Divisional perspectives on trust in generative AI: an empirical study of senior decision-makers

You’re listening to “Divisional perspectives on trust in generative AI: an empirical study of senior decision-makers,” by L.T. Do Khac and colleagues. Published in 2025.

ISSN: 1751-7575 (Print) 1751-7583 (Online) Journal homepage: the linked source

Lilian Tai Do Khac, Michael Leyer, Johannes Wichmann & Alexander Richter

To cite this article: Lilian Tai Do Khac, Michael Leyer, Johannes Wichmann & Alexander Richter (2025) Divisional perspectives on trust in generative AI: an empirical study of senior decision-makers, Enterprise Information Systems, 19:7-8, 2481557, DOI: 10.1080/17517575.2025.2481557

© 2025 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.

Published online: 27 Mar 2025.

Submit your article to this journal

Article views: 3461

View related articles

View Crossmark data

Citing articles: 13 View citing articles

RESEARCH ARTICLE

Divisional perspectives on trust in generative AI: an empirical study of senior decision-makers

Lilian Tai Do Khaca, Michael Leyera,b, Johannes Wichmanna and Alexander Richterc aChair of Digitalisation & Management, University of Marburg, Marburg, Germany; bSchool of Management, Queensland University of Technology, Brisbane, Australia; cSchool of Information Management, Victoria University of Wellington, Wellington, New Zealand

ABSTRACT.

ChatGPT opened a new chapter for using artificial intelligence (AI). Our results from 500 senior corporate decision-makers, focus on trust in introducing company internal generative AI applications. We find that decision-makers’ perspectives differ between organisational departments, particularly those that face each other. Counterintuitively, ability is the most important factor to explain trust for decision makers in sales departments while for non-IT departments in organisations integrity is most important. We contribute to understanding how technology is trusted in different divisions of organisations when integrated into internal data sources and processes. Our results help to support different divisions when introducing AI applications.

1. Introduction.

ARTICLE HISTORY

Artificial intelligence; generative AI; ChatGPT; trust; intention to use; decision makers

ChatGPT has opened a new chapter in the utilisation of generative artificial intelligence (genAI), a subcategory of AI, almost anywhere, regardless of the industry. Thereby, setting a significant milestone mark on the evolutionary ladder of digital services and products (e.g. T. Wu et al. 2023). Beyond public interest and individual usage, corporations are interested in implementing such genAI capabilities in their processes to exploit the full potential of their data. Since it requires a considerable amount of effort and detailed knowledge of AI, specifically in the realm of genAI, most companies need additional expertise or resources to develop such models independently (S. Wu et al. 2023).

However, the question is how senior management reacts to introducing genAI capabilities in the corporate setting for their subordinates. As for AI in a broader sense, studies have shown that a lack of trust is an essential factor responsible for no intention to use AI and therefore not delegating tasks to an AI. However, these studies focus typically on users and not on organisational decision makers who decide on the organisational introduction of AI tools such as ChatGPT. Therefore, we ask in this study:

Which trust factors of managerial decision-makers in organisations lead to the intention to use an AI in the organisation?

What are the differences of these trust factors concerning different functional units of organisations?

We address the two research questions by conducting a quantitative empirical study with 500 managerial decision-makers in German corporations. As a specific example of genAI, we use ChatGPT as it has become the most popular (generative) AI application recently. The questionnaire to collect empirical data was distributed in February/March 2023. The results indicate that the importance of trust factors differs between organisational units. This paper contributes to understanding trust in the organisational adoption of AI by focusing on senior decision-makers responsible for introducing genAI in their divisions. It highlights that trust factors, such as ability and integrity, vary significantly across departments, with practical implications for AI integration strategies.

The findings emphasise that fostering trust is crucial for AI adoption, guiding tailored organisational support and governance structures.

The paper is organised as follows: First, we describe the nature of genAI and AI lifecycles and examine the concept of trust in AI to achieve human-AI collaboration, specifically with genAI applications. We ground our analysis on an established trust model and derive our research model. We then operationalise our research model with an empirical study conducted among 500 decision-makers in Germany. We then present and discuss our results. This is followed by lessons learned regarding the introduction of genAI in different organisational units. The paper closes with implications regarding introducing genAI on the operational level between human and AI.

2. Background.

2.1. Generative AI.

Generative AI (genAI) is a result of recent developments, particularly large language models (LLMs), which demonstrate that modern AI models can be applied more broadly to a variety of tasks. Notably, the impressive performances on cognitive tasks, which demonstrate world knowledge and semantic recognition capabilities, showcase the new generation of AI’s potential to contribute significantly to the workforce. However, the ownership of such (gen)AI models comes with a high cost, mainly driven by size, management, and quality of datasets, model size in terms of parameters, and training volume. Furthermore, the availability of computing and hosting resources and the talent to create such models must be considered.

