How locus of causality shapes human-AI decision-making Open Access

Global interest in Artificial Intelligence (AI) has surged, driven by the rapid development of generative AI technology, attracting $25.2 billion since 2023—over a quarter of all AI funding (Stanford Institute for Human-Centered Artificial Intelligence, 2024). Adoption is particularly notable in marketing and sales, where 34% of organizations use generative AI, compared to 23% in product development and 4% in manufacturing (McKinsey, 2024). AI is integrated into marketing for customer-facing tasks like content creation, personalized marketing, and customer contact (Ferraro et al., 2024). Conversational agents increasingly handle customer service interactions, replacing human representatives thanks to their advanced communication capabilities (PwC, 2024; Kot and Leszczyński, 2020). Intelligent agents add the ability to perform actions on behalf of company representatives, helping them to use a service or an ancillary service related to the implementation of the agent. This way, conversational agents have a direct impact on a supplier’s customer relationships (Kot and Leszczyński, 2022).

Replacing humans with AI may streamline processes and reduce costs for a company implementing such a solution. However, it may be negatively perceived by business partners (Ciechanowski et al., 2018) who need to interact with AI instead of humans. This raises certain constraints and may lead to negative consequences related to blurred responsibility in human-AI collaboration. The phenomenon has been defined as the responsibility gap (Hindriks and Veluwenkamp, 2023).

The responsibility gap asks the questions: Who is ultimately responsible for AI’s assistance in human work? How much responsibility can people attribute to an algorithm? When AI systems are integrated into decision-making processes, diffusion of responsibility can strain relationships between organizations. When managers rely on AI to execute tasks humans used to do, understanding the nuances of responsibility attribution is critical to maintaining trust and collaboration (Nabavi et al., 2024). The responsibility gap is seen as one of the most significant challenges in AI today (AI Act, Regulation (EU) 2024/1689), as it significantly impacts human-AI collaborations (Kunz and Wirtz, 2024), may even intensify ethically questionable actions (Bleher and Braun, 2022) and drastically alter long-term business relationships.

So far, experimental research on human-AI interactions and responsibility indicates that evaluating these interactions depends on whether the algorithm’s performance has a positive or negative outcome (Shin, 2020). However, recent studies rooted in attribution theory suggest that people may evaluate the same outcome differently (positive or negative) depending on the attribution of credit or blame. This raises the issue of the responsibility gap for both positive and negative outcomes of AI-supported work and its behavioral consequences.

This study investigates how managers attribute responsibility for successful or unsuccessful outcomes when interacting with AI or humans. We focus on internal (i.e. company representative) or external (i.e. situational or coincidental) factors driving attributions and their influence on satisfaction. The research framework is embedded in attribution theory and collaboration with AI in organizations, focusing on the locus of causality (Weiner, 1986) ^[1]. Perceived locus of causality refers to an individual’s perception of who or what is the leading cause of success or failure. Empirically, this study conducts three experiments with B2B marketing managers buying services from providers (N = 355; recruited on the Prolific platform, Palan and Schitter, 2018). Study 1a was a scenario-based experiment (vignette) with 2 (agent: AI vs. human) × 2 (performance: success vs. failure) design, with locus of causality as the dependent variable. Study 1b was a real-interaction experiment in which managers conversed with a genuine AI agent. Participants evaluated the AI’s performance as either successful or unsuccessful. We then tested how this performance evaluation influenced locus of causality as the dependent variable. Study 2 was a vignette, one-factor design, with anticipated satisfaction as the dependent variable and agent (AI vs. human) as the independent variable. The vignette studies provided control, while the real-interaction study added ecological validity, offering complementary insights into managers’ judgments in AI-assisted B2B interactions.

Our findings contribute to the literature on human responses to AI and managerial decision-making. First, we show that evaluating an agent’s performance depends on the perception of it as an AI, and on responsibility attribution, challenging the algorithm aversion paradigm (Dietvorst et al., 2015). Second, we demonstrate that the responsibility gap applies to both positive and negative outcomes, extending the research focused mainly on failures (Hindriks and Veluwenkamp, 2023). Third, our results suggest that managers can share responsibility with AI, perceiving it as having agency, contributing to the debate on AI as a tool versus as an actor (Nabavi et al., 2024). Fourth, we identify responsibility attribution as a key driver of satisfaction in business interactions, linking it to algorithm aversion/appreciation in business. Finally, our research offers practical guidance for managers on implementing AI in daily operations.

The collaboration between B2B actors and AI agents is increasingly shaping managerial decision-making. While AI systems can enhance human capabilities by performing routine tasks and providing analytical insights, their involvement also introduces new dynamics in control, responsibility, and trust (Jarrahi, 2018; Shin et al., 2023; Belanche et al., 2020). Managers’ perceptions of AI as either a supportive tool or a decision-making partner affect their willingness to collaborate with these systems (Kot and Leszczyński, 2022; Teng, 2024). Research shows that when AI is integrated into collaborative processes, managers tend to diffuse responsibility (Bleher and Braun, 2022). Beyond these operational dynamics, the replacement of a human businessperson with an AI agent may influence how managers evaluate the outcome of an interaction. Managers may judge the same performance differently when it is delivered by an AI as opposed to a human, and automation bias can lead to over-reliance on AI (Dzindolet et al., 2003). Additionally, algorithm aversion can lead to more critical evaluations when AI fails (Dietvorst et al., 2015). These tendencies suggest that AI’s presence may not only shape managers’ decision-making processes but also color their assessments of the success or failure of business interactions.

