Latest / 1 More Paper / Understanding Generative AI Usage Behavior in the Workplace: Focusing on the Moderating Role of Subscription to Premium Version — Kim & Kim (2026)
Transcript
- You’re listening to “Understanding Generative AI Usage Behavior in the Workplace: Focusing on the Moderating Role of Subscription to Premium Version,” by B. Kim and S. Kim. Published in 2026. Abstract. As generative AI becomes more integrated into organizational workflow, comprehending the mechanisms that promote its continued utilization by employees has become essential for management. This study investigates the factors influencing organizational members’ usage behaviors of generative AI in the workplace, incorporating perceived security and perceived risk into the expectation– confirmation model. Moreover, this study identifies the moderating effects of premium subscription on employees’ continuance decision-making. The theoretical framework was evaluated using data from 193 organizational members. The results reveal that perceived security exhibits no significant direct effects on satisfaction or continuance intention. Perceived risk negatively impacts generative AI use, user satisfaction, perceived usefulness, and perceived security, yet does not directly influence continuance intention. Premium subscription significantly enhances the relationship between continuance intention and generative AI use, while mitigating the adverse effect of perceived risk on generative AI use. INTRODUCTION. Generative artificial intelligence (AI) has rapidly emerged as a foundational technology across various domains, significantly transforming both individual daily experiences and the work practices of organizational members. Generative AI, by facilitating advanced capabilities such as content generation, data analysis, programming support, and decision-making support, is progressively recognized as a strategic asset for enhancing organizational productivity and operational efficiency, with recent industry reports underscoring its transformative potential. In line with these expectations, organizations globally are increasingly adopting generative AI technologies to secure a competitive advantage. A survey conducted by global consulting group PwC among CEOs worldwide indicated that 56% of CEOs surveyed have experienced an enhancement in operational efficiency through the adoption of generative AI. In South Korea, large corporations such as Samsung, LG, and POSCO have progressed from merely utilizing publicly available generative AI tools to developing proprietary generative AI customized for their specific business needs and security requirements. However, despite its rapid diffusion, the mere implementation of generative AI does not guarantee that organizations will attain their expected benefits. Nonetheless, the sustained and effective utilization of generative AI by organizational members is essential for converting technological potential into actual performance gains. User behavior regarding innovative technologies in the information systems (IS) literature has traditionally been elucidated through adoption-oriented frameworks, such as the technology acceptance model (TAM) and the unified theory of acceptance and use of technology (UTAUT). Although these models effectively explained initial intention to adopt technologies, they were relatively constrained in accounting for users’ post-adoption behaviors and long-term usage decisions. As organizational members acclimate to a technology, the factors influencing their sustained use of AI may significantly diverge from those prompting its initial adoption. To address this limitation, the expectation-confirmation model (ECM) was extensively utilized to explain users’ post-adoption behaviors by emphasizing the influence of usage experiences on satisfaction and the intention for continued usage. According to the ECM, users establish initial expectations before adoption, assess IS performance during actual use, and derive satisfaction to the extent to which their expectations are validated, subsequently influencing their intention to persist in using the IS. The ECM has been thoroughly validated in diverse IS contexts as a robust theoretical framework for explaining continuous usage behavior. Prior studies have effectively utilized ECM in technologies, such as mobile services, e-commerce platforms, wearable devices, and AI-based services, demonstrating its explanatory power in post-adoption contexts. Accordingly, this study utilized the ECM as the primary theoretical framework to analyze the mechanisms underlying organizational members’ sustained intention to use generative AI. The ECM offers an appropriate lens for comprehending how organizational members evaluate generative AI based on post-adoption beliefs and satisfaction after gaining practical usage experience in their work settings. The implementation of generative AI within organizations presents considerable issues regarding data security and privacy. Due to generative AI systems’ reliance on extensive user and organizational data from learning and output generation, employees may be concerned about the potential leakage or misuse of sensitive personal and corporate information. Yoon et al. (2025) noted that although generative AI enhances work efficiency and creativity, it also poses a double-edged sword effect due to issues related to data reliability, the hallucination phenomena, and security concerns. Accordingly, this study integrated perceived security and perceived risk as a fundamental post-adoption belief within the ECM framework. In organizational contexts, where employees frequently handle confidential information, perceptions of insufficient security may serve as a psychological barrier that deters ongoing usage, despite the technology being regarded as beneficial. Most existing studies on the ECM assumed a homogeneous user population, treating all generative AI users as equivalent regardless of whether they access the technology through a free or a paid premium plan. This assumption is increasingly untenable: the contemporary generative AI market is structurally bifurcated with free versions (e.g., ChatGPT Free and Gemini Basic) and premium versions (e.g., ChatGPT Plus and Claude Pro) offering substantively different functional capabilities, security architectures, and data governance frameworks. In organizational workplaces, where employees routinely handle sensitive data and face performance accountability, this service-tier distinction has direct and consequential implications for post-adoption behavior that existing research has yet to examine. Premium subscribers gain access to advanced reasoning models, larger context windows, priority system availability, and enterprise-grade data governance protocols unavailable to free users. These structural differences are expected to alter three critical post-adoption mechanisms that this study empirically tests: the strength with which continuance intention translates into actual usage behavior reinforced by sunk-cost psychology and reduced access friction unique to paid subscribers; the buffering of perceived risk against actual usage because premium users associate paid access with institutional safeguards and formal data governance; and the boundary conditions under which perceived usefulness