It seems economically evident that there would be a demand for genAI models, a new vulnerability vector, and open production of such models with consequences on their usage. An AI lifecycle, also applicable to genAI development, can be mapped with the technology development phases in accordance to the industry standard CRISP-DM (Cross-

Industry Standard for Data Mining). According to the ENISA (2021), there are six vulnerabilities across the AI lifecycle (see Figure 1) that can negatively impact trust. ‘Model data disclosure’ refers to processes that could lead to data or information leakages. ‘Poisoning’ and ‘Evasion’ are attacks on algorithms by manipulated data sets. ‘Compromise of ML application components’ refers to events such as the log4j zero-day attack, which abused a vulnerability in open-source applications. ‘Failure or malfunction of ML application’ refers to external factors that could be caused by users themselves or through ‘denial of service’ events caused by wrong inputs.

Furthermore, given the previous production cycle of AI, there are additional risks during operation, such as data security-related issues and particularly for genAI ethical and social risks.

2.2. ChatGPT.

Our investigation focuses on ChatGPT as the most popular example of genAI applications. ChatGPT is a software application developed by OpenAI and released for public research intentions at the end of November 2022. ChatGPT is built upon GPT-3 (at the time of the study) and was fine-tuned with deep reinforcement learning for human feedback (DRLHF) such that it can interact more naturally with humans. It is pre-trained with unsupervised learning on a large text dataset and has 175 billion parameters. Since its release, ChatGPT has been massively adopted globally by millions of users, which makes it more remarkable than previous AI releases. This kind of modern AI application demonstrated very good capabilities such that its utilisation was significantly adopted among various kinds of user groups.

ChatGPT was made available for free, and while using ChatGPT, a significant share of personal identifying information (PII) such as interaction content data, interaction location, or end-device information was collected and declared for possible further usage in creating new services or provision to other third parties (see

Appendix). PII data in Europe is subject to data privacy laws such as the General Data Protection Regulation (GDPR), and therefore rights such as the ‘right to know’ and ‘right to erasure’ along with appropriate data management and data governance measurements have to be in place. Since ChatGPT processed the PII data at cloud computing locations in the U.S.A., several EU member states initiated data privacy compliance investigations because of potential data privacy violations. In terms of business use cases, ChatGPT has been used in various companies to improve products and services. As such and in terms of best practices, the multinational energy company Uniper used ChatGPT to strengthen their internal audit function and determined that in terms of efficiency, a variety of their processes can be improved by 50 to 80%. Further, ChatGPT has different influences on companies’ departments.

As such and for human resource management, LLMs may reduce workloads for standardised processes that have to be personalised, such as writing policy documents on long-term absence due to illness or in constructing job descriptions. Hence, human resource departments require ChatGPT to provide meaningful documents for their daily work while the source of information is not that important. In contrast and for research and development, the requirements are different, as those departments also require ChatGPT to provide meaningful contents to be helpful, but also to use scientifically proven sources. Further, the database that is used to train ChatGPT for certain purposes is important, as, if the database is not sufficient, ChatGPT could provide incomplete or inaccurate information for maintenance departments.

2.3. Trust towards AI.

Mayer, Davis, and Schoorman (1995)’s organisational trust concept defines trust in the context of human workforce organisation within an organisation as ‘the willingness of a party to be vulnerable to the actions of another party based on the expectation that the other will perform a particular action important to the trustor, irrespective of the ability to monitor or control that other party’. During this procedure, the authors identify three trustworthiness factors of a trustee causing trust as significant to form a trust relationship with a trustor: ability, benevolence, and integrity. These are modulated through a trustor’s propensity to trust and the perceived risk and results into an outcome due to trust.

Mayer, Davis, and Schoorman (1995)’s has been broadly used and cited across multiple academic disciplines and in various sub-disciplines of business administration (marketing, accounting, finance, economics, supply chain management), proving its generally accepted applicability and validity. In the context of supplier assessment, this model has been utilised in Prior et al. (2022) as an evaluation framework for potential suppliers’ capabilities and to gauge the buyer’s firm perception of the supplier’s trustworthiness. It aims to capture and emphasise the trustworthiness factor in the supplier selection process, especially in high-risk contexts, where it is postulated to be of paramount importance.

