*Article* **Impacts of Digital Technostress and Digital Technology Self-Efficacy on Fintech Usage Intention of Chinese Gen Z Consumers**

**You-Kyung Lee**

Department of Business Administration, College of Management and Economics, Dongguk University, Gyeongju 38066, Korea; yklee2329@dongguk.ac.kr; Tel.: +82-54-770-2329; Fax: +82-54-770-2469

**Abstract:** The role of digital technostress and self-efficacy in digital marketing research is seldom discussed and even more rarely examined among Gen Z consumers. This study investigates the relationships between four sub-dimensions of technostress (complexity, overload, invasion, and uncertainty), digital technology self-efficacy, and fintech usage intention. Data from a total of 266 Chinese Gen Z consumers were used in multiple regression analysis. The results of the study generally support that all sub-dimensions of technostress were negatively related to fintech usage intention. Related to the moderating effects of digital technology self-efficacy on the relationship between the four sub-dimensions of technostress and fintech usage intention, significant interaction effects with complexity and overload were found. Finally, the study discusses the theoretical and managerial implications of the research findings.

**Keywords:** digital technostress; digital techno self-efficacy; fintech usage intention; Chinese Gen Z consumers

#### **1. Introduction**

"Fintech" is a portmanteau formed from the terms finance and technology [1]. It is currently utilized in nearly every consumer financial service—from mobile payment to online investment management service, consumer insurance, and peer-to-peer lending [2]. Fintech is rapidly revolutionizing the financial landscape with the progress of the fourth industrial revolution [3,4]. In particular, the Chinese fintech industry has evolved at a remarkable pace at which the rest of the world struggles to emulate [5–7]. Leading Chinese fintech businesses, such as mobile payment services and big data-based online lending, are at the frontier of the global fintech industry [8]. The Chinese fintech industry has evolved differently from those in developed countries in many ways. While Western countries have mainly developed fintech that focuses on cryptocurrencies or cross-border payment services, Chinese fintech businesses have focused more on consumer mobile financial services, such as mobile payment and online lending [9,10]. Therefore, for Chinese consumers, fintech is becoming a most widely used digital technology that encompasses most onlineto-offline (O2O) commerce from mobile payment to entertainment, education, cultural services, transportation, medical care, and other miscellaneous consumption areas [11]. Therefore, many digital marketing researchers have tried to find determinants of consumers fintech behavior in China as fintech has most vastly reached Chinese consumers. Zhou identified that trust, flow, and satisfaction determine the continuance intention of mobile payment [12]. Chuang et al. found that brand and service trust, perceived usefulness, and perceived ease of use positively related to the adoption of fintech service [13]. Wang et al. found that trust in fintech service and structural assurance can encourage the continuance usage intention of fintech service [14]. While many researchers identified the promoting factors of fintech behavior in digital marketing literature, few studies have focused on the constraints of fintech behavior.

**Citation:** Lee, Y.-K. Impacts of Digital Technostress and Digital Technology Self-Efficacy on Fintech Usage Intention of Chinese Gen Z Consumers. *Sustainability* **2021**, *13*, 5077. https://doi.org/10.3390/ su13095077

Academic Editors: Salvador Cruz Rambaud, Joaquín López Pascual and Marc A. Rosen

Received: 30 March 2021 Accepted: 29 April 2021 Published: 30 April 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

Digital innovations, such as fintech, offer greater convenience and efficiency to consumers, but some consumers experience digital technostress due to the rapid development of digital technology. In this digital revolution era, consumers feel the pressure to quickly adapt to a new digital technology as soon as they have adapted to the previous one. In addition, as the fintech industry evolves, risk of personal privacy infringement, financial accidents, and fraud increase, and these risks are likely to increase consumers' technostress. Technostress has early been defined as a modern disease of adaptation caused by an inability to cope with new computer technologies in a healthy manner [15]. Moreover, it has recently been defined as a physical, behavioral, and psychological strain resulting from information and communication technology (ICT)-driven changes in work environment. Many researchers have examined the impacts of technostress on organizational and personal performances at work because technostress construct was developed in the human resource management research field. However, the impact of technostress on digital technology adoption behavior from the consumers' perspective has hardly been examined. Moreover, research focusing on the technostress of Gen Z consumers—the so-called digital natives [16]—is even more scarce. Therefore, the study tries to empirically examine the relationship between digital technostress and fintech usage behavior among Chinese Gen Z consumers, who most commonly use fintech services in their daily life [16]. In addition, the study tries to verify the moderating effect of digital technology self-efficacy on the relationship between digital technostress and fintech usage behavior among Gen Z consumers. Gen Z consumers are called "digital natives" who have grown up in the digital age and so are likely to have high digital technology self-efficacy; however, they also experience technostress caused by rapidly changing digital technologies. Therefore, the study aims to find a sustainable fintech marketing strategy in the Chinese Gen Z consumer market, which is the most rapidly emerging in the world, through finding new empirical evidence of an interaction effect between technostress and digital technology self-efficacy on fintech usage intention of Gen Z consumers.

The study is expected to expand the scope of digital marketing research by examining the impact of technostress on fintech usage behavior of Gen Z consumers which, unlike technostress, has been mainly researched in terms of work and mental health. For the digital marketing research field, discovering constraint factors of consumer's usage behavior of new digital technology, such as fintech, is as important as finding promoting factors as new digital technologies are expected to be continuously developed and be more widely adopted to a variety of products and services. Digital marketers should find and manage the impediments of fintech usage behavior of Gen Z consumers, which form the most important market segment for digital company. Therefore, the study results are expected to provide practical and academic implications for the digital marketing field.

#### **2. Literature Review and Hypotheses**

#### *2.1. Fintech Growth in China*

Fintech is being used in various consumer financial services, from mobile payment to lending, stocks, insurance, remittances, and asset management. Mobile payment is the most widely used fintech service in China. It was first used in earnest for internet and mobile payment services to support e-commerce consumers in the early 2000s [17]. As of 2019, 87% of Chinese consumers were using fintech services, far ahead of Hong Kong, Singapore, South Korea (67%), and Australia (58%) [17]. Alipay and WeChat Pay, which are non-bank mobile payment business, experienced 75% annual growth between 2015 and 2019. The Chinese mobile payment business has grown in such a way that non-bank companies' mobile payment services dominate banks' mobile banking payment services [18,19]. Tencent, Alibaba, and other major tech firms have been changing the financial services landscape. The mobile payment platform is creating a variety of innovative business models both online and offline, beyond the provision of payment services [20]. The rapid growth of mobile payments by non-bank payment companies in early 2000 resulted in online shopping being quickly replaced by mobile shopping as the Chinese communication

system rapidly jumped from wired communication to wireless communication. The online payment system of Chinese banks was insufficient at that time; thus, the payment system of non-bank internet e-commerce companies, such as Alibaba, could develop significantly. Alipay currently provides total consumer financial services that support various financial activities beyond mobile payment services, including personal asset management, online insurance, loans, and stock trading [19]. In addition, fintech is becoming the digital technology most used in China in various daily consumption activities, such as cultural content, education, medical care, beauty, and housekeeping as well as financial services.

However, fintech has also created serious financial risks and social problems in China. The fintech sector is still in its early stage of development, and many fintech business models are not holistically developed. Chinese authorities are currently trying to formulate new financial regulations for balancing between innovation and stability. However, despite these efforts by the Chinese authorities, the development of fintech exposes consumers to the risk of hacking, ransomware, and financial fraud caused by personal information leakage [10]. Recently, peer-to-peer (P2P) financial transaction accidents have frequently occurred in China, recording millions of victims due to the insolvency of P2P financial companies [8]. Chinese authorities are strengthening the supervision and regulation of the fintech industry as the number of accidents of online payment and P2P lending as well as consumer concerns about financial risks have recently increased [21]. In 2019, CNNIC conducted a survey regarding the problems of most concern when using online services in which 30,000 internet users in 31 regions of China participated. As a result, respondents expressed concern in the order of personal information leakage (20.4%), online transaction fraud (17.0%), hacking or virus infection (10.7%), and account or password theft (9.9%) [11].

#### *2.2. Digital Technostress*

Stress refers to a state in which negative emotions appear in the process of responding to external threats, a physiological imbalance is felt, and involving reacting to survive [22]. In the medical field, researchers have mainly paid attention to the patients' psychological and physiological reactions and the negative effects of stress on the body [23]. Further, the academic fields of sociology, psychology, and business have also begun to pay attention to the effects of stress as the complexity of modern society and the psychological burden of people increased. In particular, many researchers in the human resource management field have paid much attention to the impact of employee job stress on organizational activities and performance [24–29]. In addition, job stress research began to focus on technical stress related to computer or internet use with the rapid development of ICT [15]. Technical stress is addressed in various terms, such as *Technostress*, *Computer Anxiety*, *Negative Computer Attitudes*, *Computer Stress*, *Technophobia*, *Computerphobia*, and *Cyberphobia*. Technostress is a compound word first used in 1982 by the American clinical psychologist Craig Brod, who defined it as a modern disease of adaptation caused by the inability to cope with the new computer technologies in a healthy manner [15]. Hudiburg also defined technostress as an adaptation-related modern disease resulting from the inability to cope with new technologies used in digital devices, such as computers [30]. Shu and Wang found that technostress is positively related to computer literacy and the acceptance of digital technologies [31]. Moreover, Arnetz and Wiholm found that employees who were heavily dependent on computers for their work were usually observed to be in a state of technostress arousal [32].

