How do AI and human users interact? Positioning of AI and human users in customer service

1 Introduction

In the wake of rapid advancements in artificial intelligence (AI), a shift is underway within service industries, marked by a proliferation of innovative AI-based service agents. These agents, manifested in live chat interfaces or voice assistant systems, utilize natural language processing to facilitate two-way communication with human users (Gnewuch et al. 2017). As an increasingly integral component of technology-mediated communication, numerous South Korean organizations and companies such as SK Telecom, Korea Telecom, and Naver are now leveraging these AI-based service agents to meet customer needs, thus replacing human agents. AI service agents have emerged as a potent source of service revolution, offering interactions of scale and consistent quality (Rust and Huang 2014; Wirtz et al. 2018).

These agents have also been observed to elicit feelings of warmth in human users (Kim et al. 2019). Wirtz et al. (2018) have explored perceptions, beliefs, and ethical implications of robot-delivered services, and other studies have also begun to examine the quality of communication, user perceptions, consumer attitudes, satisfaction, and repurchase/revisit patterns (Mero 2018). Kim et al. (2019) have found that the anthropomorphism of consumer robots increased psychological warmth but decreased positive attitudes, suggesting a balance to avoid the ‘uncanny valley’ phenomenon.

Amid these developments, this study aims to explore how human users communicate with AI-based voice assistants. Unlike research focusing on the effects of AI-based service agents, our study delves into the interactional patterns between human users and AI-based voice assistants. Our methodology builds upon the principles of conversation analysis (CA), a research tradition that interprets social interaction in terms of sequences of actions. The study considers how the AI-based voice assistants manage the balance between pre-programmed responses and the need to adapt to unpredictable elements of human communication. It also probes how these agents use linguistic and discursive strategies to maintain a ‘natural’ conversation flow.

2 Literature review

Many studies have sought to understand the possibilities of AI, concentrating primarily on its accuracy (Coniam 2014) as well as its effective use for resource management in the workplace, including applicant screening during the hiring process (Mehta et al. 2013; Park et al. 2021; Van Esch et al. 2019), and customer service (Følstad and Skjuve 2019; Xu et al. 2017). Conversational AI agents, both text-based and voice-based, emerge as pivotal in streamlining operations and reducing costs, with voice-based agents extending emotional support in specialized fields such as healthcare and therapy (Cha et al. 2021; Elgarf et al. 2021). Fuoli et al. (2021) extend this understanding to social media platforms, revealing how companies’ responses to customer feedback on Twitter, particularly those employing an affective style, can effectively manage customer relations and mitigate reputational risks. Liao and Sundar (2021) also explore how AI maintains conversation with customers based on their collected personal information. The role of AI in digital customer engagement, particularly through webcare, emphasizes the importance of communication styles in managing customer relations. Companies’ engagement in webcare to address online feedback highlights the evolving landscape of customer service, where digital interactions are becoming increasingly significant (Fuoli and Bednarek 2022; Van Noort and Willemsen 2012).

In addressing the distinct nature of human-computer interactions compared to natural conversations, it becomes evident that the former lacks the spontaneity and depth of human-to-human exchange. AI service agents, as actors in this domain, are designed to simulate these natural conversational patterns to bridge this gap. The imitation of natural dialogue is crucial in enhancing user experience and trust, making interactions with AI more intuitive and less mechanical (Chandra et al. 2022). This endeavor to replicate natural conversational dynamics is especially pertinent in service encounters over the telephone, where the absence of visual cues makes verbal interaction the sole medium of communication.

The debate over informal versus formal communication styles in this context reveals a preference for a conversational tone that fosters a more personal and engaging customer experience (Kelleher 2009; Schamari and Schaefers 2015; Sung and Kim 2018). However, the effectiveness of these styles may vary across cultural contexts and brand familiarity, suggesting a complex interplay between communication style, customer perception, and brand identity (Gretry et al. 2017; Steinmann et al. 2015). While many previous studies explore the perception and use of AI-based conversational agents, a closer scrutiny of how these conversations unfold employing a conversation analytic lens can provide deeper insights into the nuances of human-AI communication.

3 Data and methodology

We gathered data logs of customer service encounters between human customers and AI-based voice assistants provided by Korea Telecom, one of Korea’s largest telecommunications service providers. Currently, South Korea is experiencing a rapid integration of AI into customer service, making Korea Telecom an ideal site for examining the cutting-edge dynamics of human-AI communication in the telecommunications industry. Our dataset comprises transcripts of 200 telephone interactions, all originally conducted in Korean and subsequently translated into English. These 200 telephone interactions were randomly selected by the service provider, with 100 representing successful cases and 100 representing failed cases. The determination of success or failure is based on the company’s assessment of how the AI succeeded or failed to capture the intent and provide an appropriate subsequent turn. Customers are informed at the outset that their calls may be monitored and used for varying purposes. To further ensure individual privacy, all personally identifiable information except age was anonymized by the provider prior to analysis. The data had to be screened by the provider, and nonverbal cues were thereby excluded from our consideration.

Each interaction began with the human customer calling the service center, typically seeking resolution for an issue or soliciting information, with AI-based service agents offering immediate reception and guidance. The data, collected from January to June of 2022, comprises calls with longer scripts and extended dialogues, typically lasting over 3 min, while shorter interactions are estimated to last between one and 2 min. This methodological approach enables us to dissect the intricate dialogue patterns that emerge between human users and AI agents, revealing the challenges and successes of AI in replicating human-like conversation dynamics. Employing this framework, the investigation draws on over 200 verbal interactions between AI and human users, offering insights into the multifaceted dialogue patterns that emerge in human-AI communication.

4 Analysis

In an effort to gain a nuanced understanding of how these human and non-human interlocutors position each other and form relationships, our analytical focus is on the initial adjacency pair, specifically the sequence of the FPP and SPP, to understand the relational dynamics initiated by AI agents and the corresponding user responses. This examination sheds light on the strategic construction of FPP by AI to induce specific SPPs from users, thereby facilitating a study on the relational positioning between human and AI interlocutors. Moreover, we explore the varied responses from human users as they navigate and formulate their SPP, offering a rich analysis of the interactional patterns that emerge between humans and AI in service contexts. With the conversation undertaken as a collaborative endeavor, we observe dialogic engagement and cooperation between the human users and the AI-based service agents as they attempt to negotiate meaning. With the conversation being driven by the customer’s needs at hand, whether the customer was seeking information or assistance for problem-solving, our findings yield insights into the variety of patterns among their SPP.

5 Conclusions

As interactions involve the collaborative endeavor of interlocutors, the present study has sought to examine how AI-based service agents and human users participate in interactions to meet the needs of the human user. The study has paid particular attention to the first pair-part (FPP), the opening led by the AI-based voice assistants, and the second pair-part (SPP) of the human users. While the AI’s FPP created a certain hierarchical relationship between the two, the SPP of the human users indicated their view and expectation of the AI-based voice assistants. Among the four types of SPP observed, the human users most frequently responded to the AI’s FPP with keywords, potentially signaling their underlying beliefs in how they should communicate with AI-based service agents, providing keywords as if searching online to most effectively and efficiently retrieve and access the information they wanted and meet their needs. That is, the pre-existing belief and attitude of human users toward these AI-based voice assistants may play a significant role in their engagement in communication. Following the enunciation of keywords, the second most frequently observed SPPs was to show no response. Without much contextual information, it is challenging to understand how no verbal input took place. Still, this finding suggests the need to question whether the five-second wait is sufficient and the following response by AI is adequate.

This brings us to the broader cultural context that underpins these interactions. Influenced by Confucianism, Korea has a set of honorific morphemes, address terms, and grammaticalized polite forms. Here, the linguistic indexicality, particularly the use of honorifics and address terms, contributed to the positioning and distance between the AI-based voice assistants and the human users as well as the formality of the conversation. These linguistic elements do more than signal relational and social structures within society; they encode a complex matrix of expectations and norms that extend into the realm of human-AI interaction. Such indexical politeness signals the relationships between the speaker and referents in the sentence (Brown 2011).

The present study has focused on production and formatting of utterances in both FPP and SPP between the interlocutors as their exchange influences how they dynamically establish the relationship and participate in the remaining interaction. As the AI-based voice assistants establish a hierarchical relationship with the human users through the mixture of polite and humble forms, we can understand that the service provider aims to provide the human users with more power, social superiority and potential closeness. By embedding hierarchical and polite forms in AI communication, service providers may inadvertently reinforce social structures, thus affecting users’ perceptions of and interactions with technology. However, such positioning of interlocutors may constantly be in flux depending on how human users participate in a conversation.

The findings offer important implications for AI research for conversational interfaces and applied linguistics, as they demonstrate that the understanding of interactional pragmatics and conversation patterns may further enhance the quality of machine-to-human interaction. The study suggests that service providers need to pay attention to how their AI-based service agents interact with and engage in conversations with their customers in a more systematic manner. Similar to other human-to-human conversations, it is important to understand that these interactions operate in the social, cultural, and ideological worlds, influenced by the language at hand. As interests and attempts to create more human-like AI agents and to offer more human-to-human-like interactions continue or reinforce the commitment to the development of such technology, the present study has a significant applied value, as it can inform the design of conversational interfaces in machine-to-human interaction. In an environment characterized by more interactions between humanoid service robots or AI-based service agents, more research into the conversational pattern is needed, as early experiences of such a communication system would influence how human users adhere to the norms and develop rules of conversation, including turn-taking organization, adjacency pairs, and repair, thus contributing to the consolidation of a shared set of conversational principles with these AI-based service assistants.

While the Korean language has an honorific system that reflects the social relationships among participants, future studies could examine how human users and AI-based service agents interact in languages without the honorific system. Moreover, as mentioned previously, with the conversation being held over the phone, the human user might have dropped the phone when the AI-based voice assistant began talking. Thus, without having full access to the context in which human users maintain the conversation, it is difficult to conclude that there is no response, signaling the lack of participation in interaction. Finally, the present study calls for further research on how ‘natural’ AI-based service agents should be. Unveiling how human consumers envision and prefer AI-based service agents to talk, whether like humans or like AIs, would yield deeper insights into the understanding of human–AI dialogue. Such an understanding of perception is significant, as their preconceived notions and beliefs affect the unfolding interaction. Further research is needed insofar as the shift, indicative of growing reliance on AI in customer service, is projected to see substantial investment growth, underscoring the pivotal role of human-AI communication.