Later, they reflected on their work and contribution to trust research and acknowledged the dimension of emotion to their model within the context of human workforce organisation within an organisation: ‘We think the role of emotions is a very interesting area of research and will add a new dimension to the model’ (Schoorman, Mayer, and Davis 2007, 349). Solberg et al. (2022) moved on to integrate and redefine the model of Mayer, Davis, and Schoorman (1995) with various trust research results to introduce AI as new workforce in the trust formation concept. Similarly, Lee and See (2004) described the dimensions of ability, benevolence, and integrity as AI’s performance, processes, and purpose. Highlighting ‘purpose’, according to Lee and See (2004), ‘[p]urpose refers to the degree to which the automation is being used within the realm of the designer’s intent’.

Whereas ‘integrity’ in Mayer, Davis, and Schoorman (1995)’s origin model was specified for the adherence to principles. Principles are being enforced through law-making, which we consider to be conceptually different from the intents of the respective designer. The respective designer would have to adhere to principles in data automation, such as data privacy regulations (e.g. GDPR, see previous section). For these reasons regarding general applicability and contemporary relevance, we will use the original Mayer, Davis, and Schoorman (1995) model as the underlying model for our hypotheses development in the following abstracts.

Given the reflection on the role of emotions to influence trust in Schoorman, Mayer, and Davis (2007) and Solberg et al. (2022), we acknowledge that emotions can be elicited by AI such that it impacts human trust, however an AI itself does not have a feature such as emotions. Therefore, we do not consider the concept of emotions for our investigation.

Next to AI, trust towards AI is important, as trust and AI, particularly for genAI, have become popular research topics recently. As such, Praveen and Vajrobol (2023) investigated whether ChatGPT can be trusted in the professional setting of medical consulting by uncovering physicians’ perceptions using deep learning techniques. They determined that physicians either do not care about those techniques or they perceive ChatGPT negatively for medical consulting purposes. Hence, Praveen and Vajrobol (2023) propose that medical consultations performed by ChatGPT are not trustworthy. In contrast, Ye, You, and Du (2023) determined that human-robot collaboration can be improved using ChatGPT.

They designed experiments in which individuals had to control a robot arm via human speech and determined that the robot arm performed better when ChatGPT improved the orders of the individuals compared to not using ChatGPT to improve orders. Hence, Ye, You, and Du (2023) state that ChatGPT benefits human-robot interaction, as human orders are improved concerning contextual understanding, conversational feedback, and robot controls. Thus, trust in human-robot collaboration is improved using ChatGPT. To investigate antecedents and consequences of travellers’ trust towards personalised travel recommendations offered by ChatGPT, Ali et al. (2023) conducted a survey. They determined that the relevance, credibility, usefulness, and intelligence of the personalised travel recommendations issued by ChatGPT significantly and positively influence intentions.

Further, perceived trust has a significantly positive influence on behavioural intentions. Using a similar approach by investigating creepiness and trust towards recommendations of ChatGPT, Hyun Baek and Kim (2023) determined that creepiness of recommendations lowers the continuance intention to use ChatGPT, while trust increases the intention. For high trust, it is important that the recommendations issued by ChatGPT increase task efficiency and are personalised for the users. To estimate the influence of trust towards ChatGPT to not only intention to use but also actual use, Choudhury and Shamszare (2023) performed a survey and showed that ChatGPT is mainly used for information gathering, entertainment, and problem-solving. Further, they determined that trust is an even more important predictor for actual use than intention to use is.

As a moderator, trust positively influences perceived enjoyment in

ChatGPT use by students. Nonetheless and for grading purposes, students do not trust ChatGPT to grade by itself, instead using it as a collaborative resource that require human oversight in college essay assignments. From what has been cited in the paragraphs before it can be derived that within the setting of professional collaborative work between human and AI an understanding of human suitable design of the AI is a factor that causes trust and then the intention to use AI.

In the related works we identified the lack of questioning senior decision makers on a divisional level to measure trust for the constellation of human and AI working together. Therefore, we aim to investigate how trust can be established within the context of human-AI interactions, empirically, by employing a well-established trust model. We focus on genAI as a particular subcategory and ChatGPT as a case study, exploring the crucial design features that contribute to building trust. However, the amount of design features for improving trust is very high as trust itself is a very complex construct. Our study is limited to what is possible to execute as empirical design and shall be taken as a first step to contribute answers towards trust building characteristics of an AI.

3. Research model.

For this work we focus on the genAI application ChatGPT as trustee and factors developed through the design of the genAI that cause the formation of trust and therefore leading to intention to use. Following the model of Mayer, Davis, and Schoorman (1995), we formulate our hypotheses to develop our research model within the execution boundaries of an empirical design as follows.