How managers evaluate these interactions is particularly critical in B2B markets, where business exchanges are not purely transactional but embedded in interpersonal relationships between company representatives (Koponen et al., 2021). Although these interactions create mutual value, their outcomes—if a deal is closed, a service is delivered, or cooperation breaks down—are key reference points for future business decisions (Sands et al., 2022). These outcomes influence satisfaction, trust (Doney et al., 2007), and negative word-of-mouth (Roy et al., 2019). When AI agents replace human partners, the absence of relational cues can exagerrate the importance of the outcome as the primary basis for evaluation. Managers may focus more strongly on whether the interaction succeeded or failed, rather than on relationship-building or operational aspects typical of human-to-human collaboration.

This focus on outcome may intensify, because performance perceptions tend to be asymmetrical. Research consistently shows that negative outcomes elicit stronger reactions than positive ones—a well-documented bias in human judgment (Kahneman and Tversky, 1979; Barron, 2021). Negative outcomes involving AI agents may magnify this effect, managers may perceive AI as being less adaptable and less capable of relationship recovery when compared to humans (Sands et al., 2022). Negative performance cues are also generally weighted more heavily than positive ones and tend to be viewed as more credible and diagnostic in managerial decision-making (Baumeister et al., 2001; Novemsky and Kahneman, 2005). Thus, when AI performance falls short, managers may not only experience greater dissatisfaction, but also become more skeptical of AI’s potential in business interactions.

Integrating AI in decision-making raises critical questions about the responsibility gap. In this situation, the outcome gives actors reasons to blame someone, but the accountability is ambiguous due to the involvement of AI (Hindriks and Veluwenkamp, 2023). Principal-agent theory traditionally views agents as self-interested actors controlled by humans, representing human decisions, which reinforces AI’s lack of genuine agency (Parkes and Wellman, 2015). However, cyborg theory and cyberagency (Fleischmann, 2009), suggest that this agency exists on a spectrum. Advanced AI, exhibiting degrees of autonomy and decision-making capability, challenges the binary view of agency and raises questions about the locus of causality in organizational outcomes.

Scenarios like the autonomous car dilemma (Bonnefon et al., 2016) highlight situations where AI must make choices, emphasizing the need for a clear link between agency and responsibility. While the “human-in-the-loop” model suggests variable levels of human oversight (Monarch, 2021), in practice, managers often face limited control, mostly when dealing with complex, opaque AI systems (e.g. black-box models) (Coeckelbergh, 2020). Such limitations in oversight contribute to a responsibility gap, where the locus of causality is unclear, and managers are uncertain about where to place responsibility for AI-driven outcomes. This ambiguity makes the responsibility gap a critical issue. It not only challenges ethics but also profoundly shapes managerial decision-making, potentially altering trust, risk assessment, and the willingness to adopt AI in organizational contexts.

The research problem in this study focuses on understanding how company managers attribute responsibility in B2B interactions with both human and AI actors. Specifically, we investigate whether the locus of causality—whether outcomes are perceived as internally (representative-driven) or externally (situationally-driven)—varies based on the success or failure of the business interaction. This research fills a crucial gap regarding blurred responsibility in human-AI collaboration (Hindriks and Veluwenkamp, 2023) by exploring how outcome-based attributions differ for AI and human representatives.

People respond more to negative than positive outcomes, as negative information is perceived as more credible and impactful on decision-making—a pattern known as negative asymmetry (Ariely et al., 2005; Baumeister et al., 2001; Barron, 2021). This effect is particularly pronounced in B2B contexts involving AI, as failures can reinforce perceptions of AI’s limitations in adaptability and responsiveness (Sands et al., 2022).

Evaluations of both positive and negative outcomes involve responsibility attribution. Research on locus of causality suggests that people tend to credit themselves for success and blame others for failure, this is known as self-serving bias (Gioia et al., 1985). However, attribution patterns may differ when AI is involved, as machines are often seen as lacking autonomy and control over their actions (Hong and Williams, 2019). People may attribute AI failures externally, assuming that the system simply followed rules, while humans are seen as autonomous and accountable.

This pattern may shift in the context of negative outcomes. While positive performance might lead to external attributions for AI, failures could trigger stronger internal attributions due to higher expectations of AI’s precision and reliability (Logg et al., 2019). When AI falls short, its presumed infallibility may lead managers to hold it responsible, making negative performance an exception to the tendency to excuse AI’s actions. Thus, we expect that:

Research shows that causal attributions play a key role in shaping satisfaction in business interactions (Folkes, 1984). Weiner (1986) argued that satisfaction judgments are particularly sensitive to responsibility being attributed internally to the agent or externally to situational factors. In the context of buyer-seller relationships, internal attributions tend to enhance satisfaction and increase the likelihood of cooperation (Oliver and DeSarbo, 1988). Conversely, external attributions, such as success being seen as a matter of chance, can lower satisfaction by reducing the agent’s perceived contribution.

Studies confirm that locus of causality is critical when business outcomes are negative, influencing satisfaction and switching intentions (Eslami et al., 2020; Wang et al., 2022). For example, blaming wholesalers for failures has been shown to reduce recovery satisfaction (Oflaç et al., 2021). While these patterns are well-established in human-to-human interactions, less is known about how attributions influence satisfaction when AI replaces human agents.

The literature on AI agency suggests that people often perceive AI systems as lacking autonomy and intentionality, seeing AI performance as the product of pre-set rules rather than individual competence (Hong and Williams, 2019). As a result, AI successes tend to be attributed externally, reducing the sense that the agent contributed meaningfully to the outcome (Weiner, 1986; Hohenstein and Jung, 2020). This may diminish satisfaction with successful AI agents, even when their performance matches a human’s. Managers may expect AI to deliver accurate and reliable performance (Logg et al., 2019), but when it succeeds, the outcome may be dismissed as routine or coincidental, further weakening satisfaction.

Therefore, we propose that the perceived locus of causality mediates the relationship between the type of service agent (AI vs. human) and anticipated satisfaction. Specifically, managers are likely to anticipate less satisfaction with successful AI services than with humans, even when their performance is identical. This discrepancy arises because AI performance is often attributed to external factors (e.g. luck) rather than the AI’s skills. Based on these insights, we hypothesize the following:

We conducted three experiments: two vignette studies (Studies 1a and 2) and one real-interaction study (Study 1b). Study 1a tested H1a and H1b, employing a 2 (performance: success vs. failure) x 2 (agent: AI vs. human) design to examine managers’ responsibility attributions. Study 1b replicated Study 1a in a real business setting, where managers interacted with an AI agent (powered by GPT-4). Study 2 tested H2, employing a one-factor design (agent: AI vs. human) in a success-only context to examine whether responsibility attribution mediates managers’ anticipated satisfaction with the agent.

Studies 1a and 2 used vignettes. In both studies, participants were shown screenshots of email conversations between a manager and either an AI or a human agent. Study 1a varied both outcome (success vs. failure) and agent (AI vs. human), while Study 2 focused only on successful interactions with AI vs. human agents. Participants evaluated responsibility attribution (Study 1a) or anticipated satisfaction (Study 2). The vignette method ensured experimental control but relied on hypothetical scenarios (Atzmüller and Steiner, 2010).

Study 1b used a real-interaction design. Managers engaged directly with an AI agent (GPT-4) to perform a business task. They made decisions and received real-time responses from the AI. This allowed us to observe responsibility attributions during live, two-way interactions, enhancing ecological validity. Combining vignette and real-interaction methods provided both control (Studies 1a and 2) and realism (Study 1b), offering a robust test of our hypotheses.

We recruited 355 English-speaking managers, using the Prolific platform as the recruitment platform. Prolific is used to reach specific groups of professionals for online research (Palan and Schitter, 2018). Examples of studies leveraging Prolific include experiments on explainable AI (Gaczek et al., 2023a), consumer decision-making (Otterbring et al., 2020), and many other fields (e.g. Wang et al., 2024; Cascio Rizzo et al., 2024). Eyal et al. (2021) found that Prolific participants demonstrate greater attentiveness and honesty in research participation than users of MTurk or CloudResearch.

Our primary focus was on B2B marketing managers with experience in decision-making processes in companies that predominantly operate as sellers. Screening criteria ensured participants had relevant experience in directing and coordinating work and the authority to make organizational-level marketing decisions.

To ensure independent samples across studies, we excluded managers who participated in previous experiments in the sampling process. To enhance consistency and relevance, we targeted managers based in the US and UK, with English as their primary language. This aligned the sample with the study’s focus on B2B marketing decision-making in English-speaking markets. This approach helped provide a diverse yet relevant sample that reflects essential managerial perspectives in the B2B domain. Participants, who engaged in the studies for nominal compensation, were explicitly selected for their background in marketing-related decision-making.

In Study 1a, 97 managers (M_age = 33.19 years, SD = 12.85; 43% female) participated, each with an average of 4.9 years (SD = 5.36) of experience in marketing-related managerial roles. Study 1b included 57 mid-level managers (M_age = 36.44 years, SD = 10.16; 42% female) with an average of 5.95 years (SD = 6.23) of experience, overseeing an average of 7.29 subordinates (SD = 14.49). Study 2 consisted of 201 managers (M_age = 36.54 years, SD = 11.1; 55% female) with an average of 6.74 years (SD = 6.03) of managerial experience.

The general scenario across all studies involves a marketing manager working for a food company. When AI replaces their business partner, the manager makes a marketing decision (Studies 1a and 2: organizing a trade show; Study 1b: selecting a marketing communication approach). Two distinct experimental scenarios were used to enhance the external validity of our findings. The experimental designs were designed to address the main research question: the impact of blurred responsibility (for positive vs. negative performance) on satisfaction with a business interaction when AI replaces a human representative.

In Study 1a, participants were exposed to one of four (randomly assigned) email conversations between representatives of two companies (successful AI vs. unsuccessful AI vs. successful human vs. unsuccessful human; see Web Appendix A). The email thread was structured to resemble an inquiry from one company to another. The individual sending the inquiry was always a human (working for a food company), while the responder was either a human or an AI agent. The human agent was signed as “John – customer service,” while the AI agent was labeled “John – virtual agent.”

Before reading the email thread, participants received a brief description of the responder, identifying them as either “John, a human agent specializing in initial responses to requests for proposals” or “John, an AI-powered agent specializing in initial responses to requests for proposals.” In the successful conversation condition, the response indicated agreement to cooperate, while the unsuccessful condition indicated a refusal to cooperate. Study 2 follows the same procedure as Study 1a, with only one modification—the participating managers are presented with a successful scenario. In addition to collecting responses regarding responsibility attribution, the dependent variable was anticipated satisfaction.

Study 1b differs significantly from Study 1a and Study 2. It utilized an authentic, interactive AI agent based on a Large Language Model (ChatGPT 4o API) connected to a specific database to limit responses strictly to the study context and prevent AI “hallucinations” (Salvagno et al., 2023). Participants interacted with an agent autonomously, asking questions or giving commands in their prompts. The agent reacted to them using natural language processing while accessing a database (it was an open conversation).

The scenario involved a scenario where participants decided which advertising to select for a product launch. Participants received background information about the AI agent, including its access to data on ad effectiveness, along with a sample from the database (see Web Appendix A). The study instructed participants to engage with the AI agent to seek ad recommendations, allowing up to six prompts and requiring at least one. Participants (as in Studies 1a and 2) were recruited on the Prolific Platform and then directed to the study’s web page, where they could communicate with the AI agent. After interacting with the AI, participants rated its performance on a 7-point Likert scale (1 - unsuccessful; 7 - successful). Using these continuous ratings, we applied the median split technique to create “successful” and “unsuccessful” groups. We then compared the perceived locus of causality between these two groups. Across all studies, participants responded to two manipulation check questions to detect automatic and non-reflecting answers (Wilson et al., 2010).

In studies 1a and 2, we used a modified version of The Revised Causal Dimension Scale (McAuley et al., 1992) to collect participants’ perceived locus of causality (3 items; Cronbach’s alpha 0.74; M = 5.16, SD = 2.16 in Study 1a). Specifically, we first asked participants to think about the reason for success (vs. failure) of AI (vs. human), write it down, and then answer three questions regarding the source of the performance outcome (1–9 Likert scale, where: 1 - something that reflects an aspect of agent, 9 – something that reflects an aspect of the situation; 1 – something inside of agent, 9 – something outside of agent; 1 – something about agent, 9 – something about others). Here, the higher the perceived locus of causality score, the more likely the cause is to reflect external factors not related to the agent. On the contrary, the lower the score, the more responsibility would be attributed to the agent.

Since we used two manipulation checks, we only analyzed data from participants who correctly answered both questions (“Who was the responder?” and “Was the responder successful?”). In Study 2, to obtain participants’ anticipated satisfaction from working with the agent, a modified version of satisfaction with hypothetical experience (Homburg et al., 2005) was used (7-point Likert Scale, where 1 – strongly disagree and 7 – strongly disagree; 4 items Cronbach’s alpha 0.93; M = 4.88, SD = 1.23).

Since Study 1b is based on a real interaction between a manager and AI, we modified the measurement for the locus of causality. Following the conceptual framework for typical attribution research, Study 1a placed the participant in the observer role, where they see the interaction of others. Thus, when assessing the locus of causality, the participant chose between agent-related cause vs. situational cause. In contrast, Study 1b puts the respondent as an actor who actively participates in the interaction. Consequently, success vs. failure is attributed to either the AI agent or the manager (who interacts with). A lower locus score in both measurements indicates more causality attributed to the agent.

The measurement method was modified because of the realistic scenario and the desire to provide more robust results, which is the purpose of replication studies. To meet these objectives, the locus of causality was measured using a one-item scale adapted from McAuley et al. (1992). Participants answered: Who is responsible for the consequences of your decision on the advertising campaign? We used a semantic differential scale to collect their responses, where 1 – The AI agent is responsible, and 7 –I am responsible.

Compared to Study 1a, another significant procedure and measurement modification was made. Since Study 1a was a vignette, and the respondent acted as an observer of the interaction, we could manipulate the success or failure of the interaction. In Study 1b, the participant decided whether the interaction was a success or failure. The 7-point Likert scale was used to collect how strongly participants agreed with the statement that the AI agent’s performance was a success. Then, we followed the median split method at the data analysis stage to group participants into successful and unsuccessful conditions. Although the median split method receives some criticism related to the reduction of power, more recent research suggests that it is suitable for data that are not correlated (Iacobucci et al., 2015). In our case, the correlation between the dependent variable (locus of causality) and performance score was non-significant, r(55) = −0.124, p = 0.358. Please see Web Appendix B for all measurements.

Study 1a investigated whether managers’ locus of causality differs between AI (vs. human) agents. Moreover, we study whether this effect depends on the business interaction’s success or failure. The participant’s task was to read an email thread between representatives of two companies that want to cooperate, then attribute reasons for the service agent’s successful or unsuccessful performance (human vs. AI). Study 1a is a vignette, usually used to elicit participants’ judgments about different scenarios (Atzmüller and Steiner, 2010). Our idea for this study is also connected to experiments regarding actors’ and observer’s causal attributions (Malle, 2006). However, the current scenario only considers observer conditions, since participants did not interact with the real AI/human agent.

We assessed differences in the perceived locus of causality using moderation analysis using PROCESS macro SPSS (Model 1) (Hayes et al., 2017). Agent type (AI vs. human) was the independent variable, the perceived locus of causality was the dependent variable, and the performance outcome, successful vs. unsuccessful, was the moderator.

The regression analysis showed a significant negative effect of agent type (AI vs. human) on the dependent variable (b = −3.78, p = 0.007), so that performance outcome was attributed to internal factors for human agents and external factors for AI agents. Crucially, there was a significant interaction between agent type and interaction outcome (success vs. failure) on the dependent variable (b = 2.38, p = 0.008). However, conditional effects showed that agent type influenced only the attribution of success (b = −1.39; p = 0.024; LCI = −2.6; UCI = −0.18), while there was no difference in the attribution of failure due to agent type (b = 0.98; p > 0.05; LCI = −0.25; UCI = 2.22), Figure 1.

Study 1a reveals that managers’ attributions for AI and human agent performance differ in successful scenarios, but not in unsuccessful scenarios. For successful outcomes, AI success is attributed more to external factors, while human success is linked to internal characteristics, supporting H1a. However, for failures, no such difference emerged, which does not support H1b.

This pattern is consistent with attribution theory (Weiner, 1986), which emphasizes the role of locus of causality in shaping evaluations of performance. The tendency to credit human success internally aligns with research on satisfaction in buyer-seller relationships (Oliver and DeSarbo, 1988). Conversely, AI success is externally attributed, reflecting perceptions of AI as lacking autonomy and simply following rules (Hong and Williams, 2019; Hohenstein and Jung, 2020).

The absence of agent-based differences in failures contrasts with previous research (Leo and Huh, 2020; Hohenstein and Jung, 2020), suggesting that negative asymmetry may explain this outcome. Managers may focus on the negative result itself rather than the agent, as negative information is generally weighted more heavily (Kahneman and Tversky, 1979; Gaczek et al., 2023b; Taylor, 1991).

Self-serving bias (Gioia et al., 1985) offers an additional explanation: managers may distance themselves from AI success by attributing it externally, while avoiding blame for failures to protect their self-image (Taylor and Brown, 1988). This behavior reflects the responsibility gap in human-AI collaboration, where accountability for AI-driven outcomes remains unclear (Hindriks and Veluwenkamp, 2023). Together, these findings demonstrate how locus of causality shapes managers’ evaluations of AI and human performance, emphasizing the importance of attribution in understanding satisfaction and trust in AI-assisted B2B decision-making.

Study 1b was designed to replicate findings from Study 1a using a real interaction between managers and AI agents. Vignette methods have limitations due to unrealistic scenarios and contexts. So, we developed an AI agent for participants to communicate, collaborate and make managerial decisions with. We aim to increase the validity of studying human-AI interaction in a business setting. Although the method used in Study 1b differs from Study 1a, the interaction context remained the same, as did the task managers were assigned. Importantly, Study 1b considers only the AI condition, focusing on the effect of successful vs. unsuccessful interaction outcomes on the perceived locus of causality.

We assessed differences in the locus of causality in the two conditions (unsuccessful vs. successful AI performance) using a one-way analysis of variance (ANOVA). Before conducting the ANOVA, we performed Levene’s test for equality of variances based on the median. The results were insignificant, F(1, 55) = 3.908, p > 0.05, suggesting that the assumption of homogeneity of variances was met across the two conditions.

An ANOVA for the unsuccessful vs. successful AI performance as the independent variable and the locus of causality as the dependent variable revealed a significant effect of the independent variable on the dependent variable, F(1, 55) = 4.552, p = 0.037, η² = 0.076. Mean comparison indicated that when the interaction with the AI agent was unsuccessful, the managers attributed the cause to the AI agent (M = 4.95, SD = 1.8). However, when the performance was successful, managers attributed the cause to themselves (M = 5.78, SD = 1.12), Figure 2.

Study 1b provides deeper insights into managers’ responses during real interactions with an AI agent. The results confirm that causal attributions for AI’s performance depend on the success or failure of the task performed. Failure was attributed internally to the AI in unsuccessful outcomes, while credit shifted to the human manager for success. This shows how the perceived causality of AI is outcome-dependent.

A comparison of results from Studies 1a and 1b shows discrepancies related to the responsibility of the AI Agent. In Study 1a, differences in attribution were observed between AI and human agents but not between successful and unsuccessful AI Agent (M_{AI success} = 5.77 SD = 2.44 vs. M_{AI failure} = 5.09 SD = 2.02, p > 0.05, Study 1a). On the contrary, in Study 1b, we observed differences in the attribution responsibility of AI Agent in both successful vs. unsuccessful performances.

Seeking to explain this discrepancy, we consider the manager’s role in the experiment (observer vs. actor) and their level of engagement/proximity to the “decision” (Carson et al., 2022). In hypothetical scenarios, the manager acts as an observer, leading to more neutral and stable attributions. However, in real interactions, the manager takes on the role of an actor, directly engaging with AI and experiencing the outcomes firsthand (Gioia et al., 1985; Malle, 2006). This shift intensifies attribution biases due to self-serving bias (Miller and Ross, 1975), where success is attributed to oneself and failure to external factors – in this case, AI.

Moreover, algorithm aversion (Dietvorst et al., 2015) explains why managers react more negatively to AI failures when engaged. The real interaction likely increases expectations of AI performance, making mistakes more salient and triggering stronger adverse reactions. This effect is further reinforced by motivated reasoning (Kunda, 1990), as managers seeking to protect their self-image, rationalize poor outcomes by blaming AI rather than acknowledging potential shared responsibility.

Thus, we conclude that observational scenarios maintain stable, less outcome-dependent attributions. On the contrary, real interactions heighten engagement and proximity (Carson et al., 2022), activating biases that intensify responsibility judgments. The discrepancy is best understood as a consequence of self-enhancing attribution tendencies (Shepperd et al., 2008) and AI-specific skepticism (Logg et al., 2019).

Study 2 aims to test the mediating role of the perceived locus of causality in anticipatory satisfaction with AI (vs. human) agents. Attribution theory explains the causes of events, and these inferred causes can influence satisfaction judgment later, serving as satisfaction antecedents (Tsiros et al., 2004). In the domain of management studies, most of the papers focus on customers’ and managers’ attributions of failure, namely—who to blame for disappointing performance and how it translates into satisfaction and behavioral intentions (e.g. willingness to recommend service provider) (Van Vaerenbergh et al., 2014).

Causal attributions are also seen to influence personnel decisions in organizations (Struthers et al., 1998) and to determine individual resistance or acceptance of information technologies (Martinko et al., 1996). Recent studies investigate the success and failure of AI performance and internal or external attribution (Coeckelbergh, 2020; Hohenstein and Jung, 2020). In this study, we first test the meditative role of causal attributions in managers’ anticipated satisfaction with AI agents. We posit that an increase in the perceived locus of causality will lead to a decrease in expected satisfaction from the service agent (AI vs. human), which is partly confirmed in consumer research (Srinivasan and Abi, 2021).

To assess the meditative effect of perceived locus of causality on anticipated satisfaction with AI and human agents, we conducted a regression analysis with the condition (AI = 1; human = 2) as the independent variable, satisfaction as the dependent variable and perceived locus of causality as a mediator. The regression was conducted via PROCESS macro SPSS (Model 4) (Hayes et al., 2017).

The mediation analysis showed a positive direct effect of the condition (AI = 1; human = 2) on anticipated satisfaction (b = 0.45; p = 0.009) and a negative impact on the perceived locus of causality (−0.79; p < 0.001). Crucially, the regression indicated that the indirect effect of the condition on the dependent variable is no longer significant (b = 0.3; p > 0.05) after introducing the mediator, while the negative effect of the mediator on the dependent variable is significant (b = −0.2; p = 0.013). Therefore, the full mediation model is supported in Figure 3.

Study 2 builds upon Studies 1a and 1b, confirming that AI success is more likely to be externally attributed when compared to human success, lowering anticipated satisfaction. This aligns with attribution theory (Weiner, 1986), which emphasizes the role of locus of causality in shaping evaluations and motivation. When success is seen as situational rather than agent-driven, satisfaction decreases—a pattern well-documented in buyer-seller relationships (Oliver and DeSarbo, 1988) and now extended to AI agents in business contexts.

This reduced satisfaction can also be interpreted through theory of mind limitations (Shank and DeSanti, 2018) and AI agency research, suggesting managers struggle to perceive AI as capable of intention or competence (Hidalgo et al., 2021; Hong and Williams, 2019). Not attributing success to AI’s skill, managers may view positive outcomes as coincidental, undermining trust and reducing engagement with AI systems—a known barrier to AI adoption in managerial practice (Yue and Li, 2023).

By linking attribution processes to satisfaction, Study 2 highlights how psychological attributions can shape AI acceptance, emphasizing the need for AI design that fosters perceptions of competence and agency to strengthen trust and satisfaction in human-AI collaboration.

A series of three experiments tested how managers attribute responsibility (internal vs. external) for success or failure when interactions involve either other humans or a combination of the two (Table 1). The study also examined whether the internal vs. external locus of causality affects satisfaction with the business interaction. Our results support and broaden the conclusions from other fields, suggesting that people deemphasize the success of AI when compared to humans, since they attribute the AI causes externally. At the same time, managers consistently avoided responsibility for failures, regardless of whether they interacted with a human or an AI agent.

Previous studies mainly focused on responsibility for negative performance (Srinivasan and Abi, 2021), while we provide examples of positive performance in a managerial decision-making context. We point out that the lower satisfaction with AI (compared to humans) results from managers’ belief that the success of AI is due to external factors unrelated to AI. To our knowledge, this study tested the mediating nature of locus of causality in shaping managers’ satisfaction with AI collaboration for the first time. Table 2 summarizes the three studies’ results. Then, we present several contributions from these findings.

First, our findings contribute to theoretical debates on AI autonomy and human responsibility. Principal-agent theory views AI as a tool under human control, reinforcing its lack of genuine agency (Parkes and Wellman, 2015). Alternative perspectives, such as cyborg theory and cyberagency (Fleischmann, 2009), suggest that AI operates on a spectrum of autonomy or that humans perceive it as having some agency (Nass and Moon, 2000). However, these studies usually consider abstract moral questions and do not tackle the real-world implications. Our study confirms that managers perceive AI responsibility on a spectrum. This attribution is driven by the perception that AI is not just a passive tool but an independent decision-maker, shifting the locus of causality from humans to AI. We conclude that the locus of causality shapes AI’s responsibility, as a stronger perception of AI as the primary causal agent increases its perceived responsibility for outcomes.

Second, this study contributes to the discussion in management literature on factors influencing managers’ openness to accept AI for decision-making (Oppioli et al., 2023). Previous studies are conceptually rooted in a technology acceptance model, so they focus on understanding the features of managers and AI that impact the use of it. Our study adds psychological insights (internal vs. external locus) by pointing out attribution as a possible aspect of AI acceptance. Across three experimental studies, we provide theoretical insights into attribution theory, agency problems, and human-AI collaboration satisfaction. With this, we fill the research gap related to the use of AI by marketing managers and the issue of attribution of responsibility for recommendations coming from algorithms.

Third, research on attribution processes towards AI has primarily focused on attributing responsibility for errors (Hohenstein and Jung, 2020), suggesting a phenomenon of algorithm aversion (Dietvorst et al., 2015). The problem of who to blame for the AI failure is seen in the literature on ethical dilemmas (Sullivan and Fosso Wamba, 2022) and human-AI interaction (Porsdam Mann et al., 2023). Our findings extend this research by showing how professionals use AI and attribute responsibility to both positive vs. negative outcomes. Contrary to previous studies, we demonstrate that differences in locus attribution to humans vs. AI are observable in a positive context (success condition). So, it is not simply the issue of ethical challenges that contribute to discrepancies at the attribution level. By doing that, we confirm that managers perceived AI as social actors to some extent, and they believe that artificial agents can take some of the blame as if they were human collaborators (Nass and Moon, 2000). That finding suggests that AI agents are taken as potential collaborators, not only as objects or tools (Kot and Leszczyński, 2020).

Linking our findings to the classic perspective on attribution theory, we suggest that managers, when acting as observers (Study 1a) or actors (Study 1b), demonstrate similar patterns of responsibility attribution. Regardless of their role, AI success is associated with the algorithm’s performance less, and more often credited to the manager interacting with the AI (This AI cannot be so successful … it must be the human factor!). Conversely, in the case of failure, managers tend to shift blame onto the AI itself (Of course, it is the AI’s mistake, not the human’s!), but only when they actually interact with the Agent. In such involving scenarios, this behavior may reflect a motivation to preserve self-esteem by avoiding personal accountability for the AI’s shortcomings. In contrast, managers enhance their self-esteem during successful outcomes by taking credit for the positive results. This contributes to the body of research on self-serving bias within human-AI collaboration.

Finally, the literature on managers’ responses to AI mainly explains their attitudes to AI or their intention to use it or have AI as a team member (Cao et al., 2021; Haesevoets et al., 2021; Berretta et al., 2023). These authors measure the managers’ attitudes towards making decisions with AI instead of human behavior, and indicate that the primary driver of satisfaction when using AI is utilitarian gratification (Xie et al., 2024). We add to that stream of research in two ways.

We show that managers are less satisfied with AI than human agents, even when the performance is the same. Thus, we support the results of studies showing human aversion to machines (Dietvorst et al., 2015), while our observation relates directly to a management scenario. We conclude that although managers use AI for work, they are vulnerable to the same biases as regular users who lack the expertise to verify recommendations from AI. Therefore, users can expect rational benefits from working with AI (Esposito et al., 2024) and, their lower satisfaction may be due to non-rational factors, as the available research suggests (Gaczek et al., 2023a).

Our work extends the studies on AI satisfaction by focusing directly on outcomes. We challenged managers with the actions and decisions that had already been made and their positive and negative consequences. In Study 1b, we went beyond vignettes and tested participants’ responses to the conversational agents. We allowed them to experience collaboration with AI, the generative technology of Large Language Models, to add new insights into the satisfaction and dissatisfaction with the outcomes.

Our findings offer several critical insights for managers and organizations seeking to integrate AI into decision-making processes, particularly in services. These implications focus on actionable insights for improved satisfaction with AI collaboration, addressing responsibility attribution. Companies should consider implementing AI in tasks where data processing strengths and efficiency with repetitive tasks are maximized. AI can be more effective for complex data-intensive decisions. In contrast, human agents may be more effective in decisions that require high levels of social interaction or intuitive decision-making. Senior managers should clearly define and communicate the roles and responsibilities of AI versus human agents to ensure smooth integration and reduce employee resistance.

Transparent systems might facilitate collaboration with AI. Explaining AI’s decision-making processes can help managers better understand AI recommendations. This could mitigate the tendency to attribute success externally and failure internally and reduce the ambiguity in responsibility attribution. Organizations can also point out successful human-AI partnerships, which may limit the tendency to attribute positive outcomes solely to external factors. This could build managers’ confidence in AI capabilities.