drives continuance intention, shaped by the broader value perception embedded in premium-specific features. By rigorously comparing these mechanisms across subscription tiers in a workplace context, this study introduces service-tier differentiation as a novel, practically consequential moderating lens that prior ECM and AI continuance research has left unexplored. Overall, this study makes three interrelated contributions that distinguish it from prior ECM and AI continuance research. First, it advances theoretical understanding of post-adoption behavior in organizational generative AI contexts by incorporating perceived security and perceived risk as key post-adoption beliefs within the ECM. Second, it introduces subscription tier (free vs. premium) as a theoretically grounded and empirically testable moderating variable in the ECM framework. Prior studies on ECM treated users as undifferentiated, yet in practice, whether an employee used a free or premium version fundamentally shaped the resources, capabilities, and risk-governance mechanisms available during post-adoption use. To our knowledge, this study offers some of the empirical evidence on how service-tier heterogeneity conditions the key intention-behavior and risk-behavior pathways in organizational AI use. Based on these objectives, this study addresses the following research questions: What post-adoption factors influence organizational members’ usage toward generative AI? How does subscription to the premium version moderate the relationships between post-adoption beliefs, continuous usage intention, and actual usage of generative AI? From a practical perspective, this study’s findings are expected to provide in-depth insights for organizations seeking to promote sustained and effective utilization of generative AI. Specifically, the results would guide managerial strategies concerning security governance and organizational AI investment policies, ultimately assisting organizations in optimizing the value of generative AI while mitigating potential risks. THEORETICAL BACKGROUND Generative AI Generative AI is an AI technology that assimilates extensive data through natural language processing to generate new content. It predominantly employs deep learning models, particularly generative adversarial networks, transformer-based models (e.g., GPT and BERT), and probabilistic generative models, to learn patterns and structures from existing data and subsequently generate various forms of content, including text, images, audio, and video. Unlike traditional AI models that emphasize analysis and prediction tasks, such as classification, regression, and anomaly detection after learning patterns from data, generative AI is characterized by its ability to innovatively generate new data rather than simply learning from existing data. This ability of generative AI to execute creative tasks enables its application across diverse industries and fields. For example, it is used in content-creation-related fields to generate content, such as novels, news articles, and emails. Moreover, it is used to develop chatbots and virtual assistants capable of learning languages and engaging in natural conversations with users. In professional fields, such as medicine, law, and education, it can facilitate data analysis and automated report writing, in addition to diagnosis and case law analysis. However, it is crucial to acknowledge that while learning from vast amounts of data gathered from the internet, generative AI may incorporate information obtained without individual consent. For example, content posted on blogs and social media may be used as learning data without users’ permission. Therefore, to advance generative AI, it is imperative to address issues, such as user privacy infringement, data bias, and With the maturation of technology after the public launch of ChatGPT in 2022, the industry shifted to a premium business model. Free versions offer public access to basic tools, whereas premium versions, such as ChatGPT Plus and Claude Pro, have been introduced to address professional and enterprise requirements. Generally, premium versions grant access to more advanced parameters, enhancing reasoning abilities and improving factual accuracy. Moreover, premium versions enable users to upload and analyze massive documents and databases. The transition to a premium version for organizational members is frequently motivated by the necessity for enhanced data security and reliability. Enterprise-grade subscriptions provide administrative controls and encryption protocols that mitigate the risks of proprietary data leakage, a primary concern for professionals managing sensitive information. Since 2023, studies on generative AI have been actively conducted. For example, Baek and Kim (2023) investigated users’ motivations for using AI chatbots by integrating uses and gratifications theory and creepiness. They discovered that task efficiency and personalization positively affect trust and continuance intention, whereas creepiness negatively impacts these factors. Wang et al. (2024) explored the mechanism of students’ intention to utilize generative AI by integrating AI literacy into the theory of planned behavior (TPB). They demonstrated that AI literacy, which measures the comprehension, utilization, and evaluation of generative AI technology, helps shape students’ positive attitudes toward generative AI and ultimately enhances their intention to use it. Marimon et al. (2025) analyzed the impact of generative AI services on employee performance and the role of trust in these services and found that their use increases engagement via the mediation of trust in the service, ultimately enhancing performance. Kim and Kim (2026) examined the primary factors affecting organization members’ use of generative AI. They discovered that organization members’ trust in generative AI and their perceived risks significantly influence their utilization of the technology. The ECM The ECM primarily elucidates how users form intentions to continue utilizing IS after initial adoption. The ECM theory contends that users’ post-adoption decisions are influenced by a comparison between their initial expectations and the system’s actual performance experienced through use. When the actual performance of the IS meets or surpasses initial expectations, users experience confirmation, resulting in increased satisfaction and reinforcing their intention to continue using the system. In the ECM framework, post-adoption beliefs are pivotal in influencing user satisfaction and continuous usage intention. As competition has escalated in the AI service market and long-term user retention gains significance, a growing body of research has applied ECM to analyze continuous usage behavior in AI-based services. Ashfaq et al. (2020) examined continuance intention toward AI-based chatbot services by extending the ECM to include interaction needs. Lee et al. (2021) applied the ECM in the context of AI-based voice services, identifying perceived value, hedonic motivation, compatibility, and perceived security as significant post-adoption beliefs. Bhatnagar and Rajesh (2024) demonstrated that AI-specific characteristics in the digital banking domain enhance perceived usefulness and satisfaction, thereby promoting continuous usage