Riedl (2019) stated that a lack of understanding of how AI can be utilised would lead to its rejection. Cutillo et al. (2020), 1) define explainability of AI as ‘the ability to understand and evaluate the internal mechanisms of a machine, algorithm, or computational process in human terms’.. We are interested in whether users understand how genAI works and if this would lead to trust formation. According to Ali et al. (2023) usefulness has a positive influence towards trust in using ChatGPT. Hence, we hypothesise as follows:

H1 (+): The higher the perceived ability of a genAI application, the more positive trust formation is influenced.

Researchers are surprised that ChatGPT as a particular genAI is so successful, given that the underlying technology has been available since 2020. A potential explanation can be found in human-robot interaction research, which postulates that the inappropriate anthropomorphic design of an artificial agent would drop at a certain point, and humans would feel uncanny towards it. This proposition is supported by Hyun Baek and Kim (2023) who determined that increased creepiness of ChatGPT use lowers its continuance intention. In contrast, if the design is just right, humans would not feel uncanny towards it. In a more recent empirical study Agnihotri and Bhattacharya (2023) conclude that ‘[...] anthropomorphism enhances the benevolence and integrity of chatbots [...]’ (p. 1).

The underlying technology from ChatGPT did not have the human like response capabilities such as ChatGPT, which has been optimised to human references for interactions. According to Glikson and Woolley (2020), the features of an AI representation have implications for the antecedents of trust. According to the authors, they describe the representation of an AI through a continuum from fully embedded to a physical robot, of which a genAI like ChatGPT would fit the first type of AI representation. Further concerning students use of ChatGPT, Tossell et al. (2024) determined that ChatGPT is valuable for learning and the comfort of the students improved post use because of benevolent aspects of ChatGPT. Hence, our second hypothesis is formulated as follows:

H2 (+): The higher the perceived benevolence of a genAI application, the more positive trust formation is influenced.

We are interested in whether data privacy compliance of AI is associated with trust formation which is also considered as a relevant factor with AI applications. This mapping of data privacy and integrity was also investigated in Agnihotri and Bhattacharya (2023). Data privacy is a very important aspect when AI or data services are put into service on the market. This proposition is supported for ChatGPT use by Choudhury and Shamszare (2023) and Rahman et al. (2023). Hence, we formulate our third hypothesis as follows:

H3 (+): The higher the perceived integrity of a genAI application, the more positive trust formation is influenced.

Finally, we are interested in how trust leads to the actual intention to use a genAI application. This leads us to our fourth hypothesis:

H4 (+): The higher the trust formation concerning the genAI application, the more positive the intention to use the application is influenced.

Figure 2 summarises the hypotheses derived in the preceding paragraphs in our research model.

4. Research methodology.

4.1. Measures.

We considered each trustworthiness factor of Mayer, Davis, and Schoorman (1995)’s model to a measurement with respectively three measurement items. Further, we operationalised trust using four measurement items and intention to use by using three measurement items.

To measure the first independent variable, ability, we focus on the perception how much decision makers can assess how the results are derived, i.e. whether they understand the functionality of ChatGPT. Since a genAI has a multitude of abilities to produce text results, it is important for the ability assessment whether the software is providing correct answers. This explainability of an AI represents a large research field regarding trust in AI, (applied to ChatGPT) through the items X1, X2, and X3 (see Appendix). X1 is derived from Abdul et al. (2018) and Tan et al. (2018), who investigated the relationship of transparency in AI systems, thus knowing how ChatGPT works. We derived X2 from Dzindolet et al. (2003), focussing on the understanding how ChatGPT produces results.

Item X3 is based on the works of Juan Manuel and Karin Rolanda (2021) and Hasija and Esper (2022) who investigated trust in AI results, i.e. how much the respondent can assess the correctness of the results.

We operationalise the second independent variable, benevolence, with the uncanny perception of ChatGPT. Uncanniness is a highly relevant concept to judge on the benevolence for AI applications especially for embedded instantiations like ChatGPT. For the measurement we used three reflective scales from MacDorman and Ishiguro (2006) that allow to reflect on the uncanniness of ChatGPT generating text.

To measure integrity as the third independent variable, we focus on the integrity of the software keeping the information from usage within the respective company. This is an important aspect for ChatGPT as unstructured company information is entered in the software and might be transferred to OpenAI servers. For this, we relate to Smith, Milberg, and Burke (1996) and Preibusch (2013) and formulated the items by the underlying evaluation context.