The existing technostress literature presents technostress as being multi-dimensional, including work overload, invasion of individual life, high complexity of technology, and occupational crisis [15,33]. Salanova et al. and Tarafdar et al. also insisted that technostress consists in the sub-dimensions of technology overload, invasion, complexity, insecurity, and uncertainty [34,35]. Tarafdar et al. developed the technostress measurement scale and validated the construct in the US [35]. The scale consists of five sub-dimensions of technostress that computer technology users can potentially experience at work. First, techno-overload is the stress that emerges when ICTs push employees to work faster. Second, techno-invasion is the stress that emerges when pervasive ICTs invade personal life. Third, techno-complexity is the stress that emerges when the complexity of new ICTs makes employees feel incompetent. Fourth, techno-insecurity is the stress that emerged when fast-changing ICTs threaten the job security of employees. Finally, techno-uncertainty is the stress that is imposed on employees due to the constant changes, upgrades, and bug fixes in ICT hardware and software. Brillhart insisted that technostress consists of four sub-dimensions of data smog, multitasking madness, computer hassles, and burn-out [36]. Ayyagari et al. argued that technostress consists of five sub-dimensions, namely work– home conflict, work overload, invasion of privacy, role ambiguity, and job insecurity, and that they are related to users' perception of ICT's usefulness, complexity, trust, connectivity, anonymity, and development speed [37]. The sub-dimensions of technostress presented in previous research on technostress are shown in Table 1.


**Table 1.** Sub-dimensions of technostress.

Technostress has been examined by digital marketing researchers since ICT began to widely invade general consumers' daily life [36]. Lee and Lee argued that some digital device users tend to stop using digital devices, such as digital breaks or digital detox, to avoid stress, which appears as a side effect of using smart devices [39]. Çoklar and ¸Sahin examined the technostress levels of Turkish social networking services (SNS) users to find that they have a "medium technostress level" [40]. They found that technostress results from the pressure of using technology, remembering large quantities of passwords and usernames, and anxiety regarding data loss [41]. Chen et al. conceptualized technostress as a phenomenon of end users experiencing overload and intrusiveness due to too much information and communication in a short period of time when they use mobile shopping applications [41]. Perceived information overload is referred to as a kind of mental stress when people perceive the environment as a condition exceeding their ability to deal with [42]. According to the stressor–strain–outcome framework, perceived overload induces fatigue and dissatisfaction in the SNS environment, which further increase the discontinuance intentions of SNS users [43]. In addition, perceived intrusiveness lowers the chances of accepting and allowing permission marketing [44]. It was also determined that the social, hedonic, and cognitive uses of social media induce technostress and SNS exhaustion which, in turn, influence a discontinuous use intention based on the stimulus– organism–response framework [45].

New digital technology, such as fintech, provides consumers with convenience and new customer experiences, but it also induces technostress, such as pressure to adapt to new technologies and risks from technological imperfections. Consumers experience technostress while utilizing fintech services, but few studies have verified the impact of technostress from the perspective of fintech users. It is harder to find research on technostress among Gen Z consumers who are always involved in various services and products adopting fintech. Even young and educated consumers are likely to feel difficulty in constantly acquiring new digital technology as this rapidly changes from day to day. In addition to the pressure of acquiring new digital technology that is constantly updated, there are many other types of technostress, such as privacy invasion problems, digital security instability, difficulties in using complex digital devices, and pressure to replace new digital devices due to the continual updates to digital technology. Therefore, the study assumes that consumers' digital technostress negatively affects the usage intention of fintech services based on previous related research. In detail, the study assumes that four sub-dimensions of technostress—complexity, overload, invasion, and uncertainty—are negatively related to usage intention of fintech services [35,38]. Meanwhile, the study excluded insecurity (or job insecurity) as a sub-dimension of technostress which might affect fintech usage intention. Tarafdar et al. and Ayyagari et al. explained that insecurity is a stress that emerged when fast-changing ICTs threaten the job security of employees [35,37]. Therefore, insecurity is not likely to be related with the stress that Gen Z consumers feel when using fintech service. Therefore, hypotheses 1 to 4 are presented as follows:

**Hypothesis 1.** *Digital techno-complexity is negatively related to the usage intention of fintech services.*

**Hypothesis 2.** *Digital techno-overload is negatively related to the usage intention of fintech services.*

**Hypothesis 3.** *Digital techno-invasion is negatively related to the usage intention of fintech services.*

**Hypothesis 4.** *Digital techno-uncertainty is negatively related to the usage intention of fintech services.*

#### *2.3. Digital Technology Self-Efficacy*

Bandura defined self-efficacy as people's judgment of their capabilities to organize and execute courses of action required to attain designated types of performances [46]. Self-efficacy is a strong sense of personal efficacy related to better health, higher achievement, and more social integration, and it represents the key construct in social cognitive theory [46–48]. Bagozzi defined self-efficacy as an individual's confidence in their own work ability. Self-efficacy has received much attention in the business literature [49]. Gist and Mitchell found that people who think they can perform their task well show better work performance than those who think that they will fail [50]. In the organizational behavior research field, researchers found that self-efficacy is positively related to job proficiency and performance, and self-efficacy lowers the negative impact of job stress on job performance [51–53].

Meanwhile, as ICT invades every corner of people's life, such as work, school, and daily lives, technology self-efficacy is attracting much attention in many research disciplines, such as psychology, education, and business. Cassidy and Eachus presented computer user self-efficacy as a factor that contributes to success in tasks in the domain of computer technology [54]. They further adapted to cover digital self-efficacy to measure individual self-efficacy in the digital domain. Self-efficacy-related ICT is often used in terms of computer efficacy or internet efficacy [55]. Venkatesh and Davis defined computer selfefficacy as a self-assessment of one's ability to use information technology or one's belief that people can use computer or internet-related technologies well [56]. Compeau and Higgins defined computer self-efficacy as a self-judgment of one's ability to use information technology [57]. Rogers found that technology self-efficacy is a trait that is variable at an individual level and positively influences the acceptance of new technologies, and that technology self-efficacy has a positive relationship with the innovation and acceptance of new technologies of organizational leaders [58]. Table 2 summarizes the antecedents and outcome variables of technological self-efficacy used in previous studies related to technological self-efficacy. Meanwhile, many researchers pay much attention to the impact of consumers' technology self-efficacy on the acceptance behavior of ICT products or services since ICT began to be widely used for general consumers. According to Bandura's theory, people with high self-efficacy tend to believe they can perform well even if they are in difficult situations, and tend to view difficult tasks as something to be mastered rather than something to be avoided [48]. Therefore, people with high self-efficacy are

likely to put more effort into learning technological skills, while those with low technology self-efficacy are likely to put in relatively little effort or give up halfway. In addition, people with high technology self-efficacy find using new technology relatively to be less difficult and show a positive attitude toward using technology [58]. The study therefore assumes that digital technology self-efficacy positively affects the usage intention of fintech services and presents the following hypothesis 5:

#### **Hypothesis 5.** *Digital technology self-efficacy is positively related to the usage intention of fintech services.*

Perceived self-efficacy to control thought processes is a key factor in regulating stress and depression [46]. People with high self-efficacy tend to approach threatening situations with the assurance that they can have control over situations, and their efficacious thought reduces stress and lowers vulnerability to depression [46]. A significant amount of research has shown that self-efficacy acts to decrease people's potential for negative stress by increasing their belief of being in control of the threatening situations they encounter. The perception of being in control represents an important buffer of negative stress [59]. Lu et al. found that managerial self-efficacy had significant moderating effects on the stressor–strain relationship in the Chinese workplace [60]. Self-efficacy was also found to be a stress moderator in some of the stressor–work well-being relationships among employees in Hong Kong and Beijing. Some researchers have found that mobile users with high self-efficacy prefer to take more proactive behavior to deal with stressors of mobile shopping apps [41]. Although little research has tried to examine the relationship between digital technology self-efficacy, technostress, and new digital technologies' usage intention from the consumer perspective, many researchers in clinical, educational, social, business management, health, and personality psychology disciplines have found that self-efficacy lowers the negative effects of stress. People with high self-efficacy can accurately perceive their situation and self-manage themselves in stressful situations; thus, self-efficacy is positively related with an active lifestyle. Therefore, technology self-efficacy is likely to lower the negative effects resulting from people's psychological anxiety or stress caused by new digital technology. Based on previous research arguments, the study presents the following hypotheses 6 to 9:

**Hypothesis 6.** *Digital technology self-efficacy lowers the negative impact of digital technocomplexity on the usage intention of fintech services.*

**Hypothesis 7.** *Digital technology self-efficacy lowers the negative impact of digital techno-overload on the usage intention of fintech services.*

**Hypothesis 8.** *Digital technology self-efficacy lowers the negative impact of digital techno-invasion on the usage intention of fintech services.*

**Hypothesis 9.** *Digital technology self-efficacy lowers the negative impact of digital technouncertainty on the usage intention of fintech services.*

#### **3. Research Model and Methodology**

#### *3.1. Research Model*

The research model is developed based on the assumption that the four dimensions of technostress (complexity, overload, invasion, and uncertainty) resulting from rapidly changing digital technology are negatively related to fintech usage intention. Constructs rooted in the secondary evaluation procedure (digital technology self-efficacy) are also considered as determinants to fintech usage intention. In addition, it is also argued that digital technology self-efficacy moderates the relationship between technostress and fintech usage intention. For the convenience of notation, the study will use abbreviations of the constructs in the latter part of this paper. Figure 1 illustrates the research model.

**Figure 1.** Conceptual research model.

#### *3.2. Data Collection and Sampling*

The study collected data by means of an online survey administered by Wenjuan Xing (www.wjx.cn, accessed on 18 September 2020), which is a professional online survey website in China. A pilot test was conducted in July 2020 for 30 Chinese undergraduate students at D university in Korea that did not form part of the sampling frame of the main study, so as to assert the reliability of the scales used the questionnaire [61,62]. The feedback resulted from a pilot test was used to refine a final questionnaire. Data collection used a snowball sampling method in August to September 2020 in which the online survey URL was transmitted to the respondents who had previously agreed to receive it. The study collected a total of 314 responses from the participants. It excluded samples with a less than 20% response rate of all measurement items or missing responses to the outcome variable to ensure the external validity of the data, in addition to considering the subject scope. The study used 266 samples for the final analysis.

Table 2 provides demographic information on the sample. The number of male respondents, at 53.4%, was slightly higher than that of females at 46.6%. Of the respondents, 94.7% of participants were single, and 5.3% were married. More than 80% had bachelor's degrees and higher, and around 70% respondents had an average monthly personal income under CNY 2000. Furthermore, 70.7% of the respondents answered that they had used a smartphone for over five years.

#### *3.3. Construct Measurement*

The construct measurement scale employed in the study was taken from existing literature, and all constructs dealing with perceptions were measured using five-point Likert scales (1 = *strongly disagree* and 5 = *strongly agree*). The operational definition and measurement scale for constructs are as follows. The study first defined digital technostress (DTS) as a psychological pressure consumers feel from using digital technology and digital devices. The study modified the technostress measurement scale of Tarafdar et al. based on the scope and purpose of the study and used the modified measurement scale to measure the sub-dimensions of technostress: CPX, OVL, IVS, and UCT [35]. The study measured CPX as four items, OVL as four items, IVS as three items, and UCT as two items, as shown in Table 3. Next, the study defined digital technology self-efficacy (DTSE) as a psychological self-belief that people can utilize digital technology well, and developed three measurement items based on the measurement scale of Cassidy and Eachus [54]. Finally, the study defined FUI as a consumer's intention to choose and use fintech services as much as possible and developed a measurement scale for FUI based on the technology acceptance model (TAM) [63]. The full survey instrument is presented in Table 3.


**Table 2.** Demographic information of respondents.

**Table 3.** Constructs and measurement items.


#### *3.4. Research Methodology*

The data analysis methods used in the study are as follows. First, frequency analysis was conducted to investigate the demographic characteristics of respondents. Second, the feasibility and reliability tests of the measurement scale were conducted to examine the predictability and accuracy of constructs. Third, correlation analysis was conducted to examine the correlations among constructs. Fourth, moderated regression analysis (MRA) was conducted to examine the relationships between constructs using IBM SPSS 20.0. MRA is an analytic approach that maintains the integrity of a sample yet provides a basis for controlling the effects of a moderator variable; therefore, MRA can avoid the loss of information resulting from an artificial transformation of a continuous variable into a qualitative one [65]. The study adopts the MRA to build three regression Equations as follows, and it examines the equality of the regression coefficients for the following three regression equations:

$$\mathbf{Y} = \mathbf{a} + \mathbf{b}\_1 \mathbf{X}\_1 + \mathbf{b}\_2 \mathbf{X}\_2 + \mathbf{b}\_3 \mathbf{X}\_3 + \mathbf{b}\_4 \mathbf{X}\_4 \tag{1}$$

$$\mathbf{Y} = \mathbf{a} + \mathbf{b}\_1 \boldsymbol{\chi}\_1 + \mathbf{b}\_2 \boldsymbol{\chi}\_2 + \mathbf{b}\_3 \boldsymbol{\chi}\_3 + \mathbf{b}\_4 \boldsymbol{\chi}\_4 + \mathbf{b}\_5 \mathbf{Z} \tag{2}$$

$$\mathbf{Y} = \mathbf{a} + \mathbf{b}\_1 \mathbf{X}\_1 + \mathbf{b}\_2 \mathbf{X}\_2 + \mathbf{b}\_3 \mathbf{X}\_3 + \mathbf{b}\_4 \mathbf{X}\_4 + \mathbf{b}\_5 \mathbf{Z} + \mathbf{b}\_6 \mathbf{X}\_1 \mathbf{Z} + \mathbf{b}\_7 \mathbf{X}\_2 \mathbf{Z} + \mathbf{b}\_8 \mathbf{X}\_3 \mathbf{Z} + \mathbf{b}\_9 \mathbf{X}\_4 \mathbf{Z} \tag{3}$$

In the above equations, if (2) and (3) are not significantly different, then Z is not a moderating variable but a simple independent variable. If Equations (1) and (2) are not different from each other but different from Equation (3), then Z is a pure moderating variable. Lastly, if Equations (1)–(3) are not different from each other, then Z is a quasimoderating variable. The study adopts the above moderated regression analysis approach to identify the research model.

#### **4. Empirical Analysis and Results**

#### *4.1. Validity and Reliability of Measurement Instruments*

The study first assessed the validity and reliability of the measurement model. An exploratory factor analysis was conducted on the 19 items relating to variables. Six principal component factors were extracted, as they had a cut-off factor loading of 0.6 and an eigenvalue greater than 1 [66]. Of the total variances, CPX accounted for 19.44%, OVL accounted for 12.97%, IVS accounted for 12.41%, UCT accounted for 8.29%, DTSE accounted for 23.89%, and FUI accounted for 19.44%. The six factors accounted for 84.79% of the total variability. The rotated component matrix of the factor analysis is shown in Table 4. Regarding the construct reliability of the six factors, all values for Cronbach's α exceeded the threshold value of 0.7. This provides sufficient evidence for the high reliability of constructs listed above [67]. The detailed results of the validity and reliability analysis are shown in Table 4.

#### *4.2. Correlation Test*

Table 5 shows the correlation matrix between the constructs. This study used partial correlation to measure nonlinear as well as linear relationships between variables. Most variables show a relatively low correlation of less than 0.6, which demonstrates that there is little chance for multicollinearity to exist between the constructs. The relationships between variables in the correlation matrix are consistent with the direction of the hypotheses. In addition, although the constructs show low Pearson correlation coefficients, nonlinear relationships between them may still exist [68].


**Table 4.** Measurement item's loading (λ) and construct's convergent validity.

CPX: complexity; OVL: overload; IVS: invasion; UCT: uncertainty; DTSE: digital technology self-efficacy; FUI: fintech usage intention.

**Table 5.** Correlations between constructs (*n* = 266).


Note: \* *p* < 0.05, \*\* *p* < 0.01. gen.: gender; edu.: education; inc.: income; sup.: smartphone usage period; CPX: complexity; OVL: overload; IVS: invasion; UCT: uncertainty; DTSE: digital technology self-efficacy; FUI: fintech usage intention.