intention. This study extends existing research by utilizing the ECM to examine organizational members’ continuous usage intention toward generative AI, emphasizing perceived security as a key post-adoption belief. Unlike traditional IS, generative AI processes and generates content using extensive amounts of potentially sensitive organizational data, thereby rendering security concerns particularly prominent in workplace environments. As employees acquire experience with generative AI, their perceptions of the system’s security in managing confidential information, protecting against data leakage, and preventing unauthorized access will significantly influence their satisfaction and continued use. Therefore, integrating perceived security and perceived risk into the ECM offers a theoretically grounded and contextually appropriate approach for comprehending continuous usage intention of generative AI in organizations. RESEARCH MODEL AND HYPOTHESES Based on the above discussion, this study integrated perceived security and perceived risk into the ECM to explain organizational members’ use of generative AI at work. We explored the moderating effect of a premium subscription on this continuous usage mechanism. The proposed research model is presented in Figure 1. Note. AI = artificial intelligence. Continuance Intention toward Generative AI The theory of planned behavior asserts that behavioral intention is the most proximal and significant determinant of an individual’s actual behavior. This theoretical framework posits that a heightened level of intention substantially increases the likelihood that an action will be executed. An intention-behavior gap frequently exists; users may possess a strong inclination to utilize IS yet fail to do so because of situational constraints, such as inadequate specialized knowledge or insufficient organizational resources. In the rapidly evolving landscape of generative AI, examining the effect of behavioral intention on actual usage is especially crucial. Wu et al. (2025) demonstrated that behavioral intention served as a predictive variable for actual usage behavior concerning ChatGPT. When employees recognize long-term value in generative AI, their continuance intentions are likely to manifest in sustained use. H1: Continuance intention has a positive relationship with generative AI use. User Satisfaction User satisfaction is defined as a subjective and affective evaluation of a service experience relative to prior expectations. When the outputs exceed the user’s initial expectations regarding quality and efficiency, a high level of satisfaction is elicited. This positive emotional state reinforces the user’s commitment to the technology. Zhou and Ma (2025) found that satisfaction with generative AI significantly influences users’ continuance intentions. However, most literature has focused on the continuous usage of generative AI by general users, resulting in a significant gap in research concerning its usage within organizational settings. Organizational members’ professional applications of generative AI are characterized by an increased sensitivity to both the quality of generated outputs and data security. Employees’ psychological bond with generative AI strengthens when they perceive that the tool effectively reduces their workload and provides a seamless user experience. Consequently, this study posits that higher satisfaction levels within the organizational context will influence organization members’ intention to continue using generative AI. H2: User satisfaction has a positive relationship with continuance intention to use generative AI. Perceived Usefulness Perceived usefulness is defined as the degree to which a user believes that using an IS will enhance their task performance. Empirical evidence has consistently demonstrated that perceived usefulness significantly accounts for variance in user satisfaction and behavioral intention. When users observe enhancements in task performance or decision-making efficiency via the use of an IS, they are more inclined to develop positive evaluations of the IS, resulting in increased satisfaction and heightened intentions for continued use. Within the context of generative AI, perceived usefulness becomes particularly salient owing to the technology’s capacity to assist with a variety of cognitively demanding tasks. Generative AI enhances work efficiency by aiding users in document drafting, code generation, data analysis, content summarization, and creative design, thereby reducing the time and cognitive effort needed to accomplish complex tasks. Furthermore, by rapidly searching for relevant information, generating alternative solutions, and explaining complex concepts in an accessible manner, generative AI facilitates prompt decision-making processes. Contemporary empirical studies have validated the applicability of perceived usefulness in generative AI-based service contexts. Kim and Kim (2026) emphasized that perceived usefulness significantly influences the increase of generative AI usage in the workplace. Thus, perceived usefulness functions as a key antecedent of both user satisfaction and continuous usage intention within organizational contexts. H3a: Perceived usefulness has a positive relationship with continuance intention to use generative AI. H3b: Perceived usefulness has a positive relationship with user satisfaction. Perceived Security Perceived security refers to a user’s subjective belief about the extent to which an IS ensures protection, confidentiality, and safe control over data and system interactions. Prior research has consistently demonstrated that perceived security significantly influences users’ intentions to utilize certain IS. Users are more inclined to develop positive attitudes toward the system and perceive lower risk in its use when they believe their information is adequately protected and remains under their control. Beyond mere technical protection, robust security policies mitigate the “privacy paradox,” where users seek the advantages of IS while apprehensive about the exposure of personal data. Particularly in contexts involving sensitive personal or professional data, security serves as a primary trust-building mechanism, a decisive factor in establishing a long-term relationship with the IS. The potential risks are particularly pertinent in the context of machine learning and generative AI systems, which depend on the collection, processing, and generation of outputs from extensive datasets. Although such capabilities significantly enhance task performance and automation, they simultaneously introduce heightened security and privacy risks, including the risk of personal or organizational information leakage and unauthorized data reuse. Veluru (2024) cautioned that AI systems may be vulnerable to malicious manipulation via data tampering, potentially undermining system integrity and resulting in unintended or harmful outcomes. Consequently, organizational members are likely to exhibit heightened sensitivity to security issues when deciding whether to use generative AI for work-related tasks. Additionally, Choudhary and Kar (2025) observed that although generative AI significantly enhances organizational productivity, it could pose risks of sensitive organizational data leakage and the potential misuse of AI in cyberattacks. Thus, perceived security is a fundamental prerequisite for the acceptance and effective utilization of generative AI. Employees frequently handle confidential information, such as customer data, internal reports, strategic documents, and intellectual property, rendering them highly cautious regarding the systems that process such information. When organizational members perceive that generative AI provides robust safeguards against data leakage, unauthorized access, and misuse, they are more inclined to trust the technology and integrate it into their daily work routines. Conversely, concerns regarding insufficient security may prompt employees to limit or entirely avoid the use of generative AI, notwithstanding its potential productivity benefits. Jang (2024) demonstrated that users’ perceptions of enhanced control over their information and diminished risk of data exposure when utilizing AI services significantly increase user engagement. Although the direct motivational role of perceived security on continuance intention and satisfaction has been theoretically argued in prior IS research, its influence in organizational generative AI contexts may be more nuanced. Thus, this study posits that perceived security is associated with user satisfaction and continuance intention. H4a: Perceived security has a positive relationship with continuance intention to use generative AI. H4b: Perceived security has a positive relationship with user satisfaction. Confirmation of Expectation Confirmation of expectation, based on cognitive dissonance theory, indicates the extent to which users’ initial expectations are validated by their actual experiences after using an IS. Individuals typically develop beliefs regarding system performance before adoption and later evaluate the technology by contrasting these expectations with the actual outcomes observed during usage. When post-usage performance fails to meet pre-usage expectations, users experience cognitive dissonance, resulting in negative evaluations and reduced satisfaction. In the context of generative AI, confirmation of expectation occurs when the benefits derived from using the technology, such as efficiency gains, task support, or decision assistance, exceed or align with users’ prior expectations. As organizational members gain hands-on experience with generative AI, confirmation of expectations emerges as a critical cognitive mechanism influencing their post-adoption evaluations. When users perceive that generative AI delivers reliable, accurate, and valuable outputs that align with or exceed their initial anticipations, they are more inclined to reassess the system positively. Confirmation of expectation is expected to enhance perceived usefulness by reinforcing the conviction that generative AI significantly enhances task performance and work efficiency. Moreover, it enhances user satisfaction, as confirmation reduces uncertainty and validates users’ initial adoption decisions. In organizational contexts where generative AI systems manage sensitive data and support critical tasks, confirmation of expectations is likely to enhance perceived security. When the system consistently functions without security breaches and demonstrates stable data protection practices, users’ confidence in its ability to safeguard information increases over time. Lee et al. (2021) conceptualized post-usage performance in AI-based voice services via dimensions, such as price value, hedonic motivation, compatibility, and perceived security, discovering that expectation confirmation positively influenced all of these constructs. When organizational members observe that generative AI meets or exceeds expectations, they are more inclined to perceive the system as beneficial, secure, and satisfying, thereby reinforcing their intention to continue using the technology. Consequently, expectation confirmation is posited as a key antecedent of positive post-adoption beliefs and sustained generative AI usage in organizational contexts. H5a: Confirmation of expectations has a positive relationship with user satisfaction. H5b: Confirmation of expectations has a positive relationship with perceived usefulness. H5c: Confirmation of expectations has a positive relationship with perceived security. Perceived Risk Perceived risk denotes users’ expectations of potential negative consequences associated with IS use, especially concerning privacy loss, data misuse, and security threats. Users are often concerned that a technology may collect, process, or use personal information beyond their control or without their explicit consent, thereby jeopardizing personal privacy. George (2004) conceptualized perceived risk as individuals’ overall concerns about inappropriate data collection, unauthorized secondary use of information, and privacy breaches, highlighting that such concerns function as a psychological barrier that negatively influences users’ decision-making processes. Users’ uncertainty regarding data encryption mechanisms or the secondary use of their inputs for model training may undermine their confidence in the system’s technical and organizational safeguards, regardless of the actual security measures implemented. Perceived risk is particularly salient in the context of generative AI. Organizational members frequently input diverse forms of information and grant broad access permissions while using generative AI for work-related tasks, thereby increasing the likelihood of inadvertently disclosing sensitive personal or organizational data. In organizational contexts, such disclosures may expose employees to personal privacy concerns, potential violations of internal policies, legal regulations, or reputational damage, thereby amplifying perceived risk. Moreover, the opaque nature of generative AI can increase users’ uncertainty about how their inputs are processed and whether their confidential information is adequately protected. This opacity constrains users’ ability to accurately evaluate system reliability and data governance practices, exacerbating perceived risk beyond that observed in conventional IS. Kim and Kim (2026) revealed that perceived risk exerts a key barrier to facilitating employees’ usage of generative AI. Such risk-avoidance behaviors are particularly pronounced in organizational contexts, where uncertainty-related costs may impact both the individual user and the organization. Thus, perceived risk is anticipated to operate as a critical inhibitor across the entire cognitive and behavioral chain of generative AI utilization. H6a: Perceived risk has a negative relationship with generative AI use. H6b: Perceived risk has a negative relationship with continuance intention to use generative AI. H6c: Perceived risk has a negative relationship with user satisfaction. H6d: Perceived risk has a negative relationship with perceived usefulness. H6e: Perceived risk has a negative relationship with perceived security. Moderating Role of Subscription to Premium Version The subscription to the premium version of generative AI is anticipated to significantly moderate the post-adoption process by enhancing the value derived from perceived usefulness and reinforcing its conversion into sustained usage behaviors. Employees who opt for premium versions typically gain access to enhanced