The mediating variable trust with regard to ChatGPT is measured, referring to Mayer and Davis (1999). The items are adapted to the specifics of ChatGPT as an automated tool that might require human control of not trusted. The dependent variable, intention to use, is measured with the standard items derived from Ajzen and Fishbein (1975).

In addition, we used the control variables gender, age, and experience with ChatGPT (yes or no). All items can be found in the Appendix. Within the questionnaire, we randomised the order of the questions to prevent the order effects of answering by participants. The complete questionnaire can be found in the Appendix.

4.2. Procedure and participants.

Five hundred managerial decision-makers from various industrial sectors (see Appendix) in Germany on a mid-level within organisations were queried. They were polled through a company panel through a market research company between 14 February 2023 and 14 March 2023. The company panel allowed access to managerial decision-makers that are typically not available and represents a convenience sample. The size of the company panel determined the sample size of 500. The distribution regarding five business functions was 39% (n = 69) in corporate management/development, 19% (n = 95) in customer-and service management, 39% (n = 194) in IT, 14% (n = 71) in product development and marketing, and 14% (n = 71) in sales.

The sample is characterised by 357 (ca. 71.4%) male participants, 141 (ca. 28.2%) female participants, and 2 (ca. 0.4%) diverse participants who are currently full-time employed. The male participants were, on average, 40 years, and the female participants had 39 years on average (age ranged between 18 and 64 years). About 30,4% (n = 152) of the participants did not experience ChatGPT, while about 69,6% (n = 348) had experience with ChatGPT.

4.3. Model verification.

Our data analysis utilised partial least squares, a suitable method for processing structural equation models. Following the recommendations of J. Hair et al. (2022) concerning multigroup analysis in PLS-SEM, we examined five groups: corporate management/development, customer- and service management, IT, product development/marketing, and sales. To conduct this analysis, we used SmartPLS 4.0.9.2 and conducted a bootstrapping procedure and permutation tests with 5,000 resamples. Our analysis proceeded in two steps. First, all reflective variables met indicator reliability criteria, as the loadings exceeded 0.7 (see Appendix). Composite reliability was confirmed, as values were above 0.7, and the average variance extracted was higher than 0.5 (see Appendix; J. F. Hair et al. (2019)).

Discriminant validity using the heterotrait – monotrait was also confirmed, as scores were less than 0.9 (see Appendix; Henseler et al. (2014)). Second, the structural model was checked for model fit using the standardised root mean square residual. The values for the saturated SRMR (0.093) were less than the required 0.1. Further, our model is not affected by multicollinearity, as our VIF values for inner and outer model are below 3 (see Appendix, J. F. Hair et al. (2019)). To investigate effect sizes, we adapted criteria from Cohen (1988) for small, medium, and large effect sizes that we adjusted using recommendations of Kock and Hadaya (2018) using G.Power 3.1.9.7. Thus, we aligned the f2 thresholds due to the sample sizes of our groups.

Hence, the f2 -values are.009 (corporate management/development).006 (customer- and service management).003 (IT).008 (product development/marketing), and.008 (sales). We also followed the MICOM procedure according to J. Hair et al. (2022) to check for measurement model invariance and determine which variables were meaningful between all groups. We found that, besides determining significances within groups, none were significant in determining differences between groups. As a final step for our statistical tests, we performed PLS predictions for the groups using PLS prediction according to Shmueli et al. (2019) and CVPAT according to Liengaard et al. (2021). While PLS predicts and addresses early antecedents, CVPAT determines the direct influence on the dependent variable. Thus, both indications should be used to determine the model’s predictive power.

Concerning PLS prediction, our PLS-SEM Q2 prediction values were exclusively positive, and the prediction errors were not highly symmetrically distributed. Therefore, we tested whether our PLS-SEM means absolute errors (MAEs) were lower than those of the linear regression model (LM). We thus determined the following predictive powers for our groups: weak (corporate management/ development), weak (customer- and service management), weak (IT), weak (product development/marketing), and strong (sales). Concerning CVPAT, we investigated indicator averages and average loss values of the linear models of the dependent variable. Thus, we determined the following predictive powers for our groups: medium (corporate management/development), medium (customer- and service management), medium (IT), medium (product development/marketing), and strong (sales).

To verify whether our analysis is subject to common method bias, we conducted Harman’s single factor test. The results of the test reveal that the explained overall variance of sums of squared factor loadings for extraction is 30.929%, hence below the threshold of 50% showing that our study is not subject to common method bias (see Appendix,).

4.4. Ethical considerations.

All procedures performed in our survey involving human participants were in accordance with the ethical standards of the 1964 helsinki declaration and its later amendments or comparable ethical standards. As stated, the data was collected anonymously. Further, we have obtained informed consent from our participants that we may use the survey data for research purposes. We have ensured that our data analysis and presentation are exclusively anonymised and therefore no conclusions can be drawn about certain individuals. Since the data gathering did not involve any personal information and was a non-interventional study, no ethical approval was required according to German national laws.

5. Results.

Table 1 provides an overview of the mean and standard deviation values of the variables in the model per organisational unit identified. The results for intention to use were beyond average for all organisational units while the highest value emerged for the IT unit. In line with this result, the IT unit also displays the highest value for trust. Among all independent variables, those measurements for ability and integrity displayed high values, whereas, for benevolence, it was low for all units.

Figure 3 provides an overview of the research model results about the organisational units. We report the details of the results for all four hypotheses of the respective organisational units in the Appendix.

Trust describes approximately between 41% (customer- and service management and sales) and 56% (product development/marketing) of variability for intention to use and confirms former outcomes from research activities such as Meske and Bunde (2023) and Hamm et al. (2023). For all organisational units, the factor loading between.647 and.751 was strong and significant. For organisational units that are less likely to be direct customer-facing (corporate management/development, product development/marketing, IT), the factor loading was lower (approx. 53%, 47%, 50%) than those with direct customer-facing (customer- and service management and sales with approx. 58% and 65%).

All independent variables except for the chosen one for benevolence contributed to explaining approximately 47% to 65% of variability for the mediating variable trust. However, there is a difference among organisational units in which independent variables contributed more to the mediating variable trust. The variable integrity was associated with a strong significant loading for the departments corporate management/development (.499), customer- and service management (.488), and product development/marketing (.466). In contrast, the independent variable ability was more significant for IT (.437) and sales (.556). The significance of integrity was less for sales (.344) than for IT (.389).

In addition, we did not find any significant values with the control variables and thus excluded them from the analysis.

6. Discussion and theoretical implications.

The results show that for using ChatGPT, decision-makers across departments consider trust important as a requirement for their intention to use ChatGPT and thereby confirming formerly conducted research of Meske and Bunde (2023) for hate speech detection and Hamm et al. (2023) for explainable AI. Further, our insights align with propositions by

Ali et al. (2023) investigating travellers and Hyun Baek and Kim (2023) as well as Choudhury and Shamszare (2023) surveying users in general that trust has a positive influence on the intention to use ChatGPT. Particularly for the product development/ marketing department, trust was strongly associated with intention to use. In contrast to this department, the results for customer- and service management and sales were on the lower end. On a counterintuitive end the relationships are flipping when comparing the results of trust for product development/marketing and sales against the results of intention to use. Within these organisational units the lowest capture of variability on R2 of trust could be found for product development/marketing. However for the same organisational unit the highest capturing of variability on R2 of intention to use is depicted.

Regarding the organisational unit sales, it is the other way around. First, it seems that the selected variables do not explain trust for product development/marketing as well as for sales. It might be that for product sales, which can be less distinctively technical, characteristics of an AI related to the principle of ‘what you see is what you get’ are more effective. This observation is supported by the measurement for ability through explainability, which was stronger for sales versus product development/marketing. Furthermore, the measurement for integrity through data privacy, which is a security design, was stronger for product development/marketing versus sales. For future research therefore it might be interesting to contrast safety related aspects such as robustness regarding these two organisational units.

This relationship indicates a contradicting belief between two organisational units, product development/marketing (creating and delivering products and services) and sales (selling products and services), that could cause a lot of friction when AI service procurement is decided on a central level.

Further, the results show two distinct patterns for IT and sales on the one hand side and for the other three departments on the other hand side.

IT and sales departments value ability relatively higher than integrity. In contrast to the other three departments, corporate management/development, product development/ marketing, and customer- and service management value ability relatively lower than integrity. Sales have, however, a much lower intention to use than IT, which is rooted in relatively lower assessments of ability and trust as predictors. Individuals in the sales department determine ability as more important, which can be explained by their closer connection to customers and the competitive environment.

It seems that decision-makers working in IT departments have understood the possibilities of ChatGPT better, which is typically the case, as individuals working in these departments have a better understanding of information systems in general. As such, they see fewer issues with ability. However, the lower values for perceiving ability can be assumed to be traced back to having fewer insights into how ChatGPT functions.

The other three types of organisational units find integrity relatively more important. This can be explained as these departments are more critical areas in organisations in which innovations are developed that contribute substantially to competitive advantage or risks such as reputational risks affect their decisions directly. Hence, individuals working in these departments see data security as more critical, as many questions are still open. Thus, our results for corporate management/development, product development/marketing, and customer- and service management, support propositions of Hamza (2023) in that data security is crucial for ChatGPT use.

Another explanation for the diverse outcome regarding integrity among the types of departments could be rooted in the share of accountability. It is, taking a value chain view, a supporting unit to the other units, whereas the other organisational units are related as having ownership over products or services. Therefore, they hold accountability and ownership through their functional association with the data and processes, whereas IT is responsible for providing tools. This difference is prominent in that regard that the IT unit predominantly executes AI-related application development and integration. However, ownership over the business processes lies in the core processes.

On the other hand, sales are incentivised differently and probably more on a short-term basis such that ability might count more for them as integrity. Naturally, in sales, someone would like to know more about their clients such that they can better influence them or make better purchasing proposals. According to the results, sales seems to value ability more which justifies the purpose, but put data security issues second. However, using genAI to gain more information about customers is risky, because they could provide harmful stereotypes in customer consumption concerning gender or ethnic identity. Hence, we support propositions of Weidinger et al. (2021) that genAI should be used cautiously for sales purposes.

Among corporate management/development, customer- and service management, and product development/marketing, ability was valued relatively lower than IT and sales. Further, the empirical analysis results display less significance in this dimension for corporate management/development and product development/marketing. These departments usually invest a large share of their thinking and time in strategic outcomes in which integrity and trust might play a very important role to endure such long-term motivations.

Customer- and service management and sales have a close and direct face to the customer. However, their values regarding ability and integrity are quite the opposite. This might be the case because customer- and service management would be involved during their daily business along a whole customer lifecycle and would thus need to trust their tools internally and externally. Sales, on the other side, is more short-lived as to customer- and service management because they would only be active on a specific point along the whole customer lifecycle.

Summing up, theoretical implications are first that the results indicate a contradicting belief between corporate management/development, customer- and service management, and product development/marketing and IT and sales, that could cause difference in perception and therefore dissonance when introducing new technologies. Second, we contribute to literature by providing empirical evidence regarding the trust in genAI from a senior management perspective which has not been targeted so far.

7. Practical implications regarding the introduction of AI from a.

cross-organisational perspective

The introduction of technology into businesses as well as to society is highly complex and with the emergence of genAI a more fundamental push towards a more interconnected socio-technical system is introduced. Given the operational potential that can be achieved within an organisation when genAIs are applied can define competitive advantage in the future digital market. Thus, it is highly important to foster and support the introduction of AI (in general) applications in an organisation. Given the different departmental assessments and paths leading to trust with an AI application, the actions for successful implementation and usage must be distinguished.

For this purpose, we use the dimensions of structural empowerment as these describe which organisational elements are necessary to enable a successful introduction of digital platforms with AI: Access to Information, Access to Resources, Access to Support, and Access to Opportunities. Access to Information on AI is expected to increase benevolence as explanations will be provided on reducing the gap of uncanny feeling. Access to Resources can be provided to be able to better assess the abilities of AI. Access to Support is necessary to reduce problems with integrity, e.g. the legal department will provide necessary documents regarding data security and liability. And access to Opportunities allows testing AI for new application areas and thus can lead to increase trust.

Concluding the empirical results and the discussion, we emphasize the following actions to support the successful introduction and usage of AI throughout the organization (Figure 4).

According to the empirical results, access to information is not relevant as benevolence has no effect on trust and thus the intention to use. Given the relevancy of the topic, most employees inform themselves about AI individually in most of the cases. Regarding the provision of additional resources, those organisational units should receive extra training, project budgets etc., which have a lower rating of abilities of AI. Understanding what AIs are capable of in general requires often support to transfer this understanding to the specific tasks and data available. Similarly, decision makers for those organisational units indicating lower ratings of integrity, especially legal and risk-related support should be provided.

This covers not only the provision of related documents that are individualised for the respective application, but ongoing support on answering questions and providing reports. Opportunities for learning and evolving require sandbox environments that need to be provided to allow for testing AI capabilities in new tasks and areas. Given the empirical results, especially corporate management and development as well as sales require such environments to have an increase in trust.