#### *4.3. Hypotheses Test*

This study conducted a hierarchical regression analysis to find more detailed causal relationships among variables. First, the study set gender, education, income, and smartphone usage period as control variables; it then verified the influences of the control variables on FUI in Model 1. The results found that the *F* value was 1.594, and R2 was 0.024; therefore, Model 1 was not statistically significant. Next, in Model 2, regression analysis was conducted on the impacts of the control variables and four sub-dimensions of technostress (CPX, OVL, IVS, and UCT) on FUI. The results found that the *F* value was 6.225, and R<sup>2</sup> was 0.163; therefore, Model 2 was statistically significant. In detail, Model 2 demonstrated that OVL, IVS, and UCT negatively affect FUI (*β* = −0.177, *p* < 0.05; *β* = −0.151, *p* < 0.05; *β* = −0.228, *p* < 0.01). In Model 3, regression analysis was conducted to analyze the impacts of control variables, four sub-dimensions of technostress (CPX, OVL, IVS, and UCT), and DTSE on FUI. The results found that the *F* value was 12.996, and R<sup>2</sup> was 0.314; therefore, Model 3 was statistically significant. Model 3 demonstrated that IVS has a significant negative impact of on FUI (*β* = −0.133, *p* < 0.05), and DTSE has a significant

positive impact on FUI (*β* = 0.470, *p* < 0.01). Finally, in Model 4, regression analysis was conducted to examine the impacts of control variables, four sub-dimensions of technostress (CPX, OVL, IVS, and UCT), DTSE, and the four interaction variables (CPX×DTSE, OVL×DTSE, IVS×DTSE, and UCT×DTSE) on FUI. The results found that the *F* value was 12.110, and R2 was 0.385; therefore, Model 4 was statistically significant. Model 4 demonstrated that the four control variables have no significant impacts on FUI. The four sub-dimensions of technostress (CPX, OVL, IVS, and UCT) are all negatively related to FUI (*β* = −0.615, *p* < 0.05; *β* = −0.800, *p* < 0.01; *β* = −0.544, *p* < 0.01; *β* = −0.420, *p* < 0.05), while DTSE has a significant positive impact on FUI (*β* = 0.661, *p* < 0.01). Of the interaction variables, the results showed that CPX×DTSE and OVL×DTSE interactions have significant negative impacts on FUI (β = −0.357, *p* < 0.05; β = −0.498, *p* < 0.05). In addition, impacts of CPX×DTSE and OVL×DTSE interactions on FUI (β = −0.357, *p* < 0.05; β = −0.498, *p* < 0.05) were lower than the direct impacts of CPX and OVL on FUI (β = −0.615, *p* < 0.05; β = −0.800, *p* < 0.01) in Model 4. In result, DTSE lower the negative impacts of CPX and OVL on FUI.

Meanwhile, the study verified the statistical significance of direct and moderating effects of variables by comparing the regression coefficients of each model [65]. As a result of estimating the analysis model of the study with the regression Equations of Model 2 and 3, the explanatory power of Model 3 increased at a statistically significant level in comparison with Model 2 (F = 20.81 \*\*). In addition, the explanatory power of Model 4 increased at a statistically significant level (F = 3.73 \*\*) in the comparison of the explanatory power of Model 3 and Model 4 [65,69,70]. Therefore, the study finally interpreted the research results based on Model 4. In Model 4, the four sub-dimensions of technostress (CPX, OVL, IVS, and UCT) all negatively affect FUI; thus, H1 to H4 are supported. In addition, DTSE has a positive impact on FUI; hence, H5 is supported. Finally, of the interaction variables, the results of Model 4 showed that CPX×DTSE and OVL×DTSE interactions have significant negative impacts on FUI; therefore, H6 and H7 are supported. The detailed analysis results are shown in Table 6 below.


**Table 6.** Results of multiple regression analysis (MRA).

Note: \* *p* < 0.1, \*\* *p* < 0.05, \*\*\* *p <* 0.001.

#### **5. Discussion and Conclusions**

#### *5.1. Summary and Discussion*

The study aims to verify the impact of digital technostress and digital technology selfefficacy on the usage intention of fintech services among Chinese Gen Z consumers, who are the most exposed to advanced digital technologies, such as fintech [16]. In particular, as consumers are currently experiencing technostress due to the rapid development of digital technologies, including fintech, the study focused on the negative effects of technostress on the usage intention of fintech services. In addition, the study assumed that digital technology self-efficacy not only has a direct positive effect on fintech usage intention but also a moderating effect on the relationship between digital technostress and fintech usage intention. The summary of the empirical analysis results is as follows.

First, it was found that all four sub-dimensions of DTS (CPX, OVL, IVS, and UCT) had a statistically significant negative effect on FUI. The abovementioned empirical analysis results are consistent with the results of previous research [37–41]. It was found that Chinese Gen Z consumers with high perception of CPX, OVL, IVS, and UCT show a lower intention to use fintech services. Therefore, Hypothesis 1 to 4 were supported. Next, the DTSE of Chinese Gen Z consumers was found to increase their intention to use Fintech service, which is consistent with previous research results [54,55,57]. Therefore, hypotheses 5 was supported. Lastly, DTSE was found to significantly lower the negative impact of CPX and OVL on FUI, while DTSE has not shown statistically significant interaction effects with IVS and UCT on FUI. Therefore, hypotheses 6 and 7 were supported, while hypotheses 8 and 9 were rejected.

The study showed that Gen Z consumers experience digital technostress due to rapidly changing digital technology, and the digital technostress negatively affect fintech usage intention of Gen Z consumers. Therefore, the empirical results of the study are contradicted to the previous study's argument that Gen Z consumers generally show a very positive psychological response to digital technology [16,71]. According to the above study findings, digital marketers and researchers should consider novel approaches to predict fintech usage behavior of Gen Z consumers. Meanwhile, the study also found that DTSE has moderating effects on the negative impacts of CPX and OVL on FUI. The interaction effect between DTSE and technostress among Gen Z consumers is a very new finding for digital marketing research field. This seems because Gen Z consumers with high DTSE have self-belief to respond the negative effects of techno-complexity and techno-overload on fintech usage intention in the consumer's individual level. However, the study found that DTSE has no moderating effect on the negative impacts of IVS and UCT on FUI. It seems because techno-invasion and techno-uncertainty are structural problem that is difficult to respond in the consumer's individual level. Therefore, the study results can offer digital marketers with practical implications that they should actively utilize digital technology self-efficacy to manage technostress which can be handled in the individual-level, such as technocomplexity and techno-overload. In addition, digital marketers must also prepare special measures to reduce the negative impact of structural technostress, such as techno-invasion and techno-uncertainty, on Gen Z consumers fintech usage intention.

#### *5.2. Conclusions*

The study results not only offer practical implications to fintech marketers but also contribute academic implications to the digital marketing research field. First, according to results of the study, fintech marketers should develop media-based materials, such as pictures, animations, or videos, through which consumers can more easily and quickly understand the features of new digital technologies and how to use them, by considering the behavioral traits of Gen Z consumers. Second, fintech marketers should present higher level of norms and regulations for personal privacy and security issues. Third, it is important to be careful not to directly expose consumers to excessively frequent updates or digital technology changes and to establish a more meticulous marketing strategy to reduce the increased cognitive and emotional burden on consumers due to digital technology changes. Such marketing efforts can lower the digital technology technostress of consumers, contributing to forming consumers' positive attitude and behavior to a wider variety of fintech services. Finally, fintech marketers should focus on a marketing strategy that can increase Gen Z consumers' DTSE as the study found a significant positive direct effect of DTSE on FUI and moderating effects of DTSE on the relationship between DTS and FUI. Therefore, fintech marketers should provide various ways for Gen Z consumers to understand and learn new digital technologies with ease and enjoyment through various media means to increase a level of Gen Z consumers' DTSE. In addition, digital marketing researchers need to have a broader perspective to find more various impediments, such as technostress, which negatively influence consumers' adoption and usage behavior of new digital technologies like fintech. In particular, examining the impacts of new negative factors, such as technostress, in a new consumer segment like Gen Z can contribute to broadening the academic scope of digital marketing.

Despite the academic and practical contributions of this study presented above, this study has the following limitations. First, the number of samples used in this study is small compared to China's population; therefore, future research will have to collect a larger amount of data for empirical analysis. In addition, the study results should be carefully interpreted as the sample size is not large enough. Second, in the case of the technostress variable, there will be large differences according to consumers' age groups; it is thus necessary to compare different impacts of DTSE on fintech behavior between age groups in future research. Third, the study has limitations in reflecting demographic and regional diversity in China; therefore, future research should consider collecting data in various consumer segments in China to compare the distinctions of fintech usage behavior. Finally, this study used consumers' comprehensive and general usage intention of various fintech services as outcome variables. However, a wide variety of new fintech services have recently been launched in the Chinese market which are being widely accepted by consumers. Therefore, future research should consider the differences in various types of fintech services and fintech consumption behavior for a more comprehensive understanding of Chinese fintech behavior.

**Funding:** This study received no external funding.

**Institutional Review Board Statement:** Ethical review and approval were waived for this study, because the study didn't directly confront the online survey respondents and use personally identifiable information either. In addition, the study didn't collect or record "sensitive information" from the respondents, so it can't directly or indirectly identify the respondents at all.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Acknowledgments:** I would like to thank Dongguk University Gyeongju Campus for supporting the study.