Organizations should also establish clear frameworks for assigning responsibility in AI-driven processes. These frameworks should explain where to attribute responsibility for failures to avoid the “responsibility gap” and reduce reluctance to rely on AI. We also suggest developing an evaluation to assess managerial satisfaction with AI, provide feedback on AI’s shortcomings, and allow refinement of AI systems.

These implications might be significant for companies integrating AI into their value proposition to enhance business customer satisfaction. AI systems should be deployed in areas where their strengths complement human decision-making capabilities. Instead of replacing humans, AI can enhance the value of services delivered to customers through efficiency, data-driven insights, and decision-making support. Positioning AI as a partner rather than a cost-saving tool might improve customer satisfaction. However, service providers should communicate that AI might not lead to successful outcomes and openly address possible failures.

Implementing AI in decision-making also requires training managers to understand and mitigate attribution biases, particularly the tendency to attribute AI success to external factors and human success to internal factors. If not, managers can perceive AI as useless because they link its success with external factors like accidents or good luck. Such training can help increase the likelihood of objective AI performance evaluations. Training could teach managers to recognize and reduce cognitive biases towards AI and take a balanced perspective on AI contributions. Another aspect that could be clarified is how to define the roles in the decision-making process with AI: who makes decisions, who takes responsibility, and what can be expected from AI.

Table 3 suggests actionable steps and tools organizations can implement in their AI systems to limit attribution bias in AI decision-making. These help organizations recognize and mitigate blaming AI for failures and crediting humans for success.

This study has limitations that could be overcome in subsequent research to understand managers’ responses to AI. The vignette method applies controlled, semi-realistic scenarios to test hypotheses. However, it introduces several biases: responding to imagined situations rather than actual experiences, simplifying decision-making organizational environments, and missing emotional engagement from participants. Further studies could focus on interactions with actual AI to capture human-AI interactions’ dynamic and iterative nature. That could lead to investigating how satisfaction and attributions evolve as managers learn from more prolonged interactions with AI. This is particularly important when studying the responsibility gap because an actual event must occur to assess responsibility for AI actions accurately.

One limitation of our study is the use of a human-like name for both the AI and human agents, which could have influenced participants’ perceptions through anthropomorphization. Future research could examine how naming conventions or other anthropomorphic traits impact service interactions, especially in contexts where AI agents are perceived as collaborators rather than tools. It is possible that assigning human-like traits to the agent makes it seem more human, which should facilitate associating it with intentions. In turn, the perception of having intentions could shift responsibility away from the manager and toward the “human-like” AI (Ahn et al., 2024). Further research should examine the relationships between responsibility, AI agency, and perceived intentionality or autonomy. Such studies would help identify the psychological foundations of the ability to share responsibility.

Having the managers use a single platform raises concerns about the generalizability of findings. The crowdsourcing sample can be biased by participants from English-speaking countries, being more tech-savvy (Berger et al., 2021) and more comfortable with technology (Bogert et al., 2021) than average. Therefore, future studies could replicate our findings using more senior executives or highly specialized marketing managers and expand to culturally and industry-diverse samples of managers, especially in non-Western contexts, to explore whether findings generalize.

This study’s generalization is also limited by using one technology—conversational agents. Although study 1b included interactions with the generative AI (Large Language Model-based agent), the interaction was still limited to text. Further studies could extend this scenario to voice or graphical interactions with conversational agents or to different types of agents. Further studies could focus on attribution bias in decision-making with agents that maximize managers’ satisfaction or benefits (utility-based agents) or agents that improve recommendations learning from managers’ feedback (learning AI agents). Understanding how the responsibility gap diminishes or expands depending on different forms of communication between the manager and the type of AI agent may be particularly important.

Additionally, the discrepancy between Study 1a and Study 1b raises concerns about the generalizability of our findings in real-world settings. In Study 1a, where managers acted as observers, attribution patterns remained stable regardless of outcome. However, in Study 1b, where managers were directly engaged with the AI, attribution became more outcome-dependent. This suggests that observational studies may not fully capture the attribution patterns that arise in real-world interactions. Direct engagement likely amplifies biases such as self-serving bias and algorithm aversion, which may influence managerial judgments of AI responsibility. Future studies should consider the role of engagement and decision proximity to ensure findings extend to practical AI applications.

Our findings indicate that satisfaction with an AI agent is lower due to the role of external factors unrelated to the AI itself. A pathway for future research is to identify the specific factors contributing to this effect, particularly those typical in attribution studies, such as task difficulty and behavioral stability. Furthermore, the latest research highlights the mixed involvement of AI in decision-making processes—other humans can control AI. In our experiments, the AI operated independently, which significantly influenced perceived responsibility. It would be intriguing to explore how the responsibility gap evolves when a manager works with an algorithm, knowing another person controls it. Finally, we propose introducing a new mediating variable—intentionality—to enhance our understanding of responsibility attribution to AI. Future research could explore how “intentional” AI affects perceptions of agency and whether this increases the attribution of responsibility.

This research was entirely funded by the National Science Centre, Poland 2022/47/B/HS4/01153.

The supplementary material for this article can be found online.