features, such as advanced data analysis, faster response times, extended context handling, and deep research capabilities. These advanced features offer a differentiated value proposition that transcends baseline functionality, enabling users to integrate generative AI more deeply into complex and mission-critical work tasks. For premium subscribers, perceived usefulness is embedded within a broader premium-specific value perception, in which higher expectations of reliability, system performance, and security assurances accompany functional benefits. Consequently, when these users evaluate generative AI as useful, they are more likely to continue using the technology because discontinuing the service would result in the loss of both functional benefits and the value associated with paid access. In contrast, employees using free versions encounter minimal financial or procedural switching costs. Even when they perceive generative AI as useful, the lack of economic commitment and feature differentiation makes them more inclined to consider alternative free tools, thereby weakening the correlation between perceived usefulness and continuous intention to use. Accordingly, the premium subscription is anticipated to positively moderate the relationship between perceived usefulness and continuous intention to use generative AI, resulting in a stronger relationship for premium subscribers compared to free users. H7a: A premium subscription moderates the relationship between perceived usefulness and continuous intention to use generative AI. Beyond intention formation, a premium subscription is likely to influence the extent to which continuous intention translates into actual generative AI use. For premium subscribers, the sunk cost of paid access increases motivation to actively utilize the system in daily work practices to justify the investment. This economic and psychological commitment reinforces the intention-behavior linkage, rendering continuous intention a more reliable predictor of actual use. Conversely, free users may express an intention to continue using generative AI, yet fail to consistently translate that intention into behavior, as they can easily disengage or switch platforms without incurring significant losses. Therefore, a premium subscription is expected to positively moderate the relationship between continuous intention to use generative AI and generative AI use. H7b: A premium subscription moderates the relationship between continuance intention to use generative AI and generative AI use. The subscription to the premium version of generative AI is anticipated to moderate the negative effect of perceived risk on generative AI utilizing by attenuating the extent to which risk perceptions translate into actual behavior. Although perceived risk typically deters technology use by heightening concerns about privacy and security threats, premium subscribers may manage these risks differently from free users. Employees who subscribe to premium versions generally correlate paid access with enhanced institutional safeguards and more defined data governance policies. These beliefs can mitigate uncertainty regarding data-handling practices and increase users’ confidence that potential risks are being actively managed. Consequently, even when premium users perceive certain risks, they may be less inclined to discontinue or refrain from utilizing the technology, as their perceived benefits and trust in the provider partially offset risk-related concerns. Accordingly, subscribing to the premium version is expected to weaken the negative relationship between perceived risk and generative AI use, such that the adverse effect of perceived risk on generative AI use is less pronounced among premium subscribers than among free users. H7c: A premium subscription moderates the relationship between perceived risk and generative AI use. METHODOLOGY. Measurement Items and Scales This study adopted a cross-sectional survey design to examine the mechanisms underlying organizational members’ continuous intention to use generative AI. To guarantee content validity and measurement reliability, established measurement items from previous studies in the IS and marketing literature were adopted and meticulously adapted to the context of generative AI usage in organizational settings. Items measuring generative AI use were adapted from Durcikova et al. (2011), who originally developed the scale for knowledge management system use. The original items captured how actively individuals rely on a system to acquire, explore, and apply knowledge in their work—behaviors that closely parallel how employees engage with generative AI in professional settings, where knowledge-seeking is a primary mode of use. Accordingly, the items were adapted to capture how frequently and purposefully employees draw on generative AI for task completion. The measurement items for continuous intention to use generative AI and confirmation of expectation were adapted from Bhattacherjee (2001). User satisfaction was measured using items from Lam et al. (2004), whereas perceived usefulness was operationalized using scales adapted from Kim and Kim (2026). Perceived security was measured using items derived from Yenisey et al. (2005), whereas perceived risk was measured using survey items from Hansen et al. (2018). All survey items were modified to explicitly reflect generative AI usage in the workplace, thereby enhancing clarity and contextual appropriateness. Two researchers specializing in IS independently reviewed the questionnaire to determine face validity and confirm that the wording accurately captured the intended constructs. Based on their feedback, minor revisions were implemented to enhance item clarity, consistency, and readability. All constructs were measured using a seven-point Likert-type scale, which ranged from 1 (strongly disagree) to 7 (strongly agree). Table 1 presents a detailed list of the survey items and their corresponding sources. continued on following page Note. AI = artificial intelligence. Survey Administration and Sample The survey was administered to organizational members employed by companies with prior experience using generative AI for work-related tasks. Given the relatively high level of generative AI adoption among firms in metropolitan areas, the data collection focused on organizational members working in companies based in Seoul. To guarantee data quality, we excluded from the analysis responses that were incomplete, exhibited response patterns indicative of low engagement, or submitted by individuals lacking experience with generative AI in their workplace. Despite the availability of various generative AI tools, this study restricted the sample to users of ChatGPT and Gemini to minimize potential confounding effects stemming from variations in system performance, interface design, and service features among different generative AI platforms. This approach facilitated increased consistency in respondents’ usage experiences and enhanced the internal validity of the findings. Following the application of the screening criteria, 193 valid responses were retained for empirical analysis. The final sample comprised 144 men (74.6%) and 49 women (25.4%), with a mean age of 38.61 years (standard deviation [SD] = 5.37) and a mean professional