**Conflicts of Interest:** The author declares no conflict of interest.

#### **References**

	- Conceptual development and empirical validation. *Inf. Syst. Res.* **2008**, *19*, 417–433. [CrossRef]

### *Article* **Toward a Chatbot for Financial Sustainability**

**Sewoong Hwang <sup>1</sup> and Jonghyuk Kim 2,\***


**Abstract:** This study examines technology effectiveness for industry demand in which artificial intelligence (AI) is applied in the financial sector. It summarizes prior studies on chatbot and customer service and investigates theories on acceptance attitudes for innovative technologies. By setting variables, the study examines bank revenue methodologically and assesses the impact of customer service and chatbot on bank revenues through customer age classification. The results indicate that new product-oriented funds or housing subscription savings are more suitable for purchase through customer service than through chatbot. However, services for existing products through chatbot positively affect banks' net income. When classified by age, purchases by the majority age group in the channel positively affect bank profits. Finally, there is a tendency to process small banking transactions through the chatbot system, which saves transaction and management costs, positively affecting profits. Through empirical analysis, we first examine the effect of an AI-based chatbot system implemented to strengthen financial soundness and suggest policy alternatives. Second, we use banking data to increase the study's real-life applicability and prove that problems in customer service can be solved through a chatbot system. Finally, we investigate how resistance to technology can be reduced and efficiently accommodated.

**Keywords:** chatbot; artificial intelligence; financial sustainability; telemarketing; cube model; voice recognition and conversion model

#### **1. Introduction**

Professor Yoshua Bengio, the winner of the 2019 Turing Award, gave a lecture on core technologies in deep learning, such as meta-learning and reinforcement learning, at the Samsung AI Forum 2020 in November 2020. He refuted what Professor Carl Benedict Frey had argued, citing success stories in the application of information technology (IT) in the financial sector. Professor Frey had argued that less than half of financial jobs were set to disappear with the increasing use of artificial intelligence (AI). However, Professor Bengio predicted, Professor Frey's arguments would lose their convincing power [1], as it had happened for Professor Zoonky Lee who had published articles in a Korean newspaper to combat prejudice against artificial intelligence (Special Series of *JoongAng Daily*, "Lee Zoonky, Ask about the Future") [2]. The common points between Lee and Frey are as follows. Considering the history of technology adoption, technological innovation should be considered as a digital transformation that changes roles rather than kills jobs. Hence, as AI grows, digital transformation occurs and people seek new roles. A chatbot, which provides advice on financial products to customers, applies AI to the financial industry. Both Lee and Frey conclude that a chatbot does not eliminate jobs; rather, humans use the chatbot system to venture into new areas. The lack of insight, imagination, and responsiveness to new variables of the chatbot algorithms require humans to resolve them. Thus, AI creates a new ecosystem within the industry, and the role of humans changes for a new era in which machines and humans coexist in a complementary way [3].

**Citation:** Hwang, S.; Kim, J. Toward a Chatbot for Financial Sustainability. *Sustainability* **2021**, *13*, 3173. https://doi.org/10.3390/su13063173

Academic Editor: Salvador Cruz Rambaud

Received: 17 February 2021 Accepted: 11 March 2021 Published: 13 March 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

Natural language processing technology and speech recognition technology are currently providing personal assistant services that communicate directly through personal mobile devices [4]. Chatbot, an interactive AI, has been widely deployed in finance, retail, public, and manufacturing industries. Apple's Siri, Amazon's Alexa, Google's OKgoogle, and Samsung Electronics' Bixby are good examples of voice conversion personal assistant services. In addition, Naver and Nugu of SK Telecom provide high-quality voice recognition services through the Korean search platform and communication market. As such, major IT companies, portal sites, and telecommunication companies worldwide are developing commercial voice recognition services and investing significant financial resources to provide AI services with higher accuracy [5].

This study investigates the role of AI in the financial industry from several aspects. First, the social demand for and expectations from artificial intelligence in the financial sector are high. The amount of investment in this field is larger compared to other industries such as the distribution, manufacturing, and public sectors [6]. Second, despite this interest, there are many misconceptions around using AI, and whether AI has been properly applied to the financial industry has been questioned [7]. Finally, the systematic criticism of AI technology applied to financial products is lacking in extant research. Recent media comments about AI describe the positive and negative effects of AI in a stimulating tone [8]. However, it is difficult to find an in-depth comparative study. Many recent studies examine the combination of AI and the financial sector because anyone who engages in economic activities is a financial consumer. Furthermore, even software used exclusively by traditional asset managers can be downloaded easily and used by ordinary people [9]. Hence, AI in the financial industry is simply a tool that individuals can access and use; it is not the exclusive property of experts. This study compares and analyzes the impact of customer service through the existing automatic response service (ARS) with the chatbot system currently being used by banks. In addition, we empirically analyze data to determine how each of these two channels (customer service and chatbot) affects bank profits and then derive practical implications based on the results [10].

The article structure is as follows. Section 2 reviews prior extant research, divided into four areas. First, through the latest research on financial chatbot systems, we investigate AI technologies and their effects in financial environments. Second, we summarize the research on problems faced by customer service counseling staff and on coping strategies and techniques to solve them. Third, we study theories of effective ways to introduce technology. Finally, we examine prior research on indicators representing bank contribution from a methodological perspective. In Section 3, we set the hypothesis for this study and perform statistical analysis based on bank data for new products and existing services by channel. We conduct quantitative analysis using statistical techniques to establish and verify the hypotheses while considering prior studies and descriptive statistics. Section 4 evaluates the theoretical underpinnings verified by data and summarizes the study. Finally, we conclude this study by revealing implications, limitations, and future research plans.

#### **2. Background**

#### *2.1. Financial Chatbot Service*

The term "chatbot" is a combination of "chatting" and "robot," which is commonly used for text messages or messengers. A chatbot is a communication software that can store appropriate answers to questions on a server, create models that continuously develop correct answers through conversations with customers, control exceptions, and provide accurate answers [11]. Chatbots create a self-learning model through computer programs and mathematical calculations and provide customers with answers and other relevant information as close as possible to user questions in real time. For companies, a chatbot is an interface that provides information required by customers and marketing through communication with financial consumers [12]. The first chatbot service in the financial sector was Bank of America's Erica introduced in May 2017. Erica's early look was similar to Apple's Siri. Erica provided simple text and voice-based responses, including transaction details, limit amount, and account balance. Additionally, it provided advanced services such as credit rating upgrade application, fund product introduction, bank loan application, interest rate guidance, utility bill payment, and fund management consulting services [13,14]. The chatbot learned customers' personal profile information, past financial product purchase history, location information, and personal routine data by applying machine learning and deep learning technologies to provide accurate and customized services. Customers could enjoy convenience by securing personalized financial services quickly and easily through a chatbot [15]. In Korea, most commercial banks and other types of financial institutions have introduced chatbot services for customers (Table 1).


**Table 1.** Financial chatbot services in Korea.

Chatbots can be classified into a retrieval model and a generative model according to the implementation method in web or mobile applications. First, a chatbot based on the retrieval model is a rule-based method that provides prepared answers according to conditions of a specific topic or question. Most early chatbot versions in financial institutions were developed based on rules [16]. However, with the commercialization of chatbot, sophisticated machine learning has become possible as industry data continue to accumulate. Second, the generative model chatbot is a deep learning method that improves the accuracy of new responses through self-evolution as customer and communication data accumulate [17]. With the latest developments in deep learning technology, the system understands the customer's question and the intent of the sentence and presents the appropriate answer to the customer [18,19]. Therefore, it is possible to recommend personalized products for customers. Studies for commercialization are being actively conducted to capture current emotions of users through individual routines and basic profiles. Despite its many advantages, cost is an issue with the generative model because it requires the accumulation of vast amounts of data for continuous self-evolution [20].

Chatbots are important in terms of technology and user interface (UI). The chatbot is a technology service that implements communication between users and AI-based on text and voice and is a representative non-face-to-face service. Most chatbot services are implemented through conversational interactions based on customer questions and chatbot responses [21]. Through the interaction process with machines, customers perceive chatbots as objects of communication rather than simple machines [22]. Therefore, the chatbot service should be designed to reflect user needs and planned as an efficient and proven system with clear interactions. Chatbot services are mostly text-based messengers in online or mobile applications [23]. Therefore, a UI that enables customers to input and check information on a small screen effectively is essential. As shown in Table 2, design elements, such as product composition, button position, and background color, may vary depending on the screen. Therefore, the chatbot's design needs to project user experience on the screen elaborately. In addition, the screen of the chatbot is a publicity vehicle that presents the image of the company [24].


