Cobotic service teams and power dynamics: Understanding and mitigating unintended consequences of human-robot collaboration in healthcare services

Robots can be defined in multiple ways (Table 1 provides multiple illustrative definitions). Wirtz et al. (2018 p. 909) provide a nuanced conceptualization of ‘service robots’, which they define as “system-based autonomous and adaptable interfaces that interact, communicate and deliver service to an organization’s customers.” The authors also consider the role of multiple important robot design attributes and characteristics, such as human-like physical embodiment (vs. virtual representation). In line with current real-world examples in services (e.g., robots such as Pepper and Nao), we study embodied, humanoid robots; that is, devices housed in mobile, human-like bodies, which operate via powerful software that enables them to perform in a rational, seemingly human way (Broadbent et al., 2009; Duffy, 2003). In our General Discussion, we will return to the importance of robot characteristics and design features as well as additional specific roles of robots (e.g., companion robots) as a platform for future marketing research (Table 4). For broader reviews of the corresponding robotics literature in marketing and beyond, see: Henschel et al. (2021), Lu et al. (2020), and Bartneck et al. (2020). For a recent review on how human and robot personalities affect human-robot-interactions, see Lionel et al. (2020).

For example, in 2018, a hospital in Singapore began using a Nao robot to guide seniors with dementia in rehabilitation sessions and recovery therapy (Mosaic 2021).

The field of cobotics is concerned with the design and evaluation of robotic systems built to collaborate with humans (Moulières-Seban et al., 2016; Peshkin & Colgate, 1999). In their review of human-robot-collaboration, Bauer et al. (2008, p. 47) note that “humans and robots collaborating on a common task form a team,” and define collaborating as working with someone to reach a common goal. Accordingly, drawing on research in robotics (e.g., Bauer et al., 2008), we study cobotic service teams that include a human and a robot, collaborating in a co-located space, to serve a customer (also see Table 1). More complex configurations of cobotic teams are possible (i.e., multiple humans x multiple robots x multiple customers), but for this initial empirical investigation, we seek to establish the predicted effects in a comparatively simple setting, the triad of one employee, robot, and customer.

Notably, this is not the first case of robots being used in roles of authority. For example, during the coronavirus pandemic, Singapore employed robotic dogs “to enforce strict social distancing” in public (Chen, 2023).

To explore this notion empirically, we conducted a study (N = 80, MTurk, M_Age = 38.05, 37 females) to examine consumers’ perceptions of doctors and assistants. Participants rated the relative extent to which power, autonomy, warmth, and competence (using the same scales as in the main studies) are usually exhibited by a doctor or an assistant in a typical medical visit (1 = “doctor has this trait” to 7 = “assistant has this trait”). The results of t-tests relative to the scale midpoint (4.0) revealed that participants more typically perceive a doctor (relative to the assistant) as more powerful (M = 2.70; t(79) = -6.91, p < .001), more autonomous (M = 2.91; t(79) = -5.89, p < .001), more competent (M = 3.61; t(79) = -2.36, p = .02) but less warm (M = 4.45; t(79) = 3.50, p < .001). These results are consistent with the notion that patient–doctor encounters are characterized by the doctor being perceived as having more power in the encounter as they provide professional diagnoses and treatment (Calnan, 1984; Schei, 2006).

For the remaining studies (3–5) we develop the corresponding hypotheses as the studies are presented.

These exercise classes are designed be relatively standardized, considering the needs of the elderly participants and their (physical) abilities. Moreover, the exercise class occurred weekly, and respondents had participated in prior classes with a human-only instructor (i.e., neither the class nor the exercises were new to the patients).

We calculated the alpha for the first wave (exercise class led by a human) and the second wave (exercise class led by a human-robot team) separately. The small sample size may account for the differences in alphas.

Our results remain the same when performing a nonparametric Wilcoxon signed-rank test (favorability Z = -2.80, p = 0.005; behavioral intentions Z = -2.04, p = 0.042).

Further discussions of analyses related to the role of autonomy, warmth, and competence for this study and our other studies are provided in Web Appendix E. We explored these variables because an alternative perspective in robotics examines the extent to which humans perceive robots as social actors (Dautenhahn, 1999) and, therefore, judge robots along social perception dimensions, such as those that humans use to assess other humans (van Doorn et al., 2017). Considering this lens, some of our studies included measures of a robot’s warmth and competence, which represent the two fundamental dimensions of social perception (see Cuddy et al., 2008). We present these exploratory results in Web Appendix E as a potential inspiration for future work on cobotics.

For completeness, for Study 1, we also examined the mediation path doctor type anxiety behavioral intentions (Hayes, 2015, Model 4, 5000 resamples). For Study 1, the analysis reveals a total indirect effect (a × b = -1.0556, 95% CI: [-1.5263, -.6344]).

Preregistration is available at https://​osf.​io/​jqzph/​?​view_​only=​c6de335a4921470b​8d199e44d821fdcc​

In a mediation analysis (Hayes, 2015, Model 4, 5000 resamples), we use doctor type as the (multi-categorical) independent variable, anxiety as the mediator, and behavioral intentions as the dependent variable. The indirect effect of the human-led versus the robot-led cobotic team though anxiety was significant (a × b = -1.2298, 95% CI: [-1.6675, -.8192]), consistent with Study 1. The indirect effect of human-only team versus robot-led cobotic team was significant (a × b = -1.0430, 95% CI: [-1.4300, -.6711]). The indirect effect of the human-only team versus human-led cobotic team was not significant (a × b = .1868, 95% CI: [-.1373, .5251]).

For completeness, a mediation analysis (Hayes, 2015, Model 7, 5000 resamples) that included doctor type as the IV, choice as the moderator, anxiety as the mediator, behavioral intentions as the dependent variable, and pandemic concerns and gender as control variables revealed significant moderated mediation (a × b = .6564, 95% CI:[.1270, 1.2513]). Anxiety mediated the effect of doctor type on behavioral intentions in the control condition (a × b = -.5694, 95% CI:[-1.0593, -.1281]), as in previous studies, but not in the choice condition (a × b = .0871, 95% CI:[-.2944, .4470]). Anxiety also mediates the effect of choice on behavioral intentions when the doctor was a robot (a × b = .4551, 95% CI:[.0079, .9222]) but not when the doctor was a human (a × b = -.2100, 95% CI:[-.5681, .1381]).

We conduct a mediation analysis (Hayes, 2015, Model 4, 5000 resamples), with doctor type as the independent variable, anxiety as mediator, and behavioral intentions as the dependent variable. We confirm the mediation path (a × b = -.5427, 95% CI: [-.9239, -.1905]).

However, under some circumstances, customers with high power distance beliefs might feel superior over technology and therefore might be less likely to accept a robot physician; future research could examine this alternative idea (we thank one of the five reviewers for raising this point).

We requested N = 150 MTurk participants, but we received N = 153. Prior to exposure to the main study, 12 participants failed an attention screener and were directed to the end of the study without participating, leaving a final sample of N = 141.

For completeness, a mediation analysis (Hayes, 2015, Model 4, 5000 resamples) that included team leader as the independent variable, anxiety as the mediator, behavioral intentions as the dependent variable, and pandemic concerns and gender as control variables revealed significant mediation when examining the human and robot (a × b = -1.1221, 95% CI: [-1.6039, -.6665]) and human and robot with performance information team leader conditions (a × b = -.2330, 95% CI [-.4066, -.0791]). As expected, the mediation for the robot and robot with performance information team leader conditions was not significant (a × b = .4069, 95% CI: [.0205, .8121]).