Service encounters with virtual agents: an examination of perceived humanness as a source of customer satisfaction

2.1 Perceived humanness

The point of departure here is the notion of humanness – the extent to which an individual has characteristics that are typical for humans. This has been referred to as the “human nature” aspect of humanness (Haslam, 2006; Haslam et al., 2008a), and it reflects a view of humanness as a variable rather than a dichotomy. With this view, then, one specific human individual may be seen as having “more” of humanness than another human individual. Previous research with perceived humanness as a social perception dimension has shown, for example, that we humans have a tendency to attribute more humanness to the self than to others (Haslam and Bain, 2007) and less humanness to out-group members than to in-group members (Epley et al., 2013; Leyens et al., 2000). Perceptions of others as having less humanness (i.e. dehumanization) is indeed likely to be a main rationale behind a long list of horrible acts in the history of mankind (cf. Bastian et al., 2012; Epley, 2018).

The variable nature of humanness also implies that non-humans can be seen as having various degrees of humanness. For example, we humans do anthropomorphize animals, but we do not do this to the same extent for all animals; monkeys and dogs are seen as having more mind than chickens and fish (Bastian et al., 2012). In any event, humanness as a dimension of social perception allows for the possibility that it may not distinguish humans from non-humans in a discrete way (Haslam et al., 2005). Or, as noted by Epley and Waytz (2010), ordinary perceivers are capable of treating their pets as people and their enemies as animals. Empirical results stressing the lack of a clear human vs non-human dichotomy appear in Longoni et al. (2019), a study of medical services showing that human service providers were not ascribed different levels of humanness than computerized service providers.

In the present study, the authors use perceived humanness as an overall judgment made by a perceiver in relation to a specific stimulus individual that is not human (i.e. a VA). Previous research, such as Kim and Sundar (2012) and Powers and Kiesler (2006), shows that humans can indeed attribute humanness at an overall level to non-humans. One main reason is that we humans are equipped with evolution-based social responses in relation to other humans – responses that we apply more or less automatically in interaction situations resembling the situations in which they were originally developed (Nowak and Biocca, 2003; Shank, 2013). In other words, we humans are preconditioned, primed and neurologically prepared to interact with humans, and reaction patterns from human–human interactions easily translate into similar reactions to non-human objects (Karr-Wisniewski and Prietula, 2010).

Epley (2018) argues that similarity between the non-human object and a human is a particularly important factor that can trigger this translation; because of the associative nature of human brains, exposure to a non-human object that is similar in some ways to a real human can make accessible and activate mental content (and its affective charge) related to real humans – and in the next step, this content is applied, more or less automatically, to the non-human object [a similar priming-based argument appears in Aggarwal and McGill (2007)]. This provides us humans with efficient information processing possibilities; we can capitalize on our knowledge about what it means to be human. And such knowledge is much more extensive – and easier to access – than our knowledge about what it is like to be non-human (Wiese et al., 2017).

Moreover, it has been argued that needs for social relations, understanding and control are additional explanations behind the tendency to perceive humanlike characteristics in non-humans. That is to say, we humans have such a strong need for social relations that we may make attempts to gain a sense of social connection by humanizing non-human objects (Aggarwal and McGill, 2007; Epley et al., 2008; Epley, 2018). We humans also have a strong need for understanding and control. As we are the most social of all primates (Epley, 2018), personal experience has equipped most of us with well-rehearsed schemes related to ourselves and other humans, and accessing them increases our ability to make sense of, and control, non-human objects (Aggarwal and McGill, 2007; Epley et al., 2007; Epley et al., 2013).

2.2 Impact of perceived humanness on customer satisfaction

A basic tenet in the present study is that the perceived humanness of an individual can have a positive impact on the overall evaluation of this individual. One first reason is person positivity bias. That is to say, for us humans, other humans typically have a positive rather than a negative charge (Sears, 1983). This bias is most likely a function of the inherently social nature of humans, meaning that other humans offer promises of social connection, belongingness and intimacy, which in turn are highly valued outcomes for most humans (Söderlund, 2016). In addition, social connections make us both more happy and healthier (Epley, 2018), and accumulated experience (personal and mass-mediated) of such outcomes may contribute to “what is humanness is good” inferencing. It has also been argued that we humans indeed need other humans for both practical and existential issues (Epley et al., 2008). This means that it makes sense, from an evolutionary point of view, to equip humans with an innate liking for other humans. In any event, a positivity bias of this type has materialized in consumer settings in such a way that the mere presence of a human employee in a store environment can boost customers’ positive emotions (Söderlund, 2016).

Second, as already indicated, humans have well-rehearsed schemes, based on a lifetime of experience, to make sense of humans. The efficiency with which such schemes are used is sometimes astonishing; for example, humans need only few visual clues to conclude that an object is a human (Johansson, 1973). And only thin slice of a person’s behavior is enough to produce remarkably accurate judgments about the person (Ambady and Rosenthal, 1992). Thus information about other humans is often subject to highly fluent information processing, which means that anthropomorphizing a non-human object can result in effortless sense-making (Rauschnabel and Ahuvia, 2014). This, in turn, is likely to influence evaluations; a high level of processing fluency in relation to an object can boost evaluations of the object (Reber et al., 1998). Similarly, it has been suggested that a non-human object’s perceived congruity with the schema for what is human provides a sense of satisfaction that may carry over, in a valence-congruent way, to evaluations of the object (Aggarwal and McGill, 2007; Chandler and Schwarz, 2010).

Third, given that a human is likely to perceive himself or herself as more similar to another human than to a non-human, and given that perceived similarity typically has a positive influence on evaluations (Cialdini, 2007), perceived similarity in relation to a non-human agent can enhance the evaluation of this agent. In addition, perceptions of an agent’s humanness may boost trust in the agent (Castelo et al., 2019; Hadi, 2019), and trust is typically influencing overall evaluations such as customer satisfaction (Anderson and Narus, 1990).