tenure of 8.89 years (SD = 4.23). The gender composition and age distribution of our sample were broadly consistent with the demographic characteristics of the South Korean corporate workforce. According to a national survey of 150 major Korean corporations conducted by the Korea CXO Research Institute (2024), male employees constituted approximately 75% of the total workforce across industries, with female employees accounting for roughly 25%—a ratio closely mirroring our sample. The predominance of male respondents in our sample therefore reflected a structural feature of the South Korean workforce rather than a sampling artifact. Similarly, the mean respondent age of approximately 39 years aligned with national workforce trends, as the 35–44 age cohort is the most actively employed segment in South Korea's full-time workforce. A total of 79 respondents reported utilizing the premium versions of generative AI for work-related tasks. Table 2 summarizes the detailed demographic information for the respondents. ANALYSIS The analysis was conducted using SmartPLS 4.0, a software tool designed for partial least squares structural equation modeling. Measurement Model The measurement model was evaluated to assess the constructs’ adequacy regarding reliability, convergent validity, discriminant validity, and potential common method bias. First, construct reliability was examined by calculating composite reliability (CR) and average variance extracted (AVE), adhering to the criteria proposed by Fornell and Larcker (1981). The CR reflects the internal consistency of the measurement items, whereas the AVE indicates the extent to which a construct accounts for the variance of its indicators. Prior research has indicated that CR values exceeding 0.70 and AVE values exceeding 0.50 indicate acceptable reliability. Table 3 reports that all constructs in the proposed research model exceeded the recommended thresholds, thereby indicating a satisfactory level of reliability. Subsequently, convergent validity was assessed by examining the factor loadings of each measurement item. Convergent validity necessitates that the factor loading values for each factor exceed 0.70. The results presented in Table 3 indicate that all factor loadings exceeded this threshold, demonstrating that the measurement items appropriately converged on their underlying constructs. Discriminant validity was subsequently evaluated to confirm that each construct was empirically distinct from the others. Discriminant validity is supported when the square root of the AVE for each construct exceeds the correlations between that construct and all other constructs in the model. Table 4 illustrates that the square roots of the AVE values displayed along the diagonal exceeded the corresponding inter-construct correlation coefficients. These results demonstrated that each construct captured unique aspects of the conceptual framework and that discriminant validity was satisfactorily established. Finally, due to the collection of all data through self-reported questionnaires from identical respondents at a singular moment, the possible impact of common method bias was evaluated. Consistent with prior methodological recommendations, Harman’s single-factor test was conducted to determine if a single factor accounted for the majority of the covariance among the measurement items. The results indicated that the largest factor accounted for 39.07% of the total variance, which is below the commonly accepted threshold of 50%. This finding indicated that common method bias was improbable to significantly undermine the validity of the results. Additionally, to assess multicollinearity among the predictor constructs, variance inflation factors were examined for all inner model paths. All variance inflation factor values ranged from 1.035 to 3.509, which were well below the recommended threshold of 5.0, indicating that multicollinearity was not a concern in the structural model. Overall, the measurement model demonstrated adequate reliability, convergent validity, and discriminant validity, with no significant evidence of common method bias identified. Given these satisfactory measurement properties, the study proceeded to evaluate the structural model to test the proposed hypotheses. Note. AI = artificial intelligence; AVE = average variance extracted; CR = composite reliability. continued on following page Note. AI = artificial intelligence; diagonal elements are the square root of the average variance extracted. Structural Model Analysis The structural model analysis was conducted using SmartPLS with the bootstrapping technique involving 5,000 sub-samples. The analysis results are depicted at Figure 2. Note. AI = artificial intelligence. Consistent with our expectation, organizational members’ continuance intention to use generative AI was significantly related to generative AI use. User satisfaction was significantly associated with employees’ continuous intention to use generative AI. Perceived usefulness significantly positively influenced continuous usage intention, yet it did not significantly affect user satisfaction. Unexpectedly, perceived security was not significantly related to neither user satisfaction nor continuance intention to use generative AI in the workplace. As expected, the confirmation of expectations positively influenced user satisfaction, perceived usefulness, and perceived security. Perceived risk significantly negatively impacted generative AI use, user satisfaction, perceived usefulness, and perceived security; however, it did not significantly affect continuance intention to use generative AI. Table 5 summarizes the results. Note. AI = artificial intelligence. The mediation tests are reported in Table 6. The indirect effect of perceived usefulness on continuance intention to use generative AI through user satisfaction was non-significant. Similarly, the indirect pathway from perceived security through user satisfaction to continuance intention and the chained pathway from perceived risk through perceived security and user satisfaction to continuance intention were both non-significant, with 95% confidence intervals that included zero. In contrast, two paths involving perceived risk were statistically significant: the serial mediation path from perceived risk through perceived usefulness and user satisfaction to continuance intention and the indirect path from perceived risk through user satisfaction to continuance intention. These results provide empirical support for the argument that perceived risk dampens continuance intention by eroding perceived usefulness and user satisfaction. continued on following page Note. COI = continuance intention to use generative artificial intelligence; PSE = perceived security; PRI = perceived risk; PUS = perceived usefulness; USS = user satisfaction. Figure 3 shows that a premium subscription does not significantly moderate the relationship between perceived usefulness and continuance intention to use generative AI. For both premium and free users, perceived usefulness positively correlated with continuance intention, and the slopes of the two groups appeared largely comparable. This indicated that the correlation between perceived usefulness and continuance intention was not significantly affected by subscription type. Figure 4 illustrates the moderating effect of subscription to the premium subscription on the correlation between continuance intention to use generative AI and actual generative AI use. Although continuance intention demonstrated a positive association with actual use for both premium and free users, the intensity of this relationship varied significantly by subscription type. Specifically, the slope was steeper for employees utilizing the premium version. Continuance intention was more effectively manifested in actual usage when premium access was provided. The figure illustrates the positive moderating role of a premium subscription in post-adoption generative AI usage behavior. Figure 5 illustrates the moderating role of subscription to the premium version in the correlation between perceived risk and generative AI use. The negative relationship appeared weaker among employees who subscribed to the premium version than among those using the free version. This result indicated that premium subscriptions may have mitigated the adverse effect of perceived risk on actual generative AI use. Thus, the figure provides visual evidence that subscribing to the premium version shaped how perceived risk translated into usage behavior. Note. AI = artificial intelligence. Note. AI = artificial intelligence. DISCUSSION. Theoretical and Practical Implications This study investigated the mechanism underlying employees’ use of generative AI in the workplace. By integrating perceived security and perceived risk into the ECM, this study offers a more nuanced understanding of how organizational members determine whether to continue using generative AI for work-related purposes. Moreover, this study investigated the moderating effects of a premium subscription on their continuance usage decisions. Our proposed research model accounted for 55.1% of the variance in generative AI use and 48.7% of the variance in continuous intention to use generative AI, suggesting that the model effectively elucidated the mechanism underlying employees’ use of generative AI at work. Based on these findings, this study’s academic and practical implications are as follows. First, the results indicated that employees’ continuance intention to use generative AI was significantly related to their actual use. The analysis results indicated that continuance intention served as a key proximal determinant of generative AI use in the workplace. Once employees established a stable intention to continue using generative AI, this intention was likely to translate into habitual usage behaviors as the generative AI became integrated into their regular work practices. Our analysis indicated that user satisfaction with generative AI significantly enhanced employees’ continuous intention to use the technology. These findings highlighted a sequential mechanism wherein positive usage experiences fostered user satisfaction, which in turn strengthened employees’ intentions to continue using generative AI, ultimately leading to actual usage. From a managerial perspective, this implied that the employees’ organizations aimed to enhance employees’ satisfaction by delivering reliable outcomes and clear guidance on appropriate use cases. By ensuring that employees have delightful experiences with generative AI, organizations can promote its sustained integration into daily work activities. Second, the results showed that perceived usefulness significantly enhanced employees’ continuous intention to use the technology, whereas it did not significantly influence user satisfaction. These results indicated that organizational members primarily approached generative AI as an instrumental work tool, evaluating its value based on task efficiency and performance improvement. In workplace contexts, employees typically determined the continuation of technology usage based on its tangible impact on their job performance. Generative AI facilitates various work-related tasks, including information retrieval, document summarization, content drafting, and problem-solving assistance. By alleviating the cognitive and temporal burden linked to these tasks, generative AI enabled employees to execute their work with greater efficiency and effectiveness. Consequently, employees who regarded generative AI as useful were more inclined to establish a strong intention to continue using the technology. However, these instrumental benefits did not inherently translate into user satisfaction, particularly in organizational contexts where efficiency gains, were frequently perceived as expected outcomes rather than sources of enjoyment or emotional reward. From a managerial perspective, organizations must focus on enabling employees to autonomously integrate generative AI into their work processes and clearly recognize its task-related benefits. When employees are given discretion over when and how they use generative AI, perceived usefulness is more likely to translate into continued use. Next, perceived security did not have a significant effect on either user satisfaction or continuance intention, a finding that contrasted with much of the extant literature. This suggested that, in organizational settings, perceived security may not function as a proactive motivational driver of sustained generative AI use. Specifically, while inadequate security perceptions can inhibit use and undermine cognitive evaluations, merely fulfilling security expectations does not inherently enhance satisfaction or reinforce employees’ intentions to continue using the technology. This reconceptualization challenged the prevailing assumption that security is a primary antecedent of satisfaction and intention, highlighting the need to reconsider its role within post-adoption frameworks in organizational AI contexts. Rather than acting as an independent motivational force, security may be better theorized as a boundary condition that shapes the strength of other relationships. Consequently, this study contributed to the IS and generative AI literature by empirically finding that the role of perceived security was more nuanced than previously thought; it appeared that once baseline security expectations were met, they ceased to function as a driver of continued engagement. Fourth, the confirmation of expectations was shown to enhance user satisfaction, perceived usefulness, and perceived security in the context of generative AI use. This result aligned with the core propositions of the ECM. When organizational members perceive generative AI as superior to their initial anticipation, they were more inclined to evaluate the technology favorably. The positive relationship between expectation confirmation and perceived security indicated that meeting expectations diminished uncertainty associated with data handling and system reliability. Specifically, the confirmation of expectations significantly enhanced user satisfaction by reinforcing employees’ overall positive assessment of their experiences with generative AI. When the generative AI delivered outputs that aligned with anticipated levels of task support, efficiency, and reliability, employees were more inclined to feel confident and content with its application in their daily work. For organizations aiming to promote sustained use of generative AI, it is critical to manage expectations before