 0.18 vs. EMMsha red hum an/A I agenc y = 4.49, SE = 0.12, t (291) = 3.03, p = 0.007, d = 0.54 [0.19, 0.90]; integrity: EMMcomplete human agency = 5.50, SE = 0.15; EMMshared human/AI agency = 5.01, SE = 0.10, t (291) = 2.78, p = 0.015, d = 0.50 [0.14, 0.86]), suggesting moderate effect sizes. Further, participants rated the online retailer higher on integrity in the complete human agency condition as compared to the complete AI agency condition (EMM = 5.09, SE = 0.08, t (291) = 2.42, p = 0.041, d = 0.42 [0.08, 0.76]), indicating a small effect size, while differences between the complete human agency and the complete AI agency condition were nonsignificant for benevolence ratings (p = 0.088). Comparisons of perceptions of benevolence and integrity between the complete AI agency and the shared human/AI agency condition proved to be nonsignificant (ps ≥ 0.337). Although a significant main effect of AIMC was observed for competence perceptions, Tukey post hoc comparisons did not reveal significant differences between the conditions (ps ≥ 0.084). Figure3 shows the EMMs for benevolence, integrity, and competence across the AIMC conditions, with the SE, 95% CI, and significant group comparisons. Given these findings, H2 stating that customers' perception of benevolence and integrity is higher when the ecommerce retailer's email is labeled as being sent by a service employee compared to when the email is labeled as being sent by a service employee assisted by AI or by AI on behalf of the service employee could be largely accepted (see Table2 for a summary of the study results). 4.6 | Moderating Effect of Service Criticality on the Relationship Between AIMC and Customer Trust (H3) To test H3, proposing that service criticality moderates the relationship between AIMC and customer trust such that insituations of high service criticality, the difference in trust ratings is more pronounced between the AIMC conditions compared to situations of low service criticality, we conducted ANCOVAs with the two betweensubjects factors of AIMC condition and service criticality including their interaction term. Mean customer trust ratings, behavioral intention trust measures, and trust components served as dependent variables; age, gender, disposition to trust, and attitude toward AI were entered as covariates. ANCOVAs revealed no significant interaction between AIMC condition and service criticality for any of the trust FIGURE 2 | Behavioral intention trust measures across AIMC conditions (H1). The float values above each bar represent the trust score (EMM); the vertical black lines indicate the 95% confidence intervals of the SE. 8 of 17 International Journal of Consumer Studies, 2025 measures (i.e., customer trust ratings, behavioral intention trust measures, trust components; ps ≥ 0.559). Similarly, the main effects of service criticality were nonsignificant for any of the trust measures (ps ≥ 0.090). The main effects of AIMC and the relationships between the covariates and the trust measures were consistent with previously reported results (see Sections4.4 and 4.5). Thus, H3, proposing that service criticality moderates the relationship between AIMC and trust, was rejected, suggesting that the perceived criticality of service situations did not influence the effects of AIMC on trust as hypothesized (see Table2 for a summary of the study results). 5 | Discussion 5.1 | Theoretical Contributions With the widespread accessibility of generative AI, such as ChatGPT, companies have started implementing this new technology in their workflows to realize cost savings and increase efficiency. Specifically, for marketing and CRM, generative AI appears to be promising (Chui et al. 2023). Applying generative AI to facilitate CRM, however, raises the question of how customers react to it and how AI and customer communication mediated by AI influence customer trust – a prerequisite for a company's business success (Isaeva etal.2020). With the EU AI Act, which came into force on 1 August 2024, companies in Europe are obliged to make the use of AI transparent to customers (European Commission2024) so that understanding customers' reactions toward the use of AI labeled as this is essential. To address this research gap, the present research aimed to unveil how different degrees of AIMC, covering the spectrum of complete human agency to complete AI agency, affect customer trust in an online retailer, thereby extending previous work on AIMC and trust in personal relationships (e.g., Hohenstein and Jung2020; Liu etal.2022, 2023) and online selfpresentation (Jakesch etal.2019). To gain a deeper understanding of the dynamics between AIMC and customer trust, this study disentangled customer trust in its components integrity, benevolence, and competence and explored how perceptions of these trust components are specifically affected when an online retailer uses AIMC in customer communication. By examining the trust components, this research adds to the literature on trust and AIMC, as most studies have focused on the exploration of AIMC and trust on an aggregate trust level (e.g., Dehghanpouri etal.2020; Hohenstein and Jung2020; Jakesch etal.2019; Liu etal.2022, 2023; Longoni and Cian2020). An additional goal of this research was to explore the role of service criticality in the relationship between different degrees of AIMC and customer trust, expanding former research which examined how humanchatbot interactions (i.e., complete AI agency) impact customer trust in service frontline settings (Mozafari etal.2021). On an aggregate trust level, the results of our research indicate that individuals are more likely to trust an online retailer when the CRM communication is labeled as written by TABLE 2 | Summary of the study results. Hypothesis Result Key finding H1: Customer trust in an ecom retailer is higher when an online retailer's email is labeled as being written by a service employee than when the email is labeled as being written by a service employee assisted by AI or by AI on behalf of the service employee. Partially supported • Participants reported higher trust in the online retailer when the email was labeled as written by a service employee compared to when it was labeled as written by a service employee assisted by AI (moderate effect). • Trust did not significantly differ between emails labeled as written by a service employee and those labeled as written by AI on behalf of the service employee. H2: Customers' perceptions of benevolence and integrity are higher when an