**Table 2.** Interface design elements of the financial chatbot.

#### *2.2. Telemarketing and Technical Elements of Alternative Systems*

Customer services centers provide online consultation with and for customers. They operate under various names such as customer support centers, call centers, contact centers, and customer relationship management (CRM) centers, depending on the company [25]. Initial customer service began as an organization that performed simple response services by receiving calls from customers [26]. In recent years, it has transformed into an organization that creates added value by enhancing the corporate image, providing information on products, conducting marketing and promoting activities, providing customer service, and communicating with customers. Customer service is an organization that provides non-face-to-face interactions with customers, but these interactions require emotional labor beyond face-to-face channels [27]. Customer service's emotional labor is an essential element of a company, as it can retain existing customers, attract new customers, and maintain a company's competitive advantage. However, this causes considerable stress on workers due to the incongruity of internal emotions and external expressions. These difficulties have led many companies to build systems that replace customer service [28,29].

Many technical elements are required to build an alternative customer service system. The customer service helper system must respond appropriately by inferring the meaning of the customer's question in real time [30]. For this reason, a semantic reasoning technique that can infer the meaning of a customer's query and provide an appropriate answer is essential. Semantic reasoning techniques can be classified into two broad categories according to their development process. First, knowledge-based question-and-answer (Q&A) structures are used by humans in the system using an ontological method. This method finds the result by inferring the large-scale knowledge database built in a logical form. Second, information retrieval-based Q&A orders a list of answers through probabilistic calculations by searching for answers to questions based on indexes in a large document set [31,32].

For the alternative customer service system to provide intelligent services, a Q&A method through ontology-based reasoning should be used, rather than a simple rulebased search. Recently, owing to the development of deep learning technology, ontologybased Q&A technology has been used in industries and chatbots in the financial sector (Table 3) [33]. Recently, the application of AI and advanced statistical analysis has enabled users to control local information, weather guidance, Internet searches, route searches, and product searches. These systems can provide advanced services based on user experiences [34].


**Table 3.** Components of intelligent virtual assistant technology.

Another essential element for alternative customer service systems is voice recognition technology. In 1952, the AT&T Bell Laboratory in the United States developed the first technology to convert recordings into text. Since then, various laboratories have attempted to develop speech recognition, but the accuracy has not exceeded 80%. The low accuracy of voice recognition is due to different accents, volume, degree of dialect, and background noise [35].

Figure 1 illustrates a recently developed two-step voice recognition and conversion model that leverages deep learning techniques to recover ambiguous speech and further clarifies speech semantic transmission by considering speech characteristics and the surrounding environment. Voice recognition techniques are evolving into deep learning-based systems that can recognize speech, including long sentences or dialogues [36].

**Figure 1.** Two-step voice recognition and conversion model.

#### *2.3. Intention to Accept New Technology and Its Spread*

Due to internal conflicts and external situations of the system, it is difficult to accommodate and apply innovative technologies to existing systems to create a completely different paradigm [37]. In the financial sector, especially in organizations that perform customer services using mainly call centers, considerable conflicts, along with trial and error, will occur when applying chatbot services initially [10].

This study examines five theories on technology acceptance and diffusion. First, the theory of reasoned action (TRA) is the basis for acceptance and proliferation of new technologies, which argues that consumer attitudes influence behavior and that behavior can be predicted if attitudes are accurately measured. In particular, this theory presupposes that people are highly rational and systematically use the information they have. TRA has three components—attitude, subjective norm, and intention [38].

Second, the newly defined technology acceptance model (TAM) is based on TRA and focuses on user evaluation of the technology as a model to emphasize individual characteristics or beliefs in the process of accepting technology [39]. TAM argues that the greater the perceived ease and perceived helpfulness of users, the greater the behavior and intention of using technology. Used by several researchers, TAM is recognized as an excellent model that demonstrates simple and high explanatory power in explaining users' IT acceptance and utilization behaviors [40].

Third, diffusion of innovation theory comprehensively describes the process by which a new paradigm of innovation is accepted and adopted by a particular organization or individual [41]. The theory considers the psychological rejection of accepting new techniques. Innovation resistance is the tendency of individuals to maintain their status quo. Created perceived risk is an important concept in accepting technologies. Perceived risk is the user's subjective perception of uncertainty about the future and possible negative consequences [42].

Fourth, the theory of planned behavior (TPB) is different from all of the above because it includes intentional action and strategic intention, and planned behavior control. This theory argues that control of intended and planned actions should be added to the performance of actions. TPB emphasizes that the main determinants of behavior are not the individual's attitude toward it but the intention to perform it and that it is under human control. From this perspective, we add a new concept, a critical variable called perceived behavior control, which sufficiently compensates for the weaknesses in rational behavior theory (Figure 2) [43].

**Figure 2.** Schematic of the theory of planned behavior (TPB) model.

Fifth, the unified theory of acceptance and use of technology (UTAUT) is a highly descriptive model, because it has been selected as a significant proven factor through numerous trials and verification procedures [44]. In particular, studies analyzing acceptance of fintech payment services, by applying additional variables called reliability to UTAUT, show that individual effort, social impact, and reliability have a positive impact on the acceptance of fintech services. Furthermore, studies using UTAUT in consumer use of internet banking have shown that variables, such as information security risks, uncertainty risks, and transaction efficiency, have a negative impact on the dispersion of internet banking. Prior research results demonstrate that UTAUT is suitable for measuring the intent to use chatbot services introduced by many financial institutions. Studies have found that variables, such as consumer performance expectations, social impact, and promotion conditions, have a significant effect on bank performance [45].

Through the various technology acceptance models described above, we deduce a positive effect of lowering internal resistance and encouraging pro-sustainability behavior, even though there is the disadvantage of being slightly expensive strategically as several variables are added. In addition, we expect to be able to develop models for advancing theory improvement and environmental policy formulation [46].

#### *2.4. Profitability Indicators*

Research using detailed profitability indicator data from companies is limited. However, many studies in finance and accounting have used stock market data through the Open API (Application Programming Interface) as a dependent variable. Research has been conducted on the quality of services that are difficult to determine quantitatively [47]. In addition, there are many studies on how political factors cause instability in the financial industry. An empirical analysis of the Bank of Korea's profitability determinants and policy measures that conducted a regression analysis using independent variables such as equity ratio, per capita expenses, assets per capita, total receivable growth rate, and corporate bond yields [48]. A study of profitability determinants for commercial banks in Japan empirically analyzed how the classification of ownership structures affects profitability. The study used gross asset net profit margins, return on equity (ROE), and net interest margin (NIM) as indicators of profitability [49]. Furthermore, a Korean study conducted a multi-regression analysis using major financial indicators and macroeconomic data of general banks from 2000 to 2009 to identify the profitability determinants of banks. The study found that the non-profit loan ratio (NPL) had a statistically significant effect on the profitability of commercial and local banks, and that poor loan management in banks had a significant impact on asset size [50].

Research in the financial sector, which specializes in financial profitability, examined bank profitability determinants in Europe, North America, and Australia, using gross asset net return, return on capital (ROC), and value-added return on total assets as indicators of profitability [51]. Another study identified the impact of each independent variable on the subsidiary variable using gross capital operating profit, gross capital net income, gross capital net return, and net sales net income, of which gross asset net income was the most effective indicator [52]. Other studies compared the profitability and efficiency of commercial and local banks to examine the impact on the bank's management performance and suggested ways to stabilize the profitability of local banks. This was an empirical analysis of the factors affecting profitability with time-series cross-section regression, using portfolio mix as a methodology, and using changes in stock prices and gross capital return as an indicator [53]. In addition, long-term time-series data from 22 general banks were used to ascertain the determinants of the bank's management performance using the net return on assets and the ratio of non-profitable loans as an indicator of the general bank's profitability. These results demonstrate that macroeconomic variables affect bank asset portfolio and productivity variables. Another study used approximately 10 years of accounting data from Spain, Portugal, Germany, and France to analyze the relationship between net return on assets and net return on equity and profitability on commercial banks [54,55].

#### **3. Methods**

#### *3.1. Samples and Data Collection*

This study analyzed product data from a large Korean banking company to determine the impact of customer product and service purchases on bank profits (return rate increase) based on two channels using ARS. We analyzed the statistical significance of how much each channel contributed to bank profits based on customer information using financial products and services through customer service calls or chatbot systems. We anticipate that our analysis will help banks derive measures to secure financial stability. Furthermore, we expect to empirically derive the extent to which AI-based chatbot can replace the existing customer service business for all financial affiliates, including banks. From Bank A, we collected 34,089 personal data of four major products and services sold through the chatbot channel for 36 months (on a daily basis) from January 2018 to December 2020, when the chatbot was first introduced at this bank. In addition, we collected 317,438 unstructured voice data acquired through customer service based on similar products at the same time. We standardized the unstructured data through a text conversion system and used a twostep voice recognition and conversion model. Except that each of the four products was

handled through customer service or chatbot, all conditions were completely the same; therefore, it is safe to assume that the statistical effect is controllable in advance. Bank A is a nationwide commercial bank, with its target customers individuals residing in Korea; it handled all products during the time of the study. Therefore, the conditions for recognition of region, seasonality, environment, and age are the same. In addition, statistical sampling bias is assumed to be controlled, because the data handled were not part of the extracted data but the parameter data for the entire product. However, unlike chatbots, in the case of responses through counseling staff, there may be a promotional event depending on the period. Therefore, the purchase of a product different from the original purpose may occur due to a specific promotion. However, this cannot be measured quantitatively, and it can be assumed that the effect of counselor promotion is negligible in a situation in which the response to customer purchases is the primary purpose of inbound calls. We deleted sensitive information from Bank A's customer data through blur-masking. In addition, we made the response to the information protection request by performing mixed-combination conversion of the primary key and set it as data that can be analyzed through data cleansing. Financial product information as final analysis data is shown in Table 4.

**Table 4.** Classification of sample data.


Bank A's main products are new sales of funds and home subscription savings products, loan interest payment and repayment services, and local taxes and utility expenses payment services, with a total of 351,527 cases. Since we used the analysis data based on the number of cases, we counted all duplicate product purchases. The data collected included contract channel (contract manager, chatbot unique allocation code), contract date and contract product, contractor's identification information, and contract number for the individual number of all products. Based on the data, we performed basic statistical information, data preprocessing, hypothesis setting, and testing. We used SAS University Edition, an open-source software, for statistical analysis and the Oracle virtual machine to prepare the software operating environment.

#### *3.2. Operational Definition and Preprocessing*

Information on the four financial products selected for this study and product information for each channel through offline counters and online ARS is summarized as follows. First, in the case of funds, the total assets of listed funds (ETFs) handled by six major Korean banking companies amounted to USD 50 billion at the end of 2019. Adding unlisted funds, the amount is over USD 100 billion, which is an increase of 26.1% year-on-year—this is classified into 335 domestic products and 115 overseas products. By investor entity, individual investors account for 38.6% and institutional and foreign investors, 61.4%.

The second product group is housing subscription savings. As of August 2019, the number of Korean subscribers exceeded 25 million, accounting for 50% of the population, and the total amount exceeded USD 80 billion, with savings per person averaging at USD 3000.

The third product group is service products related to loan interest payment and repayment. At the end of 2019, the total amount of personal loans exceeded USD 1.3 trillion, including the amount on credit cards. The average loan per capita is USD 60,000, and interest expenses were, on average, over USD 300 per month.

The final product group, the amount paid in utility bills including local taxes and administrative fees is not large; however, recently commercial banks are promoting a policy to increase the number or amount through promotions. These policies have positive benefits for high interest rates and currency exchange; hence, consumers are also actively using this system. Table 5 shows the data preprocessing status.


**Table 5.** Data preprocessing according to variable classification.

\* marked variable is newly created data for preprocessing.

As shown in Table 5, the \* marked variable is newly created data for preprocessing. However, in the case of the fund's approval date, it may differ from the sale date depending on the product's contract terms and the buyer's credit terms. For this study, age groups were classified as "Junior" for individuals less than 45 years old, and "Senior" for individuals of 45 years and older. Purchasing channels were classified by the employee number—58 employees of the ARS team at the bank's head office—and the unique codes of employees of five other inbound marketing service companies. In the case of the chatbot, the purchasing channels were classified with Bank A's own chatbot allocation code starting with "CB00". All amount-related variables were calculated based on the total amount received by the bank for each product in the period. To estimate the effect of product-specific returns on bank contribution, we created a new variable of net increase or net income excluding costs from profits by using the gross return on assets (ROA), which was used as a dependent variable in previous studies. We set the customer service cost formula by dividing the number of customers by the sum of labor cost and organizational operation

cost. We used the average annual depreciation cost ratio of general system infrastructure of 11.3% and general management cost for server operation to calculate the formula of development and operation cost for the chatbot. We divided this amount by the number of chatbot users and calculated the average cost per chatbot use. As a result, the final cost was set at USD 1.03 per case for customer service and USD 0.39 for the chatbot.

#### *3.3. Descriptive Analysis*

As shown in Table 6, in the specific classification of each channel-product group, among all consumers who purchased all financial products using ARS, the number of customers who purchased products through customer service was about 9.3 times more than those who purchased the same products through the chatbot. Therefore, 90.3% of the parameter data purchased products through customer service, whereas purchases through chatbot only remained at 9.7%. In terms of age groups, the purchase of products and services through customer service is higher in the Senior group (54.7%) than in the Junior group (45.3%). This trend is the same for all four products sold through customer service. In particular, in the case of housing subscription savings, the gap widens by 14.6%, which is approximately 5% more compared to the average of 9.5%. In terms of the product purchase rate, 55.8% of customers use customer service to pay utility bills.


**Table 6.** Descriptive statistics.

Regarding consumers using the chatbot, the distribution of purchases is completely different from that of customer service. First, in terms of frequency of use, the Junior group (63.7%) clearly used the chatbot more than the Senior group (36.3%). However, in terms of the product purchase rate, 61.1% of users, which is higher than customer service, used the chatbot for utility bill payment services. In addition, the frequency of purchases of funds and housing subscription savings, which are subscriptions for new products, is completely different from payment of loan interest or utility bills, which are services for existing products. The most striking statistic related to the difference between the chatbot and customer service channels is that customer service occupies a higher proportion of handling new products at 10.5% and 17.3%, compared to 7.4% and 12.8% of the chatbot.

#### *3.4. Hypotheses*

Considering the statistics in the case of new product sales, the ratio of total purchases per channel was lower in chatbot than in customer service. Conversely, in terms of loan interest payment and utility bill management, the chatbot has a higher relative ratio than customer service. Based on these data, we posit the following hypothesis to fit the assumption of the null hypothesis that there is no basis for expecting that new product purchases through customer service will have a greater positive effect on bank profits than purchases through the chatbot:

**Hypotheses 1 (H1).** *Comparing customer service and chatbot users, there is no difference in their impact on bank contribution according to product classification.*

Considering age groups, the data demonstrated that the relative proportion of seniors is considerably large for products handled through customer service than for those handled through the chatbot. Conversely, in the case of product handling through chatbot, the proportion of Junior users was higher than that of Seniors. Therefore, we expect that specific age groups will have a greater positive effect on bank profits in the division by channel, and we propose the following hypothesis to fit into the null hypothesis assumption similar to H1:

**Hypotheses 2 (H2).** *Comparing customer service and chatbot users, there is no difference in their impact on bank contribution according to customer classification.*

Finally, we examined the concrete effects of the two hypotheses. We created a cube model with a combination of four cases in the form of 2 × 2 by mixing product groups and customer age groups. We analyzed the effect of each combination on the increase or decrease in the bank's net income. We categorized the sale of funds and housingsubscription savings products as "new product sales," and categorized loan interest and payment of utility bills as "existing service provisions." Utilizing these categories and the two age groups, we developed the four areas as follows: (1) New product sales–Junior group, (2) Existing service provision–Junior group, (3) New product sales–Senior group, and (4) Existing service provision–Senior group. Table 7 presents the data of the cube combination.


**Table 7.** The relative ratio of rows and columns by cube combination.

We present the following hypotheses for each of the four combinations to investigate their bank contribution:

**Hypotheses 3a (H3a).** *In the case of the Junior group who purchased new products, there was no difference in the degree of contribution to the bank according to the classification by channel;*

**Hypotheses 3b (H3b).** *In the case of the Junior group that received the existing services, there was no difference in the degree of contribution to the bank according to the classification by channel;*

**Hypotheses 3c (H3c).** *In the case of the Senior group who purchased new products, there was no difference in the degree of contribution to the bank according to the classification by channel;*

**Hypotheses 3d (H3d).** *In the case of the Senior group that received the existing services, there was no difference in the degree of contribution to the bank according to the classification by channel.*

#### **4. Results**

#### *4.1. Statistical Hypothesis Testing*

To test Hypothesis 1, which assumes that there is no difference in the impact on bank contribution of customer service and chatbot users according to product classification, we performed an analysis of covariance (ANCOVA), as shown in Table 8.


**Table 8.** Covariance analysis test for Hypothesis 1 (H1).

The results show that both purchases of new products and existing services have a significant effect on the increase or decrease in bank profits according to customer service and chatbot channels. This means that new product-oriented funds and housing subscription savings are more suitable for customer service than the chatbot. Conversely, services for existing products, such as loan interest or payment of utility bills, are more suitable for processing through chatbot, which has a positive effect on bank net income.