Thus, several mechanisms can contribute to boosting evaluations of an anthropomorphized non-human. It must be underscored, however, that it is humanness in general that is assumed to have a positive charge. Specific human individuals, or specific groups of humans (e.g. criminals, drug addicts and the homeless) do not necessarily have a positive charge, so ascribing a non-human characteristics of a specific (and negatively charged) human is not likely to boost evaluations of the non-human (cf. Aggarwal and McGill, 2007). In any event, and in terms of findings in previous studies, a positive association between the extent to which a human person is imbued with humanness and evaluations of the person has been found by Kozak et al. (2006) and Söderlund (2020). Moreover, in consumer contexts, several studies have identified a positive influence of anthropomorphizing an offer on evaluations of the offer (Aggarwal and McGill, 2007; Delbaere et al., 2011; Rauschnabel and Ahuvia, 2014; Van den Hende and Mugge, 2014). It has also been shown that the presence of a (humanlike) avatar on a website boosts customer satisfaction compared to no such avatar (Holzwarth et al., 2006). In the light of the arguments above and existing empirical findings, then, and in a service encounter, one would expect a positive (and linear) association between the perceived humanness of a VA and the evaluation of the VA (in terms of customer satisfaction).

Some authors, however, have expressed concerns regarding the possibility of a non-linear association between a non-human’s humanlike appearance and the evaluation of this non-human. This possibility has been referred to as the uncanny valley hypothesis, which in the typical case is applied to the human observer’s reactions to non-human objects such as robots. More specifically, according to this hypothesis, increases in humanlike appearance leads to increased liking up to a point after which a too humanlike appearance becomes unnerving and results in reduced liking. It is this dip that is referred to as the uncanny valley. When the humanlike appearance increases further, however, liking is expected to increase again (Cheetham et al., 2011; Gray and Wegner, 2012; Ho and MacDorman, 2010; Kätsyri et al., 2015). Several reasons why there would be an uncanny valley effect have been suggested. For example, humanlike robots may remind us of death and prompt us to see a mind in robots – and both these aspects can be eerie (Gray and Wegner, 2012). It is also possible that ambiguity during categorization of the non-human may result in eeriness (Wiese et al., 2017). In any event, the uncanny valley hypothesis entails the possibility that the association between perceived humanness of a VA and customer satisfaction produced by the VA may be nonlinear. However, empirical findings regarding this hypothesis have been inconsistent (Cheetham et al., 2011; Gray and Wegner, 2012; Kätsyri et al., 2015), or have comprised other dependent variables such as customer satisfaction (Castelo, 2019), so in the present study, the hypothesis regarding perceived humanness of a VA and customer satisfaction is formulated in linear terms (but in the subsequent analysis, the possibility of a non-linear association is examined). In a service encounter setting in which the customer interacts with a VA, then, and when customer satisfaction is conceptualized as an overall evaluation of the VA, the following is hypothesized:

2.3 Human capabilities as antecedents to perceived humanness

Studies and theories comprising humanness have distinguished several specific facets of humanness, and a set of them – specific capabilities of humans – are discussed in this section (agency, emotionality and morality). They were selected because they appear frequently in theories and studies of humanness and anthropomorphizing of non-humans, and it can be assumed that a perceiver can attribute each of them to a non-human agent (such as a VA). In any event, for each of them, it is hypothesized that there is a positive association between the extent to which a VA is perceived to have the specific capability and the perceived humanness of the VA. It is also hypothesized that perceived agency and emotionality of a VA are likely to boost perceptions of the VA’s morality.

First, agency is often mentioned as a humanness aspect (Castelo, 2019; Haslam et al., 2008a; Morera et al., 2018). To be an agent is to be capable of influencing intentionally one’s functioning and life circumstances; it is agency that enables us humans to transcend the dictates of our immediate environment and it allows us to shape the course of our lives (Bandura, 2006). Agency also has social utility (Wojciszke et al., 2009), which makes it valuable for highly social organisms. For example, agentic qualities are instrumental for gaining status and success (ibid.). The main properties of agency are forming intentions that include action plans and strategies for realizing them; setting goals and anticipating the likely outcomes of actions to guide and motivate goal attainment; constructing appropriate courses of action and to motivate and regulate their executions; and the ability to reflect on personal efficacy (Bandura, 2006).

It is assumed here that humans are hardwired to make sense of others in agency terms, because this capability has been identified among 12- to 15-month-old infants (Johnson, 2003). And already at the age of 2.5 years, human children perform better than chimpanzees and orangutans in understanding an actor’s intentions (Herrmann et al., 2007). The capability of attributing agency to others, sometimes referred to as mentalizing and theory of mind (Martini et al., 2016), provides several advantages: it can be used to predict and explain others’ behavior (Johnson, 2003; Kozak et al., 2006; Wiese et al., 2012) and to manipulate (and deceive) others (Johnson, 2003). The exact mechanism for how we humans make sense of others’ agency, however, is an open issue; this capability appears to stem from a combination of using the self as a source of analogy and theory-driven inferences (Epley and Waytz, 2010).

Given the general tendency to attribute humanness to non-human objects, it is not surprising that several studies indicate that we humans can attribute agency also to non-humans (Bastian et al., 2012; Gray and Wegner, 2012; Martini et al., 2016). A classical study is Heider and Simmel (1944), who found that humans can attribute intentions even to geometrical shapes. Moreover, as the attribution of agency to a non-human object is likely to foster anthropomorphizing in relation to that object (Chandler and Schwarz, 2010), the following is hypothesized in a setting comprising customers’ service encounters with VAs:

Emotionality, the capability to experience emotions, is another fundamental aspect of humanness (Castelo, 2019; Epley et al., 2013; Epley, 2018; Haslam and Bain, 2007; Haslam et al., 2008a, Morera et al., 2018). Conversely, when others (such as outgroup members) are perceived to lack emotionality, they are typically denied humanness (Leyens et al., 2000). One main reason why emotionality is crucial for us humans is that it facilitates choosing between alternatives; emotion is the currency that underlies all of our decisions (Johnston, 1999). That is to say, to be able to feel negative or positive emotions has strong implications for action, given the important role of emotions for evaluations of actions (Bagozzi et al., 1999).

The capability to recognize the emotionality of a target person is another aspect of having theory of mind (Epley, 2018; Gray et al., 2012; Leslie et al., 2004) and a first step toward empathy or empathizing (Wiese et al., 2017). One main source for recognizing others’ emotionality is that others frequently display various levels of specific emotions, particularly with facial expressions. Humans’ high levels of emotion detecting skills in identifying others’ displayed emotions (Batty and Taylor, 2003) indicate that we humans are indeed able to attribute emotionality per se to others. In any event, given the general tendency to anthropomorphize non-humans, humans have been shown to be able to attribute emotionality to non-humans (Gray and Wegner, 2012; Haslam et al., 2008b; Martini et al., 2016), including computers (Johnson et al., 2006), smartphones (Wang, 2017) and robots (Wang and Krumhuber, 2018). Therefore, given that emotionality is a central aspect of humanness, the following is expected in a service encounter setting in which customers interact with VAs:

Morality is the third humanness facet to be included in the present study. It is, at a general level, the capability to distinguish what is right from what is wrong and trying to do what is right (Gray et al., 2007; Gray et al., 2012; Waytz et al., 2010). And what is right, in the light of research on moral judgments and human morality, as well as in the folk notion of morality, is typically about what does not harm others and what is just in relation to others (Graham et al., 2011). Some researchers, such as Mikhail (2007), view morality as an innate human faculty (i.e. largely pre-determined by the inherent structure of the mind), and it has been argued that its basic function is to regulate selfishness and to make social life possible (Haidt, 2008). It has also been argued that the justice aspect of morality, visible already among hunter-gatherers, is a particularly decisive aspect to foster a cooperative mode and to form long-term groups (Flannery and Marcus, 2012). Morality, then, seems highly qualified as a central facet of humanness. Empirical results pointing in this direction are provided by Goodwin et al., 2014). Its centrality is reflected in a social perception context, because it has been shown that perceptions of another person’s morality occupy a particularly privileged position for us humans when it comes to forming global evaluative impressions of this person (Brambilla and Leach, 2014; Goodwin et al., 2014).

Given again the tendency to anthropomorphize non-humans, it is expected that non-humans can be subject to morality attributions. Prior research offers support for this in terms that are relevant for VAs; humans can attribute fairness to both computers (Shank, 2012) and algorithms (Lee, 2018), and a disembodied artificial intelligence (AI) can be blamed for its actions (Malle et al., 2019). Therefore, the following is hypothesized regarding VA morality and perceived humanness in a service encounter context:

2.4 Inter-capability associations

So far, it has been hypothesized that attributions of agency, emotionality and morality to a VA are likely to boost perceptions of VA humanness. However, from a perceiver’s point of view, the three capabilities are likely to be interrelated. In the present study, given that agency and emotionality are two central aspects of having a mind, and given that attributions of mind to a target person can turn this person into a moral agent (Epley et al., 2008), it is assumed that attributions of agency and emotionality would serve as clues that contribute to attributions of morality.

In human life, agency–morality associations manifest themselves by a frequent use of agency signals to determine the morality of target persons. In the folk notion of morality, for example, agency is typically one determinant of what is morally wrong (Gray et al., 2012). Indeed, people consistently blame intentional norm violators more severely than unintentional ones (Malle et al., 2014). An agency–morality association can also be seen when people are asked to think about acts that are morally wrong: frequent answers are murder, stealing and adultery (Schein and Gray, 2015), and such acts require a perpetrator with agency. Indeed, many justice systems reflect beliefs that punishments of moral violations should be a function of the (perceived) intent of the transgressor. However, agency can also signal high morality, in the sense that many markers of high morality persons require agency; being fair, brave, self-controlled, principled, responsible, just, honest, sincere, determined, committed, disciplined and loyal are examples (Brambilla and Leach, 2014; Goodwin, 2015; Goodwin et al., 2014; Walker and Hennig, 2004).

In the light of this, and given the tendency to anthropomorphize non-humans, it is assumed in the present study that agency-morality associations in human life can transfer to a situation in which customers are exposed to a VA so that attribution of agency to a VA boosts attributions of VA morality. Hence the following is hypothesized:

Moreover, again in human life, affect is assumed to profoundly shape morality (Harenski and Kiehl, 2011; Sherman and Haidt, 2011). Indeed, emotionality can provide clues about morality in several ways. First, emotionality is a useful facilitator for making right vs wrong distinctions, because emotions provide the decision-maker with a valenced charge for various options. In other words, emotions are can inform morality assessments (Haidt, 2008; Horberg et al., 2011). Assuming that most people have experienced that emotions are associated with morality aspects when they make decisions and judgments, and assuming that the self is an important point of reference when inferences are made about others’ minds, particularly when these others are similar to us (Waytz et al., 2010), one’s own experience of the influence of emotions on morality decisions and judgments can suggest that emotionality influences morality also for others. Second, particularly when transgressions occur, the emotional response by the transgressor is likely to be informative for a perceiver who is making judgments about the transgressor’s morality – and so-called moral emotions (Moll et al., 2007) are particularly informative. A transgressor who does not appear to feel, for example, shame and guilt is likely to be subject to attenuated morality judgments. More generally, lack of emotionality, or distorted emotionality, which has been ascribed to psychopaths and sex-offenders, may explain why they can be perceived as highly immoral (Harenski and Kiehl, 2011). Third, emotionality may be a prerequisite for understanding that others have emotionality, too. And a target person who signals that he or she understands that others can experience emotions provides material for perceivers’ assessment of the person’s morality. For example, a harmful activity carried out by a target person who not only had the intent to carry out the activity, but also understood that others would suffer from it, is typically considered immoral (Gray et al., 2012). A target person’s emotionality, however, can also signal high morality. Markers of a person’s high moral standing derived from naturalistic conceptions comprise several characteristics that seem to require that the person has emotionality; examples are being compassionate, caring and kind (Goodwin, 2015; Walker and Hennig, 2004). In addition, altruism and generosity are other markers of a person’s elevated moral status (Bai, 2017) and they seem to require emotionality, too.

Given this, authors who are convinced that artificial and mechanical agents (such as robots) do not have emotionality typically conclude that they are non-moral agents (Coeckelbergh, 2010). In the present study, however, which is concerned with social perception, and given again humans’ tendency to anthropomorphize non-humans, it is expected that emotionality–morality associations in human life can transfer to a situation in which customers are exposed to a VA. The following, then, is hypothesized:

2.5 Mediation issues

The reasoning above involves several assumptions of mediation. First, given that agency, emotionality and morality are assumed to influence perceived humanness (i.e. H2–H4), which in turn is assumed to influence customer satisfaction (H1), perceived humanness is seen as mediating the influence of the three capabilities on customer satisfaction. To assess this explicitly, the following is hypothesized:

Second, the assumption that agency and emotionality can be seen as antecedents to morality (i.e. H5 and H6), which in turn is assumed to influence perceived humanness (H4), implies that morality mediates the influence of agency and emotionality on perceived humanness. To explicitly assess this, the following is hypothesized:

2.6 Moderation issues

With respect to moderator variables, it has been argued that extant research on the effects of anthropomorphism in consumer settings often overlooks the role of consumption context (Wan, 2018). Nevertheless, previous research indicates that humans prefer a human service provider rather than a computerized service provider in situations when the motives for seeking service are highly involving (Longoni et al., 2019). Given a view of humanness as a variable, and given that a human is capable of having more humanness than a non-human, Longoni et al.’s (2019) results suggest that task involvement may moderate also associations in which the perceived humanness of a service provider is the independent variable. Results that seem to point in this direction have been provided by Holzwarth et al. (2006); the presence of an (humanlike) avatar produced a higher level of satisfaction for high involvement visitors to a website than for low involvement visitors.

More specifically, in the present study, it is expected that the strength of the hypothesized association between the perceived humanness of a service provider and customer satisfaction (i.e. H1) can be moderated by task importance in such a way that high task importance results in that customers give more weight to the perceived humanness of a service provider when they form overall evaluations. The main reason is that high task importance is a situation in which the customer needs more advanced information, and needs to process it more carefully (Zaichkowsky, 1985), to arrive at a satisfying decision. The authors of the present study assume that the extent to which such needs can be satisfied is a function of the perceived humanness of a service provider. That is to say, given that effective transferring of information requires at least some theory of mind (Epley and Waytz, 2010), it is assumed that satisfying a receiver’s need for more detailed and personalized information would demand more theory of mind (and thus more humanness) from the sender. Conversely, when task importance is low, and thus the customer’s need for information is not high, the level of the humanness of the service provider is likely to be relatively less relevant from an information exchange point of view – and therefore less influential in the process in which overall customer satisfaction is formed. It should also be noted that assessing other people’s minds is cognitively taxing (Epley and Waytz, 2010), and low task involvement is likely to be less motivating to do so than high task involvement. The following, then, is hypothesized for customers who are interacting with VAs in service encounters:

Taken together, then, there are nine hypotheses to be tested in the present study and an overview of them is provided in Figure 1.

3.1 Data collection and participants

As already indicated, one main idea in the present study was that it should comprise customers’ experience of interacting with existing VAs and thus that the study should capture natural VA variation that exists in the field. Therefore, the authors invited participants who had interacted with a VA, defined as “a computer-generated character, sometimes powered by artificial intelligence, which provides customer service.”

The participants were asked to think about one specific and memorable service encounter in the past when they – as consumers – had been interacting with a VA that was representing a firm. This part, then, is somewhat similar to studies based on critical incidents (cf. Bitner et al., 1990) in the sense that it generates one particular incident (here: an encounter with a VA) for each participant. In the present study, however, the participants were not asked to think about a critical incident (e.g. an encounter that generated a very low or a very high level of satisfaction), because the idea was to allow for full variation in satisfaction. Then, in the second part, and with open-ended question, the participants were asked to specify which type of VA they had interacted with, which type of firm it represented and for what task they interacted with the VA. They were also asked to describe what happened in the interaction. This part mainly served the purpose of making the selected encounter come alive again in the participants’ minds and to allow for descriptive data regarding the VAs. The third part, for which the participants were asked to have the selected interaction in mind, comprised standardized questionnaire items to measure the variables in the hypothesis. That is to say, each participant selected one VA encounter from his or her life as a customer, all participants responded to (the same) standardized questions (presented below) about a selected encounter, and the responses to the standardized questions were used for testing the hypotheses.

The data were collected online from members of the Prolific panel, a platform explicitly tailored for researchers who collect data online (cf. Palan and Schitter, 2018). From this panel, 225 members who were UK residents and who had experience in online shopping were invited, and those who passed two attention checks were used in the analysis. The first attention check was the item “Please select the number 8 below” (with 5, 8, 13 and 15 as response alternatives); the second check was the item “Please select the number 14 below” (with 5, 8, 12 and 14 as the response alternatives). This approach was inspired by Jones et al. (2015), who refer to such items as “trap questions.” After removing those who failed the attention checks, the result was 209 participants (M_age = 37.53; 149 women and 60 men) to be used for testing the hypotheses.

3.2 Virtual agents in the sample

As for the VAs selected by the participants, the majority were chatbots encountered online on firms’ websites. In addition, some VAs were encountered in telephone calls to firms and some appeared in physical environments such as airports, train stations and clothes stores. The firms that these VAs represented comprised banks, health-care providers, automobile companies, supermarkets, energy providers, clothing retailers, airlines, insurance firms, restaurants, travel agents and telecom firms. And the tasks that were covered in the VA interactions had to do with, for example, seeking information, booking tickets, cancelling orders, checking accounts and order tracking. With respect to physical characteristics of the 209 VAs in the sample (and as reported by the participants when asked about them), 53% had a name (e.g. Alice, Andy, John, Sara and Sally), 55% had a gender and 24% were represented with a face.

3.3 Assessment of the measures

In a first step, the properties of the measures were assessed with a structural equation modeling approach (SmartPLS 3.0 was used). The items used for the multi-item measures are presented in Table 1, in which also the reliability and validity of the measures are reported in terms of Cronbach’s alpha (CA), composite reliability (CR) and average variance extracted (AVE). For a discriminant validity assessment, the heterotrait–monotrait ratio of correlation was used with 0.85 as a threshold, and all ratios in this assessment were <0.85. All items were scored on a scale ranging from 1 to 10. For all items except the satisfaction items, the endpoints were labeled as 1 = do not agree at all and 10 = agree completely. The endpoint labels for the satisfaction items are reported below.

As for the participants’ perceptions of VA capabilities (i.e. the antecedents to perceived humanness; H2–H4), it should be noted that the focus in the present study is on a first step in human social sense-making of others, namely, the extent to which another individual (here: a VA) is perceived to have the capability per se (cf. Martini et al., 2016; Wiese et al., 2012; Wiese et al., 2017), not on perceptions of levels of particular expressions of the capability.

Agency was measured with items similar to what was used in Gray and Wegner (2012), Kozak et al. (2006), Morera et al. (2018) and Waytz et al. (2010). For emotionality, similar items have been used by Kozak et al. (2006) in a mind attribution scale, by Wang (2017) in an anthropomorphism scale and by Wang and Krumhuber (2018) in a study of the extent to which emotionality was attributed to robots. It may be noted that Klein and Hodges (2001) argue that women are more motivated than men to perform better at tasks involving empathic accuracy (i.e. inferring the feelings of another person). This suggests that they may be more motivated to infer emotionality, too. The data in the present study do indicate this; female participants attributed a higher level of emotionality (M = 3.93) to the VAs than did male participants (M = 2.89). This difference was significant (t = 2.49, p < 0.05). This, then, serves as a validity indicator for the emotionality measure. As for morality, Gray et al. (2007), Gray et al. (2012) and Morera et al. (2018) have used similar items. Perceived humanness items of the type used in the present study have been used previously by, for example, Aggarwal and McGill (2007), Söderlund (2020) and Thompson et al. (2011).

For customer satisfaction, the three Fornell (1992) items were used: “How dissatisfied or satisfied are you with the virtual agent?” (1 = very dissatisfied and 10 = very satisfied), “To what extent did the virtual agent meet your expectations?” (1 = not at all and 10 = totally) and “Imagine a virtual agent that is perfect in every respect. How near or far from this ideal did you find the virtual agent? (1 = very far from and 10 = cannot get any closer). To assess the validity of the customer satisfaction measure, the responses to the open-ended item used for asking the participants to describe what happened in the interaction were used. More specifically, the Linguistic Inquiry and Word Count (LIWC) software was used to compute the emotional tone of the texts provided by the participants [for a similar approach, see Cavanaugh et al. (2015)]. LIWC produces an emotional tone score that takes on values in the range from 0 (very negative) to 100 (very positive). Given that customer satisfaction is expected to be positively associated with positive affect (Mano and Oliver, 1993), it was assumed that a positive association between the two variables would indicate validity in the satisfaction measure. The two variables were indeed positively and significantly associated (r = 0.48, p < 0.01). Moreover, customer satisfaction is typically assumed to have a positive impact on repatronize intentions. As an additional validity check, then, the question “How likely is it that you would interact with the same virtual agent again?” (1 = very unlikely and 10 = very unlikely) was included. The correlation between the satisfaction variable and this intention variable was positive and significant (r = 0.72, p < 0.01), which provides further evidence of the satisfaction measure’s validity.

For task importance (i.e. the potential moderator involved in H7), the participants were asked to describe, in their own words, for what task they interacted with the selected VA. After this, the following question appeared: “How important was it for you to complete this task?” (1 = not important at all and 10 = very important). The responses to the itemized rating scale was used as a measure of task importance (M = 8.25, SD = 1.77).

To be able to test the uncanny valley hypothesis, as an extension of H1, the authors of the present study used a measure of VA uncanniness comprising the participants’ feelings about the selected VA interaction in terms of a ten-point scale and the adjectives “uncomfortable,” “uneasy” and “eerie” (CA = 0.93; M = 3.10, SD = 2.31). Similar adjectives to capture the uncanniness of a robot were used by Gray and Wegner (2012) and Thompson et al. (2011).

Finally, as already indicated, the VA descriptive data comprised (dichotomous) data about VA features (i.e. if they had a name, gender and face). These features are indeed possible antecedents to perceived humanness, given that non-human similarity with humans is assumed to contribute to anthropomorphization (Epley, 2018). In the present study, however, the three feature variables were only weakly and non-significantly associated with perceived humanness, so the subsequent analysis comprises only agency, emotionality and morality as antecedents to perceived humanness.

Taken together, the measures of the variables in the hypotheses had acceptable levels of reliability and validity, so the next step was to assess the relationships between them. A structural equation modeling approach (with SmartPLS 3.0) was the main tool for this assessment.

4.1 Perceived humanness as a higher-order construct?

It should be underscored that the present study has conceptualized (and measured) perceived humanness as a unidimensional variable, in which the measure is free of content related to specific humanlike capabilities. This was assumed to facilitate contact between literatures in which the perceived humanness of agents with different features and capabilities are studied – such as brands, physical products, VAs, robots and even real humans. However, this is not the only option; several studies have conceptualized (and/or measured) perceived humanness as a higher-order construct comprising several capabilities.

The latter has several advantages; for example, it reduces the number of relationships in a model and it can be a useful option in cases with high collinearity between independent variables (Sarstedt et al., 2019). One disadvantage, however, is that it can provide less precision for interventions. As an example, a practitioner who wants to know if VA agency or emotionality should be emphasized to boost satisfaction would find it less helpful if these two aspects are not separate variables. In any event, before turning to the tests of the hypotheses from the theory section, an examination of what happens if perceived humanness is conceptualized and measured as a higher-order construct in the present study is warranted.

This can be done in many ways, but one particularly common way is to think of a higher-order construct as of the reflective-reflective type (Sarstedt et al., 2019). In the present case, this means that perceived humanness was modeled as reflective of agency (measured with four items), emotionality (measured with three items) and morality (measured with five items), and an antecedent to customer satisfaction (measured with three items). SmartPLS 3.0 was used for the analysis. In terms of overall fit, however, this alternative model (standardized root mean residual [SRMR] = 0.10) was inferior to a model with the hypothesized associations between variables (i.e. H1–H6 and H9) as depicted in Figure 1 (SRMR = 0.05). The alternative model also reduced the explained variance in customer satisfaction (R² = 0.408) in relation to the proposed model (R² = 0.43). Given model fit and criterion-related validity as two ways to substantiate the usefulness of higher-order multidimensional constructs (Johnson et al., 2012), it can be contended that it was less useful to conceptualize perceived humanness as a higher-order construct in the present case. The content of higher-order constructs, however, should be determined by theory (Johnson et al., 2012). Several theories do suggest that agency, emotionality and morality are central human capabilities, but few authors would argue that these are the only capabilities that determine perceived humanness – and to date there is no consensus regarding a complete list of such capabilities. Thus additional capabilities need to be assessed before the final word is said about perceived humanness as a higher-order construct.

4.2 Testing the hypotheses

The hypothesized associations (see Figure 1) were tested with a structural equation modeling approach (SmartPLS 3.0 was used). That is to say, the associations covered by H1–H6 and H9 represented the proposed model, and its overall fit with the data was good (SMRM = 0.05, NFI = 0.902). As for H1–H6, the path coefficients are reported in Table 2.

For H1, the results indicate a positive and significant association between perceived humanness and satisfaction (b = 0.45, p < 0.01). This means that H1 was supported. It should be recalled, however, that several authors have discussed the possibility of an uncanny valley aspect of perceived humanness (i.e. the possibility of a non-linear association between perceived humanness and overall evaluation variables). Therefore, the authors of the present study modeled the association as a third degree polynomial function (Y = a + b₁x + b₂x² + b₃x³), which is capable of capturing the two extreme points to be expected if an uncanny valley relationship is at hand, as depicted by, for example, Ho and MacDorman (2010) and Kätsyri et al. (2015), and compared it with a linear model (Y = a + b₁x). This comparison, made with SPSS 26, however, yielded a well-neigh linear pattern also for the third degree polynomial and a virtually identical level of explained variance (R² = 0.434) in relation to the linear model (R² = 0.428). It should also be noted that the association between perceived humanness and the explicit measure of uncanniness was weak (r = −0.12, p < 0.10). Taken together, then, it must be contended that the perceived humanness of the VAs in the present study did not elicit an uncanny valley effect.

Turning to the hypotheses about the three capabilities and their influence on perceived humanness (i.e. H2–H4), there were significant and positive associations for agency (b = 0.15, p < 0.05), emotionality (b = 0.39, p < 0.01), and for morality (b = 0.33, p < 0.01). This means that H2–H4 were supported. In a human-to-human context, this is perhaps not surprising, because each capability has been discussed in the literature as a fundamental aspect of what it means to be a human. The results of the present study, however, show support for the hypotheses in a human–VA context and the results thereby indicate that not only easily observed features of a VA (such as a face or voice), which have been examined in previous studies, can contribute to the perceived humanness of a VA. It may be noted that agency was subject to a weaker association with perceived humanness than emotionality and morality. Thus, in the present study, agency appeared to be less diagnostic of perceived humanness than the other capabilities. One reason may be that attributions of agency to non-human entities are likely to occur more frequently in everyday life than attributions of emotionality and morality (e.g. in relation to objects such as cars and computers), which can make agency a less distinct characteristic of humanness than the other two capabilities.

H5 and H6 have to do with relations between the three capabilities, and the results indicate that there was a positive and significant association between agency and morality (b = 0.51, p < 0.01) and a positive and significant association between emotionality and morality (b = 0.39, p < 0.01. This means that both H5 and H6 were supported. These results are consonant with the argument that that both agency and emotionality are needed for morality (Gray et al., 2012). This part of the results also indicates that agency was more diagnostic than emotionality for the attributions of morality. One possible reason is that the agency–morality link may be activated more frequently than the emotionality–morality link in everyday life, and this could have made it more accessible in a VA encounter. If so, this seems to reflect that agency has a particularly strong link to moral reasoning, in the sense that, for example, intentional norm-violators are blamed more than those that are unintentional (Malle et al., 2014) and that children, who may appear less autonomous than adults, are held less responsible for transgressions (Bigman et al., 2019).

For the first mediation hypothesis (H7), stating that perceived humanness mediates the influence of agency, emotionality and morality on customer satisfaction, the procedure suggested by Nitzl et al. (2016) and Sarstedt et al. (2020) was followed. That is to say, and again with SmartPLS 3.0, three links were added to the proposed model (i.e. direct links between agency and satisfaction; emotionality and satisfaction; and morality and satisfaction). This was done to be able to control for direct effects and, if the indirect links are significant, to assess the type of mediation. With this approach, as in the perhaps more familiar Hayes PROCESS approach, mediation is at hand if there is a significant indirect effect between an independent variable and a dependent variable. This, in turn, is indicated by the confidence interval for the coefficient for the indirect effect (it should not comprise a zero). Nitzl et al. (2016) recommends a biased-corrected confidence interval for the assessment, and this was used in the present mediation analysis. This analysis showed that there was a significant indirect effect in the following chains of influence: agency–humanness–satisfaction (b = 0.06, p < 0.05), emotionality–humanness–satisfaction (b = 0.15, p < 0.01) and morality–humanness–satisfaction (b = 0.13, p < 0.01). Here, the relatively less diagnostic role of agency for perceived humanness made itself visible again, and this time in terms of a relatively weaker indirect effect of agency compared to the other capabilities. Moreover, the direct impact of agency on satisfaction was not significant (p = 0.19), the direct impact of emotionality on satisfaction was significant (p < 0.01) and the direct impact of morality on satisfaction was not significant (p = 0.85). This, then, indicates that perceived humanness fully mediated the agency–satisfaction association and the morality–satisfaction association, while mediation was complementary in the emotionality–satisfaction association (cf. Zhao et al., 2010). H7 was thus supported.

The second mediation hypothesis (H8) was based on the assumption that agency and emotionality are likely to provide clues about morality and thus they can be seen as antecedents to morality (cf. Section 2.4). To test this, the same extended model as above was used (i.e. a model with direct links between agency and satisfaction; emotionality and satisfaction; and morality and satisfaction) in a mediation analysis with SmartPLS 3.0. In this analysis, there was a significant indirect effect (b = 0.06, p < 0.05) of agency on satisfaction in the chain in which also morality is present as a mediator (i.e. agency–morality–humanness–satisfaction). Similarly, emotionality had a significant indirect effect (b = 0.04, p < 0.05) on satisfaction when also morality served as a mediator (i.e. emotionality–morality–humanness–satisfaction). This, then, provides support for H8 and adds further evidence in support of the argument that morality seems to require both agency and emotionality (Gray et al., 2012).

Finally, and for the moderator hypothesis (H9), an initial inspection of the data revealed that the association between perceived humanness and satisfaction was weaker when task importance was low (r = 0.52) compared to when task importance was high (r = 0.66). This is thus the type of moderation pattern that was hypothesized. To test this explicitly, H7 was first assessed with the moderator analysis module in SmartPLS 3.0, which produced a non-significant result at the 5% level (p = 0.09). Second, the potential for moderation was re-examined with the PROCESS procedure for SPSS version 3.3 and SPSS 26. Hayes (2012) Model 1 with 5,000 bootstrap samples was used (perceived humanness was the independent variable, task importance was the moderator and customer satisfaction was the dependent variable). The interaction term from this analysis (b = 0.013), which was used as an indicator of moderation, was not significant (p = 0.70). Thus both tests point in the same direction, so H9 must be rejected.

5.1 Summary of main results

The results indicate that the more the VAs in service encounters were attributed agency, emotionality and morality, the more they were attributed humanness. Perceived humanness, in turn, was positively associated with VA-generated customer satisfaction (and there were no indications of an uncanny valley effect when perceived humanness increased from low to high).

5.2 Contributions

The present study contributes to the literatures on anthropomorphization and service encounters by showing that VAs in service encounters can be anthropomorphized in terms of fundamental human capabilities (agency, emotionality and morality) that must be inferred, because they cannot be directly observed. It should be noted that the levels (on a ten-point scale) reached by the three capability variables were relatively low (i.e. M_agency = 4.95, M_emotionality = 3.58 and M_morality = 4.69), which indicates that many of the VAs in the present study have not reached an advanced stage. Nevertheless, the capability variables were still able to boost the perceived humanness of the VAs. It should also be noted that the perceived humanness variable reached a relatively low level (M_humanness = 4.79). This may suggest that the VAs in the study were not perceived as very humanlike. It should be recalled, however, that previous research has examined the perceived humanness of real humans, and results from such studies show that real humans, particularly out-group members, but also store employees, rarely receive “full” perceived humanness scores (cf. Söderlund, 2020). One main reason is likely to be that the minds of others are never fully accessible. In any event, in previous research, some groups of real humans appear to have received basically the same level of perceived humanness as the VAs in present study (cf. Table 1 in Kteily and Bruneau, 2017).

The present study also contributes by identifying a positive association between the perceived humanness of a VA and the evaluation of the VA (in terms of customer satisfaction). This indicates that we humans are not only influenced by heuristics of the type “what is beautiful is good” (Dion et al., 1972) and “what is healthy is good” (Pazda et al., 2016) in a social perception setting; we are also influenced by a more fundamental and holistic heuristic of the type “what is humanness is good.” It should be underlined that there was no evidence of a so-called uncanny valley effect in the association between perceived humanness and customer satisfaction. Presumably, there is less potential for eeriness as consumers’ experience in interacting with VAs is increasing – and such experience appears to already have reached a level at which many consumers are familiar with VA interactions (Reeves, 2016). After all, virtual assistants such as Apple’s Siri and Amazon’s Alexa have been around for some years, and an increasing number of consumers are encountering virtual humans in entertainment areas – such as virtual influencers (e.g. Lil Miquela), pop stars (e.g. Hatsune Miju) and supermodels (e.g. Noonouri).

Moreover, as VAs powered by AI are developing fast, and are expected to replace or complement human employees in many roles (Fen, 2019), there is an ongoing discussion of what AI is and is not capable of in relation to humans’ capabilities. Several observers, for example, Fen (2019) and du Sautoy (2019), appear to be convinced that AI-powered entities are not – and will not be – capable of agency and other fundamental human capabilities of the type explored in the present study. Be that as it may, the present study contributes to the general discourse on AI and the human–machine distinction by showing that humans can perceive VAs as having relatively advanced capabilities and that such perceptions add to the overall perceived humanness of a non-human – and to the overall evaluation of VAs. More specifically, given that we humans are expected to spend more time interacting with VAs in the near future, the present study contributes to the AI discourse by highlighting factors that can facilitate human–machine interactions in commercial settings and, presumably, also blur the boundary between humans and machines.

5.3 Managerial implications

In marketing, particularly in offline settings, there are many examples of firms that make attempts to “humanize” their products and brands in the hope of more positive evaluations. For example, brands are imbued with personalities, they are referred to as “she” or “he” instead of “it,” and they are presented as talking directly, in the first person, to the receiver of an ad (Aggarwal and McGill, 2007; Van den Hende and Mugge, 2014). Previous research indicates that customers’ anthropomorphizing in such cases can pay off in terms of positive product attitudes (van den Hende and Mugge, 2014).

As service encounters are becoming increasingly online based, and comprise interactions with various forms of VAs, firms are striving for anthropomorphization also in this setting (Steinhoff et al., 2019). With respect to such online practices, and in the light of the results in the present study, it does appear as if anthropomorphizing VAs in service encounters has a similar positive effect on evaluations as in the case of physical products. Given the VA’s role in a service encounter as a representative of a firm, and given that service encounters typically mean that the representative is the firm from the customer’s point of view (Bitner et al., 1990), the results of the present study imply that customer satisfaction-seeking firms that are using VAs would benefit if the VAs are designed so that they receive high scores in perceived humanness.

More specifically, and in addition to previous research indicating that various VA features (e.g. a name and a humanlike face) can enhance perceived humanness, the present study shows that perceived VA agency, emotionality and morality have a positive influence on perceived VA humanness. To produce VAs with such capabilities would of course be a Herculean task, both for a typical service firms that encounters customers online and for full-time VA designers, yet it should be underscored that the present study has dealt with perceptions of such capabilities. It should also be noticed that many contemporary VAs are visible to the customer only in terms of the text that they produce in a chat window on a screen; a typical encounter with a VA, then, resembles an exchange of SMS messages. These two aspects make life easier for managers who want to boost perceptions of humanlike capabilities of VA, because text can signal important aspects of a sender’s psychological world (Pennebaker et al., 2003). For example, recent research shows that simple linguistic elements in texts sent by VAs can boost their perceived happiness (Söderlund et al., 2021), which would signal that they have emotionality. Similarly, if the mere movement of a geometrical shape signals that it has agency (Heider and Simmel, 1944), one would assume that a VA in a service encounter that uses phrases such as “My goal is […]” and “What we need to do, to find a suitable XYZ for you, is to […]” can signal agency, and that phrases such as “I believe it is right to […]” and “It would not be fair to […]” can signal morality. Not much existing research is available to give specific guidance about what exactly a VA should say to boost perceptions of agency, emotionality and morality, but firms with online service encounters can relatively easily create their own VA guidelines by experiments online with their visitors (cf. Thomke, 2020).

5.4 Limitations and suggestion for further research

Several limitations characterize the present study – and several possible alternative options for the study were not chosen. An overview of options for further research is provided in below:

• Examine different types of VAs with respect to their level of perceived humanness.

• Examine different ways of measuring perceived humanness.

• Examine perceived humanness and technological aspects of VAs with respect to the impact on overall evaluations.

• Examine additional capabilities that may influence perceived humanness.

• Examine additional moderator variables.

• Examine other downstream variables than customer satisfaction, such as trust, perceived hedonic value, money spent, prosocial behaviors and performance in joint action tasks.

With respect to the VAs in the sample, the authors of the present study did not make systematic attempts to compare different types (e.g. VAs that appear on firms’ websites and VAs that exist in commercial environments such as hotel lobbies and airports). There are indications, however, that different types of VAs may be subject to different levels of anthropomorphism. For example, embodied synthetic agents (such as three-dimensional robots) prompt more anthropomorphism than agents represented in two dimensions (Kiesler et al., 2008). Further research should examine this by explicitly comparing different types of VAs.

Moreover, previous studies indicate that it is easy to prime humans who are exposed to a non-human object so that they indeed attribute human characteristics to the object. This calls for caution when researchers examine humanness perceptions. It is not unlikely that merely asking participants to assess the extent to which a non-human object is perceived to be human (e.g. with questionnaire items of the type used in the present study) can set in motion thoughts for which the downstream result is that anthropomorphization indeed occurs. If this happens, it is an example of self-generated validity (Feldman and Lynch, 1988). To avoid this, other (and less obtrusive) measurement approaches may be beneficial. One option may be the “ascent of humans” measure used by Kteily and Bruneau (2017). Another option is to do as Chandler and Schwarz (2010) did; they asked participants to describe their cars with their own words, and the resulting texts were content analyzed (e.g. by noting if a car was described by personal pronouns and with personality traits) to generate a measure of anthropomorphization. Further research is needed to establish the relationship between itemized rating scales for the measurement of perceived humanness and less obtrusive measures.

In addition, although anthropomorphization of non-human objects is a pervasive human tendency, presumably grounded in our extensive experience of ourselves and of other humans (which provides us with easily accessible schemes to be applied to other objects), computers have become ubiquitous. Indeed, some contemporary humans may spend more time with various screen-based devices than with other humans. It has resulted in the emergence of “inverted anthropomorhization”; we humans increasingly describe ourselves and others in terms of our technologies (Johnson et al., 2008). This is clearly also the case when researchers (including the present authors) approach the human brain with terms such as “information processing” and “hardwired.” Thus we humans are able to “technomorphize” humans – and we can do so because we are able to attribute various technical characteristics (e.g. size of the screen and capacity of the hard disc) to our technologies. And such characteristics are likely to influence our overall evaluations. This indicates that the humanness of a VA is unlikely to be the sole predictor of VA-generated customer satisfaction. Further studies should therefore explore if “technomorphizing” of VAs exists and, if it does exist, how it contributes to overall evaluations.

When it comes to antecedents to perceived humanness, the present study included only some capability variables (i.e. agency, emotionality and morality). To expand the knowledge in this area, other capabilities should be examined. It has been suggested, for example, that openness and warmth are relevant capability variables within a humanness framework (Haslam and Bain, 2007). Similarly, the capabilities to have thoughts and to learn (Martini et al., 2016), intelligence (Johnson et al., 2006), exercising control (Johnson et al., 2008), the capability to recognize and understand that there are differences between human individuals (Longoni et al., 2019) and sensitivity to how we are perceived by others (Epley and Waytz, 2010) are aspects of humanness, too. It would also be fruitful to examine more advanced capabilities. For example, it has been argued that one of the critical markers of a fully developed theory of mind is the capability to recognize another person’s false beliefs (Epley and Waytz, 2010), and this appears to be relevant in a service encounter context. Presumably, customers can have many false beliefs that interfere with their consumption satisfaction, and a clever human frontline service employee would be able to address this already in the pre-purchase stage of the consumer’s decision-making process (e.g. when a mountain guide assesses, and corrects, a potential client’s beliefs about the effort and skills needed to reach one particular peak). It should be relatively easy to include this aspect in a study of participants’ perceptions of VAs (e.g. with questionnaire items such as “The VA identified that I was wrong about some aspects related to the offer” and “The VA corrected my beliefs”).

With respect to variables that may moderate the perceived humanness–satisfaction association, the present study examined task importance (yet it did not influence the association). Other task-related variables, however, may indeed serve as moderators. Castelo et al. (2019) report findings suggesting that the type of task can moderate the impact of the humanization of an algorithm and trust in the algorithm. More specifically, under the condition of low algorithm humanization, algorithms produced more trust for objective tasks (i.e. quantifiable and measurable facts exist) than for subjective tasks (i.e. personal opinion and intuition are needed). Under the high humanization condition, however, these differences disappeared, thus suggesting that humanization can be used to build trust when algorithms are used for subjective tasks. Given that trust can produce satisfaction (Anderson and Narus, 1990), further research should explore if the type of task (objective or subjective) moderates also the humanness–satisfaction association. Moreover, Hadi (2019) found that for customers who entered a chatbot interaction with anger, chatbot humanization had a negative impact on satisfaction. This indicates that the customer’s initial emotional state can moderate the influence of VA humanization on customer satisfaction. Presumably, this may reflect an inability of current chatbots to empathize with customers who are in a negative state of mind, which in turn may reduce satisfaction. Given that basic emotions are discrete, in the sense that they feel different and have different antecedents and consequences (Söderlund and Rosengren, 2004), further research is needed to explore if other initial emotions than anger have a moderating potential. Previous research also indicates that customers differ with respect to the extent to which they need or want personal assistance in a service context (McGoldrick et al., 2008), and it seems likely that those who do not want this are less influenced by the humanness of a VA.

Finally, the main dependent variable in the present study was customer satisfaction. However, other outcome variables (i.e. potential consequences of perceived humanness) do exist, particularly with respect to customer behavior in a service encounter before it is completed. For example, it has been argued that a strong form of anthropomorphism entails behaving toward non-human object as if it possesses humanlike traits (Epley et al., 2008). Previous research indicates that perceived humanness/anthropomorphism can affect behaviors that are relevant in a service encounter context, such as prosocial behaviors (less cheating and more generosity) and performance in joint action tasks (Wiese et al., 2017). It can also affect how much money the customer spends (when a slot machine is anthropomorphized; Riva et al., 2015), and it may influence customers’ replacement decisions (Chandler and Schwarz, 2010). Presumably, the perceived humanness of a VA can also make the VA easier to remember (cf. Kiesler et al., 2008) and evoke more trust (Etemad-Sajadi, 2016), and these two aspects are likely to have implications for repatronize decisions. Moreover, the results presented by Holzwarth et al. (2006), showing that the presence of an (humanlike) avatar increases perceptions of entertainment value for website visitors, indicate that the perceived humanness of a non-human can be entertaining. This, in turn, may boost the customer’s experience in terms of the notion of an experience in the literature on experiential offers (Brakus et al., 2009). Outcomes such as these, then, should be examined to add more details to the influence of the perceived humanness of VAs in service encounters.