online retailer's email is labeled as being written by a service employee than when the email is labeled as being written by a service employee assisted by AI or by AI on behalf of the service employee. Largely supported • Participants reported higher perceptions of the online retailer's benevolence and integrity when the email was labeled as written by a service employee compared to when it was labeled as written by a service employee assisted by AI (moderate effect). • Participants reported higher perceptions of the online retailer's integrity when the email was labeled as written by a service employee rather than by AI on behalf of the service employee (small effect). • Benevolence perceptions of the online retailer did not significantly differ between emails labeled as written by a service employee and those labeled as written by AI on behalf of the service employee. H3: Service criticality moderates the relationship between AIMC and customer trust in an online retailer, such that the difference in customer trust between AIMC conditions is more pronounced when service criticality is high compared to when service criticality is low. Not supported • The level of service criticality did not significantly influence customer trust levels for different AIMC degrees. 15 of 17 Thormudsson, B. 2023b. Rate of Generative AI Adoption in the Workplace in the United States 2023, by Industry. Statista. https:// www. stati sta. com/ st ati stics/ 1 3 612 51/ gener at i v e - a iadopt ionr at eatworkbyi ndus tryus/ . Tuffley, D. 2019. “Human Intelligence + Artificial Intelligence = Human Potential.” Griffith Journal of Law & Human Dignity 2019: 170– 189. https:// doi. org/ 10. 69970/ gjlhd. v6i3. 1043. Vanneste, B. S., and P. Puranam. 2021. “Artificial Intelligence, Trust, and Perceptions of Agency.” INSEAD Working Paper. 2023. 1–46. https:// doi. org/ 10. 5465/ amr. 2022. 0041. Walther, J. B., and M. T. Whitty. 2021. “Language, Psychology, and New Media: The Hyperpersonal Model of Mediated Communication at TwentyFive Years.” Journal of Language and Social Psychology 40, no. 1: 120–135. https:// doi. org/ 10. 1177/ 02619 27X20 967703. Webster, C., and D. S. Sundaram. 1998. “Service Consumption Criticality in Failure Recovery.” Journal of Business Research 41, no. 2: 153–159. https:// doi. org/ 10. 1016/ S0148 - 2963(97) 00004 - 0. Wikipedia. 2024. “Liste Der häufigsten Familiennamen in Deutschland [List of the Most Common Surnames in Germany].” https:// de. wikip ed ia . org / w ik i/ List e_ der_ hä uf i gs t en_ Fam il ienn a men _ in _ D eut s chland. Xu, Y., Z. Chen, M. Y.- P. Peng, and M. K. Anser. 2020. “Enhancing Consumer Online Purchase Intention Through Gamification in China: Perspective of Cognitive Evaluation Theory.” Frontiers in Psychology 11: 581200. https://doi.org/10.3389/fpsyg.2020.581200. Xu, Y., C.- H. Shieh, P. Van Esch, and I.- L. Ling. 2020. “AI Customer Service: Task Complexity, ProblemSolving Ability, and Usage Intention.” Australasian Marketing Journal 28, no. 4: 189–199. https:// doi. org/ 10. 1016/j. ausmj. 2020. 03. 005. Yang, Y., Y. Liu, X. Lv, J. Ai, and Y. Li. 2022. “Anthropomorphism and Customers' Willingness to Use Artificial Intelligence Service Agents.” Journal of Hospitality Marketing & Management 31, no. 1: 1–23. https:// doi. org/ 10. 1080/ 19368 623. 2021. 1926037. Zanzotto, F. M. 2019. “Viewpoint: HumanInTheLoop Artificial Intelligence.” Journal of Artificial Intelligence Research 64: 243–252. https:// doi. org/ 10. 1613/ jair.1. 11345 . Zhang, G., L. Chong, K. Kotovsky, and J. Cagan. 2023. “Trust in an AI Versus a Human Teammate: The Effects of Teammate Identity and Performance on HumanAI Cooperation.” Computers in Human Behavior 139: 107536. https://doi.org/10.1016/j.chb.2022.107536. Supporting Information Additional supporting information can be found online in the Supporting Information section. Data S1: Supporting Information. 16 of 17 International Journal of Consumer Studies, 2025 Appendix A Scenario Description and Stimulus Material Contains the scenario descriptions and stimulus materials used in the study. The scenarios were designed to manipulate the degree of AI agency in CRM communication and the level of service criticality in a product return context. Participants were presented with a screenshot of an email response from an online retailer and asked to imagine themselves as the customer. The email varied in AI agency (complete human agency, shared human/AI agency, or complete AI agency) and service criticality (high vs. low). The following section provides the exact wording of the stimulus materials for two selected scenarios: (a) a highcriticality scenario with complete AI agency (FigureA1) and (b) a lowcriticality scenario with complete human agency (FigureA2). The remaining scenarios follow the same structure and can be inferred from the method section in our manuscript. High Criticality Scenario–Complete AI Agency Condition The following shows you an email communication from an online retailer to a customer. We are interested in how you perceive this email. Please imagine that you have received this email yourself from an online retailer. Background You recently purchased a new and very expensive laptop from an online retailer. Upon first use, you discover that the laptop does not work. However, you urgently need the device to create an important presentation for your job or your studies. Subsequently, you contacted the online retailer's customer service via email. This email was written by an advanced AI (Artificial Intelligence) system on behalf of the service employee. You receive the following reply: FIGURE A1 | Screenshot of the high criticality scenario–complete AI agency condition. FIGURE A2 | Screenshot of the low criticality scenario–complete human agency condition. 17 of 17 Low Criticality Scenario–Complete Human Agency Condition The following shows you an email communication from an online retailer to a customer. We are interested in how you perceive this email. Please imagine that you have received this email yourself from an online retailer. Background You recently purchased a USB charging cable from an online retailer. The cable was not expensive, as it was on sale. Upon first use, you discover that the cable does not work. It is not urgent since you have other charging cables at home, but it is still annoying. Subsequently, you contacted the online retailer's customer service by email. This email was personally written by the service employee. You receive the following reply: