AI assistants have established a place in our lives and are applied millions of times worldwide (Statista, 2021), for which a specific and increasing number is taking place in the financial sector and the banking industry (Kochhar et al., 2019). Artificial Intelligence as an advisor in financial decisions is becoming a reality in everyday business (Ludden et al., 2015; Vincent et al., 2015; Jung et al., 2018; Shanmuganathan, 2020). AI can advise on whether to make bank transfers, deposit money, buy stocks, or to make other investments, sometimes in a highly speculative manner (Fein, 2015; Tertilt and Scholz, 2018). For instance, to promptly respond to price fluctuations, many individuals execute such financial decisions directly on their smartphones or tablets (Varshney, 2016; Azhikodan et al., 2019; Lele et al., 2020). Privacy considerations are crucial (Manikonda et al., 2018; Burbach et al., 2019), and the level of autonomy of a financial bot may impact its acceptance and adoption. As financial AI assistants have access to sensitive data and may be able to make critical decisions independently, users may be hesitant to use it if they perceive a lack of control over the system's actions.

Research on human-machine interaction shows that individuals engage in social behavior toward machines by applying heuristics from interpersonal relations (Nass et al., 1997; Christoforakos et al., 2021). Trust is a critical component in social interactions, as it allows individuals to rely on and believe in the actions and statements of others (Rotter, 1980; Scheuerer-Englisch and Zimmermann, 1997; Neser, 2016). Social psychologists view trust as a multidimensional construct that is composed of cognitive, affective, and behavioral dimensions (Johnson and Grayson, 2005; Righetti and Finkenauer, 2011; Lyon et al., 2015). This means that trust involves positive expectations toward a trustee based on an emotional bond, and it expresses itself in concrete actions of both the trustor and the trustee. Trust also depends on the existence of risk and the individual's willingness to accept vulnerability, and it is associated with expectations of a positive outcome and positive prospective behavior from the other person (Mayer et al., 1995; Flores and Solomon, 1998).

Cognitive trust (as one route of trust formation) is characterized by the assessment of an individual's reliability and dependability (Rempel et al., 1985; McAllister, 1995; Colwell and Hogarth-Scott, 2004), based on the trustor's rational evaluation of the trustee's knowledge, competence, or understandability (McAllister, 1995; Kohn et al., 2021). In contrast, affective trust is an advancement of cognitive trust (Chen et al., 1998; Kim, 2005), which involves mutual interpersonal care and concern, and is characterized by an emotional bond and feelings of security between individuals (Rempel et al., 1985; Johnson and Grayson, 2005). Emotional dependence is a critical determinant in which the impact on trust is more closely associated with personal experiences as compared to cognitive trust. Both the emotional aspect and task-specific abilities may lead to either a positive or negative evaluation of an agent's trustworthiness. Components such as competence and knowledge gained from interpersonal interaction research (McAllister, 1995) are highly important in HCI and affect the level of trust (Chen and Terrence, 2009; Christoforakos et al., 2021) and ultimately the acceptance of automated systems (Ward et al., 2017).

When interacting with digital assistants and AI applications, it may be difficult to accurately evaluate trust-relevant characteristics (Rai, 2020). Often labeled as “black box” services, users have less knowledge about what actually is transpired in the service performance which then requires a high portion of trust (Rosenberg, 1983; MacKenzie, 2005; Devlin et al., 2015). Appropriate trust calibration is one important factor which should be considered in the use of AI powered assistant tools. Low levels of trust can lead to disuse (Lee and Moray, 1992; Parasuraman and Riley, 1997), while very high levels of trust in automated systems may be associated with over-reliance and over-trust (Freedy et al., 2007; Parasuraman et al., 2008; Körber et al., 2018). Additionally, users often evaluate their own decisions in comparison to automated solutions but may also develop a tendency to over-rely on automated systems (Young and Stanton, 2007; Chen and Terrence, 2009). Conversely, people who overestimate their performance tend to exhibit under-reliance on AI systems (similar to the Dunning-Kruger Effect in social psychology), which hinders effective interaction with AI-powered assistant tools (He et al., 2023).