adoption and to ensure consistent performance after implementation. Fifth, the results showed that perceived risk was negatively associated with generative AI use, user satisfaction, perceived usefulness, and perceived security. In work environments, organizational members frequently utilize generative AI to perform tasks, such as summarizing emails, drafting documents, or analyzing internal data. Because the data collection, storage, and secondary use mechanisms of generative AI are not entirely transparent, employees may experience heightened concerns regarding unauthorized data reuse or unintended information leakage. These risk perceptions undermine users’ cognitive and affective evaluations of generative AI. Organizational members who are more likely to question whether the system is beneficial for their work feel less satisfied with its outcomes and doubt the adequacy of its security protections when they perceive higher levels of risk. Such risks also lead to reduced actual use of generative AI, as employees may avoid or limit usage to minimize potential negative consequences. This finding highlighted that perceived risk operates as a potent inhibitor of the evaluation and use of generative AI. However, the direct relationship between perceived risk and continuance intention to use generative AI did not align with our hypothesized direction. The path coefficient was small but marginally significant in the positive direction, running opposite to the negative effect we anticipated in H6b. H6b was therefore not supported in the hypothesized direction. One plausible interpretation in the workplace context is that organizational members who are more sensitive to potential risks do not necessarily intend to abandon generative AI; rather, they may continue to engage with it deliberately as a way of monitoring outputs, validating information, and maintaining personal oversight over how the technology is used. Two indirect pathways emerged as significant in the hypothesized negative direction: perceived risk reduced continuance intention serially through perceived usefulness and user satisfaction, as well as through user satisfaction alone. This finding highlighted the need to reconsider the role of perceived risk within post-adoption frameworks for organizational AI contexts: rather than being modeled as a direct inhibitor of continuance intention, perceived risk is better theorized as an indirect force operating through its negative effects on satisfaction, usefulness, and security perceptions. Organizations should seek to promote effective and sustained generative AI use by mitigating perceived risk. They should seek to enhance transparency regarding data usage, clarify data governance practices, and establish clear organizational guidelines for information protection to encourage more active and confident use of generative AI in work practices. Lastly, the moderating findings of this study represent its most novel theoretical contribution. The non-significant moderation of premium subscriptions on the perceived usefulness-continuance intention relationship is, in itself, theoretically, informative. It reveals that the cognitive value of generative AI is assessed equivalently by both free and premium users, confirming that perceived usefulness is a tier-invariant driver of intention. In contrast, the significant moderation of premium subscriptions on the continuance intention-actual use relationship constituted a key theoretical advance. Existing research on the ECM typically treated the intention-behavior linkage as uniform across users. However, our results revealed that in organizational AI contexts, this linkage is structurally conditioned by the subscription tier. Premium subscribers exhibited a stronger intention-behavior correspondence, driven by sunk-cost psychology, reduced platform friction, and the motivational force of justifying ongoing investment—mechanisms that were entirely absent for free users. This finding reframed the intention-behavior gap as a function of service-tier commitment, opening a new explanatory pathway for ECM research. Equally important is the moderation of premium subscriptions on the perceived risk-actual use relationship. Whereas prior research often treated perceived risk as a uniformly inhibiting factor, our findings showed that the subscription tier functions as an organizational risk buffer. Premium users experienced a significantly attenuated translation of risk into avoidance behavior because paid access signals institutional safeguards and formal data governance that free versions do not convey. This challenged the prevailing assumption of risk homogeneity in AI use research, suggesting that the same level of perceived risk can produce meaningfully different behavioral outcomes depending on the service tier. In summary, the premium tier matters most where behavioral enactment and risk-related restraint are at stake, and it matters least when users are forming cognitive judgments about usefulness. This bounded framing sharpens the study's theoretical contribution. It clarifies for practitioners that the primary value of providing premium access in organizations lies in strengthening intention-behavior follow-through and buffering risk-driven disengagement, rather than in altering how employees judge the usefulness of generative AI. Limitations and Future Research Directions Several limitations should be acknowledged as follows. Firstly, while the sample's gender composition and age profile broadly reflected the structural characteristics of the South Korean corporate workforce, several features of the sampling design limited the generalizability of the findings. The geographic restriction to Seoul-based organizations may not have captured regional variation in organizational culture, generative AI adoption readiness, or autonomy norms across other metropolitan and non-metropolitan areas of South Korea. Given the small sample size and the concentration of respondents in particular regions and industries, external validity was limited. The findings are best understood as evidence specific to the sampled context, and replication studies across diverse national, regional, and industry settings are warranted before stronger generalizability claims can be advanced. Second, the use of generative AI can be influenced by organizational culture. In other words, organizational culture would influence how organizational members perform their tasks. Therefore, future research examining whether organizational culture actually affects organizational members' continuous use of generative AI is necessary. Lastly, this study focused on organizational members’ usage behaviors of generative AI, neglecting the potential variability of the proposed antecedents across different task types and task characteristics. For instance, privacy-related concerns may have exerted a weaker influence in tasks that infrequently involve the handling of sensitive information. Thus, future studies should investigate how the impacts of key antecedents differ based on the specific types of work and tasks involving generative AI. FUNDING STATEMENT This research received support by the 2026 Yeungnam University Research Grant. COMPETING INTERESTS STATEMENT The authors of this publication declare there are no competing interests.