Hypothesis 2 assumes that there is no difference in the impact on bank contribution of customer service and chatbot users according to customer classification. We performed ANCOVA, as shown in Table 9, to test two or more elements, as in Hypothesis 1.


**Table 9.** Covariance analysis test for Hypothesis 2 (H2).

Hypothesis 2 secured model suitability according to the F-test result (F = 70.1013). From the results (Table 9), we conclude that both Junior and Senior customers have a significant effect on the increase or decrease of bank revenues according to the two customer channels—customer services and chatbot. In the case of product purchase through customer service, the proportion of Seniors was higher, while the proportion of Juniors was larger for the chatbot. In conclusion, the age group that occupies a relatively large proportion has a positive effect on bank profits.

The total number of samples in Hypothesis 3a is 43,955, which are Junior group customers purchasing new products. The dependent variable is the net increase in bank revenue. We tested the statistical significance of the difference according to the classification by channel.

In the case of Hypothesis 3a (Table 10), the assumption of equal variance is satisfied by the F test (F = 8.12). Therefore, we refer to the pooled t-test, and the test result accepts the hypothesis (t = 1.4352). Hence, when comparing customers who purchase products through customer service and customers who purchase products through chatbot, that there is no difference in the bank net profit (New products–Junior group).


**Table 10.** Two-sample *t*-test for Hypothesis 3a (H3a).

Hypothesis 3b is classified by product–customer, and the total number of samples is 121,466 users: Junior customers receiving existing services. The dependent variable is the net increase in revenue for the bank. We tested the statistical significance of differences in channel classification.

In the case of Hypothesis 3b (Table 11), the assumption of equal variance is satisfied by the F test (F = 6.19). Therefore, we referred to the pooled t-test, and the test result rejected the hypothesis (t = 18.2142). That is, when comparing customers who purchase through customer service and those who purchase through chatbot, bank net profits from the customer groups (Existing service–Junior group) are statistically different. In the case of the Junior group receiving only existing service, the bank profit was higher from the chatbot group than from the customer service group. The junior group's handling of small amounts of multiple utility bills through the chatbot has a positive effect on bank finances due to the regular transaction costs of customer service. Therefore, inducing the use of chatbots with low operating costs is a positive contribution to the bank, due to the nature of existing services involving a small amount of money but a larger number of transactions.

**Table 11.** Two-sample *t*-test for Hypothesis 3b (H3b).


In Hypothesis 3c, the total number of samples classified by product and customer is 51,313 because they are customers who purchase new products and belong to the Senior group. The dependent variable is the net increase in bank revenue. We tested the statistical significance of differences in channel classification.

For Hypothesis 3c (Table 12), the assumption of equal variance was not satisfied by the F test (F = 23.73). Therefore, we refer to the *t*-test of the Satterthwaite method, and the test result rejects the hypothesis (t = 34.1223). When comparing customers who purchase through customer service and those who purchase through chatbot, bank net profits of these group customers (New product purchase–Senior group) are not statistically equal. This result is due to the large number of Senior group customers who purchase new products such as funds and savings through customer service. Additionally, the amount of fund products is large. This increases the average bank receipts. In addition, since the housing subscription savings product has regularity, it is expected to have a positive role in terms of bank contribution.

**Table 12.** Two-sample *t*-test for Hypothesis 3c (H3c).


In Hypothesis 3d, when categorized by product and customer, the total number of samples is 134,793; they are users receiving existing services and customers belonging to the Senior group by age. The dependent variable is the net increase in revenue for the bank. We tested the statistical significance of differences in channel classification.

In the case of Hypothesis 3d (Table 13), the assumption of equal variance is not satisfied by the F test (F = 12.09). Therefore, we refer to the Satterthwaite method t-test, and the test result rejects the hypothesis (t = −12.1025). Contrary to Hypothesis 3b, in the case of the Senior group, handling existing services with high transaction frequency and small monetary amounts through customer services has high transaction costs and a negative effect on bank revenue.


**Table 13.** Two-sample *t*-test for Hypothesis 3d (H3d).

#### *4.2. Cube Model Interpretation*

To plot the results of Hypothesis 3, the combination of two conditions by product and by age was made into a 2 × 2 cube model. The X-axis is divided into the age group of customers, and the Y-axis is divided into product characteristics. In addition, we divided the channels into customer service and chatbot. We plotted the four combinations and analyzed the effect of each combination on bank revenue. The analysis results for each combination are shown in Table 14.


**Table 14.** Interpretation of the hypotheses from the cube model.

As for X1–Y1, the hypothesis of the study was adopted, so there is no difference in the effect on the net profit of banking operations between the two channels. In the case of X1– Y2, the analysis result was significant, because multiple small transactions were able to save labor and management costs through automated processing. Additionally, X2–Y1 positively affected contribution based on the behavior of the Senior group purchasing new products with large amounts of money. Finally, in the case of X2–Y2, multiple micro-transactions using a chatbot rather than using customer service positively affect bank finances.

#### **5. Conclusions**

This study conducted an empirical analysis to pursue the expansion of the use of AI-enabled chatbot in banking financial products and bank policy changes, based on the ARS data of leading banks. For empirical analysis, we summarized the practical implications through the results of hypotheses setting and testing. First, we empirically analyzed the effect of the AI-based chatbot system and suggested policy alternatives to strengthen the financial soundness of large banks. We evaluated the performance of the chatbot system, newly introduced to the existing ARS system in January 2018. In addition, we presented alternatives on how this system contributed financially to banks and what aspects should be supplemented to optimize customized profits in the future. The findings indicate that reinforcing customer service expertise according to product and age classification increases bank profits. In some chatbot cases, the increase is greater. Second, companies, especially in the financial sector, are furiously building AI platforms. However, applying new technologies to the field, including acceptance and adaptations, requires considerable time and public relations, and may result in internal friction. This can affect short-term profits and may lead to economic opportunity losses. If companies fail to make the right investment at the right time, they may forfeit future opportunities. Therefore, this study categorized whether banks are investing with an eye to profits and analyzed the effectiveness of these investments. This study can be applicable to financial institutions other than banks in the future.

We examined previous studies in four dimensions and in that backdrop, summarize the academic contributions of this study. First, considering the financial chatbot system, we examined AI technologies and effects introduced in various financial environments through prior research. Second, in relation to the ARS system, we summarized the practical problems of customer service counseling staff and the countermeasures and techniques to solve them. Third, we studied the properties of resistance to the introduction of technologies and theories related to alternatives that help reduce the resistance and increase acceptance. Fourth, we investigated prior research on actual indicators representing bank contribution from a methodological perspective. Thus, this study provides a real-world situation through data and meaningful statistical inference.

Despite the various academic significances and practical contributions described above, there are problems and limitations of this study. First, data handled at offline counters that account for most product management were excluded. Banks sell bankspecific savings and loan products, and they offer specialized products such as insurance and bonds. The percentage of products sold through ARS is less than 5% of the bank's total sales. Of these, sales through chatbot are insignificant, less than 10%; hence, it may be unreasonable to closely associate them with bank profits. However, building a new infrastructure for a chatbot is an important factor, considering the unknown impact for the new era. Therefore, continuous research on the introduction of the AI financial system is necessary. Second, the four products and services presented in this study are all parameters of the data accumulated for two years after the chatbot was introduced. These data were developed through trial and error at the time of initial settlement, and the stability of the sample is poor. In addition, there are many macro-environment variables that should be considered along with the impacts presented in this study. This is expected to be a problem that can be resolved naturally as data are continuously accumulated and the system is stabilized in the future. Nonetheless, it remains practically and academically necessary to continuously correct these problems for research. Third, we also need to design an experiment by separating the cases of failure from the success cases in the chatbot service and additionally analyze the service failure factors [56]. In other words, we need meticulous research to control situations that are unfamiliar to customers through further investigation of chatbot service failures. Fourth, we overlooked dealing with digital governance issues. The main challenge in digital governance is not technical but the people involved in the decision-making process [57]. In other words, it is important to create a governance structure so that people participate in decision making and at the same time do not fall into the trap of knowledge issues. Therefore, we need to provide multiple processes at different levels for a sustainable transition to digital governance. Finally, we omitted the study of distorted trust between social cognition and the cognitive ability of chatbots [58]. In other words, we need to list the significant negative impacts of a number of faulty interfaces that could be considered in the conceptual model of a chatbot and provide reasonable evidence of its impact on users. We expect that through the process of closing this set of limitations, we will be able to more accurately relocate the contributions of our research to the digitization of society through chatbots.

**Author Contributions:** Data curation, S.H. and J.K.; formal analysis, S.H. and J.K.; funding acquisition, J.K.; methodology, S.H. and J.K.; project administration, J.K.; visualization, S.H.; writing original draft preparation, S.H. and J.K.; writing—review and editing, J.K. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**