In addition to certain voice characteristics such as gender, naturalness or accent (Nass and Brave, 2005; McGinn and Torre, 2019; Moradbakhti et al., 2022; Schreibelmayr and Mara, 2022), other features on the behavioral level of assistance tools like agency are becoming more and more important (Kang and Lou, 2022). Agency pertains to an AI assistant's self-control ability (e.g., Gray et al., 2007), whereas autonomous AI usually has high agency, making decisions based on data, while low agency AI relies on consistent human input. Machines and systems work on a highly autonomous level and according to the situation, offering their help adaptively and in real time (Qiu et al., 2013; Profactor, 2021). Even though users want to experience a proactive style when interacting with a chatbot (Medhi Thies et al., 2017; Pizzi et al., 2021), they also crave some sense of control over the chatbot's or autonomous system's actions and may feel threatened if they behave too autonomously (Złotowski et al., 2017; Stein et al., 2019; Seeber et al., 2020a,b). By way of example, proactivity for irrelevant information is viewed negatively (Chaves and Gerosa, 2021), but it is largely unclear which degree of agency is actually desirable and demanded for AI. Various studies indicate the effects of agency on interactions between humans and virtual characters or robots (Guadagno et al., 2007; Beer et al., 2014; Fox et al., 2015; Pitardi et al., 2021) and Stein et al. (2019) have shown, for example, that participants experienced significantly stronger eeriness if they perceived an empathic character to be an autonomous AI. Considering the phenomenon of the uncanny valley (MacDorman and Ishiguro, 2006; Mori et al., 2012), the concept of a thinking robot that autonomously generates ideas, desires, and expresses needs (Parviainen and Coeckelbergh, 2021; Hanson Robotics, 2023) is unsettling and evokes strong feelings of eeriness alongside fascination (Gray and Wegner, 2012; Stein and Ohler, 2017; Appel et al., 2020). The degree of agency differs depending on technical possibilities and improvements and more research has to be done taking various perspectives into account to create a user-centric point of view. Based on the literature, the degree of agency of an AI banking assistant should be considered in the present study to examine the impact on trust and intention to use.

In the current research, we asked participants about their perceptions of an artificially intelligent banking assistant. As the utilization of assistance systems in the financial sector is becoming more important and poses potential risks related to trust theories (Mayer et al., 1995; Lee and See, 2004), we chose this particular context to examine participants' impressions after they viewed a video introduction of a fictitious AI banking assistant. We explored impressions that were evoked by the AI banking assistant across different user groups (low vs. high trust) by asking the participants to write down how they would describe the presented AI. Additionally, participants answered closed self-report questionnaires typically used in empirical studies on trust in automation, technology acceptance, and the uncanny valley phenomenon. As experimental manipulation, we incorporated two levels of autonomy into the AI banking assistant. Factors of the technology-acceptance model (TAM, Venkatesh and Bala, 2008) were included and discussed in the light of the Trust in Automation literature (Körber, 2019).

The sample size required for the present online experiment was calculated by a power analysis, which helps to determine how many study participants are needed, using G^*Power (Cohen, 1992; Faul et al., 2007). For the calculation, a medium effect size of d = 0.50 was assumed and α error probability was set to 0.05. In order to achieve a power (1 – β) of 80% (β represents the probability of committing a Type II error, which is the error of failing to reject a null hypothesis that is actually false), the analysis resulted in a recommended sample size of at least N = 128 to run a test procedure with two independent groups. The participants were recruited through a snowball approach¹ at the campus of the Johannes Kepler University Linz, Austria. Overall, 136 participants took part in the online experiment. After excluding 9 participants as their indicated age was below 16 years or because of insufficient German skills, the final sample consisted of 127 participants (62 females, 61 male, 2 non-binary, 2 no specification), aged between 17 and 82 years (M = 30.55, SD = 15.47). 49.6% of participants had completed compulsory education with a high school diploma (Matura or Abitur), 38.6% had completed university education with a bachelor's level or higher, and 11.8% had some other type of vocational training.

The experiment was conducted as a 2 x 1 between-subjects design, involving 2 groups of participants experiencing distinct levels of the independent variable to analyze its effects on the dependent variable (N_Highagency = 61, N_Lowagency = 66), and was designed with the online survey software (Questback, 2020). At the beginning of the study, all participants were instructed to either use headphones or keep their computer/laptop audio on a high volume for the duration of the study. After that, participants read an introduction, confirmed their consent, filled out demographic information (including age, gender, and level of education) and were asked to fill in personality questionnaires (openness to experience, neuroticism, and propensity to trust). Once these scales were completed by the participants, brief instructions appeared followed by one out of two AI banking assistant videos, which were randomly assigned to each participant. Additionally, every video was divided into two parts (~1 min each). Subsequent to the first part, the participants were queried about the degree of human-likeness and eeriness of the AI banking assistant. Items regarding AI banking assistant's agency (manipulation-check) were asked after the first part of the video as well. After the second part of the video, the participants were requested to assess the perceived level of trustworthiness they had in the assistant featured in the video. Then, the participants were requested to write down attributes of the AI banking assistant and their intention to use it. Finally, some check items were asked to make sure that the participants had clearly understood the sound of the video (it was technically not possible to skip the video), that their German language skills allowed them to understand the content of the study, and that they had answered all questions honestly and conscientiously. The average processing time for the entire survey was approximately 15 min and 31 s.

Open-ended descriptions of the AI banking assistant. We asked the participants to provide a verbal description of the AI banking assistant using three terms in an open comment format. (“If you were to tell someone else about the AI banking assistant in the video, what terms would you use to describe it?”).

Trust was measured with two items of the trust in automation subscale (Körber, 2019). The items were modified slightly to match the content of the AI banking assistant (e.g., “I would trust the AI banking assistant.”, ranging from 1 = not at all to 5 = very much) (Cronbach's α = 0.760). Find all items in Appendix B.

Competence (4 items, Cronbach's α = 0.598), understandability (2 items, Cronbach's α = 0.744), and intention of developers (2 items, Cronbach's α = 0.738) were measured with the help of shortened scales taken from the trust in automation model (Körber, 2019). The items were modified slightly to match the content of the AI banking assistant. Find all items in Appendix B.

Intention to use was measured with the help of two items based on the intention to use items from the Technology Acceptance Model (TAM3, Venkatesh and Bala, 2008). A 5-point Likert-scale (1 = not at all to 5 = very much) was used and the items were slightly adapted to fit the context of the current experiment: “I could imagine using the AI banking assistant in the future.” and “I would like to be informed about products that are similar to the AI banking assistant.”. The reliability of the items was high, with a Cronbach's α of 0.804.

Human-likeness of the AI banking assistant, which refers to the extent to which a system resembles human attributes, behavior, appearance, or cognitive capabilities, was assessed with five items on a five-point semantic differential scale [e.g., 1 = synthetic, 5 = real; 1 = mechanical, 5 = organic, adapted from Ho and MacDorman (2010), which yielded an excellent reliability with Cronbach's α = 0.832.]

Uncanniness of the AI banking assistant, which describes an unsettling sensation or a sense of discomfort and creepiness (e.g., Mori, 1970; Mori et al., 2012; due to the blurring of human and non-human traits referring to the uncanny hypothesis), was measured with three items on a five-point semantic differential scale [e.g., 1 = scary, 5 = comforting, as example of an inverse coded item, adapted from Ho and MacDorman (2010), Cronbach's α = 0.837].

Agency, consisting of two items, was used as a manipulation check for the different agency levels (low/high) of the two conditions (e.g., “The AI banking assistant has the ability to act in a self-controlled manner”; Cronbach's α = 0.725). A 5-point Likert-scale was used.

We used the two dimensions of the Big Five personality traits, openness to experience, which refers to an individual's inclination to engage with novel ideas, experiences, and intellectual or artistic pursuits (Neyer and Asendorpf, 2018) and neuroticism, which is characterized by a tendency toward experiencing negative emotions such as anxiety, depression, moodiness, and emotional instability (Neyer and Asendorpf, 2018). Both scales were explained by three items each, based on the 15-items short-scale (5 dimensions) from the Socio-Economic Panel (SOEP, see Schupp and Gerlitz, 2014). A 5-point Likert-scale was used. Internal consistencies were good (openness to experience: Cronbach's α = 0.708, neuroticism: Cronbach's α = 0.713).

Propensity to trust, as a personality trait characterized by an individual's innate inclination to believe in the reliability and good intentions of others in various situations (Patent and Searle, 2019), was assessed with the help of three items (5-point Likert-scale) taken from the trust in automation model (Körber, 2019), showing acceptable internal consistency (Cronbach's α = 0.503).

After the participants wrote down three terms (e.g., nouns, adjectives) that described the AI banking assistant in an open comment field, two to three independent raters evaluated these terms. In total, 358 (23 missing) descriptions were recorded. The raters had to determine for each participant whether at least one of the three terms reflected specific categories (useful, automated, financial savings, artificial, uncanny, manipulative, dangerous).

These seven categories were created collaboratively post-hoc based on the most frequently occurring categories identified. Per category, each user-generated answer was assigned either 0 (if none of the terms matched the category) or 1 (if at least one of the terms matched the category). In cases where the raters did not agree, a decision was made collaboratively. The interrater reliabilities demonstrated substantial to nearly perfect agreement (Landis and Koch, 1977): Useful: κ = 0.727, automated: κ = 0.977, financial savings: κ = 0.832, artificial: κ = 0.895, uncanny: κ = 0.899, manipulative: κ = 0.725, dangerous: κ = 0.634.

Furthermore, the raters assigned a separate category for an overall impression (sentiment) to each test participant. They rated this category based on whether the terms used were only neutral, at least one positive, at least one negative, or a combination of both negative and positive terms. The sentiment was rated on a scale from −1 (negative) to 0 (neutral) to 1 (positive), depending on the terms used (14 ambivalent combinations including both negative and positive terms were excluded from the analysis, 5 VPN were additionally excluded due the fact that they didn't write down any descriptions). Subsequently, a mean score was calculated based on these evaluations, which represented the overall impression (sentiment) and was utilized for comparing the different user groups (low vs. high trust) in further analysis.

We conducted a chi-square test to determine whether there was a significant difference between the two groups (low trust vs. high trust) across the seven aforementioned categories (useful, automated, financial savings, artificial, uncanny, manipulative, dangerous). The percentage frequency of term occurrences for the categories useful, X²(1,96) = 11.068, p < 0.001, ϕ = 0.34, and dangerous, X²(1,96) = 7.634, p = 0.006, ϕ = −0.28, show significant differences between the low trust and high trust group. People who trusted the AI banking assistant perceived fewer risks and danger (5.7%) in using it, as opposed to the user group that distrusts the assistant, and thus used more terms to describe the potential dangers of its use (29.5%; e.g., “deceptive”, “suspect”). Similarly, the group that trusted the AI banking assistant regarded it as useful (45.7%; e.g., “optimization”, “workload reduction”) whereas a substantial number of those who distrusted the AI banking assistant viewed its assistance as less useful (14.8%) and, as previously stated, risky and dangerous (29.5%).

In the remaining five categories, no statistically significant differences can be found, whereby three of them reach a borderline area on an alpha level of 0.05: Automated, X²(1,96) = 2.942, p = 0.086, ϕ = 0.175; financial savings, X²(1,96) = 3.749, p = 0.053, ϕ = 0.198; artificial, X²(1,96) = 3.286, p = 0.070, ϕ = −0.185; uncanny, X²(1,96) = 2.079, p = 0.149, ϕ = −0.147; manipulative, X²(1,96) = 0.617, p = 0.432, ϕ = −0.080. Approximately 20% of the participants belonging to the low trust group perceived the AI banking assistant as uncanny, and around 34% associated terms such as “artificial” with the assistant. Conversely, in the high trust group, the number of participants who described the AI banking assistant as scary is just under 9% and only around 17% described the assistant as artificial. Furthermore, in the high trust group, ~31%—more than twice the number in the low trust group (15%)—associated the assistant with financial savings.

Comparing the mean sentiment score (overall impression) between the low trust (M = −0.48) and high trust (M = 0.39) groups also showed a significant difference. Specifically, participants who trusted the AI banking assistant exhibited a significantly more positive attitude (e.g., “practical”, “expectant”) toward the assistant compared to those who distrusted it. Conversely, the latter group demonstrated a negative attitude, as evidenced by their use of negative associations and terms toward the AI banking assistant (e.g., “unnecessary”, “dubious”).

In the subsequent analysis, we compared the two user groups low trust vs. high trust based on the scale indices of the variables. Wilcoxon-Mann-Whitney tests revealed significant differences in mean values for the following variables: Trust (U = 0.00, Z = –8.378, p < 0.001, r = 0.74), intention to use (U = 353.50, Z = –5.687, p < 0.001, r = 0.50), competence (U = 323.00, Z = –5.880, p < 0.001, r = 0.52), understandability (U = 313.50, Z = –6.016, p < 0.001, r = 0.53), intention of developers (U = 318.50, Z = –5.959, p < 0.001, r = 0.53), human-likeness (U = 631.00, Z = –3.592, p < 0.001, r = 0.32), and uncanniness (U = 615.50, Z = –3.713, p < 0.001, r = 0.33). Except for the variable uncanniness, the participants from the high trust group show higher mean values in all variables. In the variable uncanniness, the mean value from the low trust group (M = 3.52, SD = 0.85) is significantly higher compared to the high trust users (M = 2.71, SD = 0.96). All means and standard deviations including all participants (N = 127) can be found in Appendix Table 1. All means calculated separately for the two user groups (high trust vs. low trust) are shown in Figure 2.

There was no significant difference observed between the low trust and high trust user groups regarding the variables neuroticism (U = 965.50, Z = –1.118, p = 0.264, r = 0.10) and openness to experience (U = 1,108.50, Z = –0.056, p = 0.955, r = 0.05). The personality trait propensity to trust (U = 546.50, Z = –4.264, p < 0.001, r = 0.38), on the other hand, differed significantly between the two groups. Specifically, participants who displayed a high level of trust in the financial assistant also had a higher tendency to trust in general, as compared to those who displayed low levels of trust in the AI banking assistant. No significant differences were found between the low trust vs. high trust group in terms of the distribution of age (U = 972.00, Z = –1.064, p = 0.288, r = 0.09) and gender (U = 875.00, Z = –1.559, p = 0.119, r = 0.13).

To verify the effectiveness of our manipulation of the two conditions (high/low), participants were asked to rate the level of autonomy exhibited by the AI banking assistant using the agency scale. We computed the mean values and detected significant differences between the groups (U = 687.00, Z = –6.451, p < 0.001, r = 0.57), indicating a successful manipulation of the conditions. The high agency bot was perceived as a highly autonomous assistant, while the low agency bot was perceived as a less autonomous assistant.

However, a subsequent analysis investigating differences in levels of trust (U = 1,943.00, Z = –0.343, p = 0.732, r = 0.03) and intention to use (U = 1,943.50, Z = −0.339, p = 0.735, r = 0.03) showed no significant differences between the low and high agency condition. Furthermore, no significant differences were observed between the low and high agency conditions in terms of the other variables, competence (U = 1,822.00, Z = –0.928, p = 0.353), understandability (U = 1,860.00, Z = –0.753, p = 0.452, r = 0.07), intention of developers (U = 1,973.00, Z = –0.196, p = 0.845, r = 0.02), human-likeness (U = 1,922.50, Z = –0.439, p = 0.661, r = 0.04), and uncanniness (U = 1,722.00, Z = –1.414, p = 0.157, r = 0.13).