Moreover, if an organisation introduces a company specific AI version, the measures should be distinguished between voluntary and mandatory usage. In the first case, managers from the different departments can decide whether to introduce the application in their department. In contrast, the second case is rather related to whether a mandatory introduction is applied. AI can typically be seen as tools capable of replacing human tasks within augmentation and automation, depending on the type of activity. Hence, an organisation should foster new roles for decision-makers in terms of responsibility for mandatory usage while quite some responsibility for process and outcome remain in voluntary settings. In the case of automation, decision-makers’ interaction with software and designers should be trained for a voluntary introduction, which is not necessary in a mandatory setting.

From the results of the analysis at hand, we can draw the implication that different organisational units and their different perception of trusting AI products lead to a diverging awareness. A managerial risk mitigation approach would be to raise awareness among their employees through training (sensitisation) and applying AI governance on the whole supply chain.

Regarding data privacy, especially with the introduction of the GDPR in Europe, there is still a significant gap in training and awareness. This topic has been overly simplified and as a result remains misunderstood up until now. The first point to address for building trust is the oversimplification of data privacy and its implications. The ethical and privacy considerations surrounding AI technologies are multifaceted and directly correlated to data management practices and legal frameworks. It is thus crucial to relate these concerns to the underlying data management and cloud technology. In accordance with our sample, we have surveyed companies in Germany for which the EU GDPR is the applicable law.

Most of the cloud and AI technology is sourced through US companies, creating a unique friction between US data laws and European data laws due to the following: The implementation of the Patriot Act in 2001, a direct response to the tragic events of September 11, granted authorities in the United States extensive powers to access data processed by American companies, even when they are operating internationally with hosting resources outside of the United States (e.g. in Frankfurt am Main, Germany). This legal framework was further reinforced by the Cloud Act in 2018, which explicitly requires American companies to provide personal data upon request, also even when the data is processed outside of the United States.

With a broad scope, the European GDPR applies to all personal data processing activities, regardless of whether they occur within or outside the EU, as long as they involve EU individuals. Under the GDPR, individuals whose personal data is being processed have several rights to control and protect their information. Thus, personal data can only be processed with a legal basis and a valid decision. Furthermore, the security and the individual’s right to their data must be ensured. This kind of so-called adequacy decision has been a challenge for the U.S.A. from the EU’s perspective, as they have not always adhered to these principles. Currently there exists an adequacy decision between the US and EUR since Sommer 2023. However, there are existing data agreements in place, for example, between Canada, Japan, and Germany.

Indeed, this extensive background knowledge is often sparsely conveyed within companies, creating a need for further education and awareness. The second point to address is data privacy by design. In an incident involving Volkswagen (VW), it was made public that

VW had collected data about their end-customers and their vehicle usage patterns, including geospatial data, in a manner that was opaque to the end-user. Additionally, this data was stored insecurely. This incident highlights the importance of transparent data collection practices and the secure handling of sensitive information. A practical strategy for fostering trust and strengthening the relationship between trust and data management practices is therefore to implement privacy by design into any kind of data-driven application.

Recent breaches of trust involving genAI have underscored the imperative of trust-building initiatives to catalyse broader adoption and strategic alignment of AI products and services with organisational goals. It is evident that initiatives focused on enhancing and showcasing trust can wield significant influence over the acceptance and integration of genAI and AI technologies at large. One potent strategy to signal this commitment is through adherence to harmonised standards and third-party audits, culminating in relevant certifications. In the realm of AI, the ISO 42,001 standard emerges as a management system framework crafted to nurture trustworthy AI. A part of this standard is the identification, control, and mitigation of risks associated with AI, which we also discussed in Figure 1.

A handful of AI companies have embarked on a journey towards achieving ISO 42,001 certification, thereby demonstrating their dedication to quality in the development of AI, among other pivotal activities (see Anthropic (2025), IEC 2023).

Emotions are an integral and fundamental aspect of human nature, and this understanding has been successfully applied to the corporate domain. While we acknowledge the absence of emotional capabilities in AI, as discussed in Section 2.3, it is important to recognise the significant role emotions play in human-AI interactions and the implementation of AI technology. As emphasised by Rust et al. (2021), emotions are a crucial factor in these interactions. The emotional dimensions of trust, particularly in the context of AI, and their influence on organisational culture and employee engagement deserve further investigation. By exploring these areas, future research can provide valuable insights into the emotional and psychological dynamics of AI interactions within corporate environments.

This knowledge will be instrumental in guiding the transformation of the workforce as we integrate AI technologies.

8. Conclusion.

Our study has shown how genAI is perceived and trusted by decision-makers in different organisational units. The results highlight the challenges of trust in AI adoption within organisations, emphasising the need for tailored training, governance frameworks, and sandbox environments. Those need to take different divisional backgrounds into account to enhance employees’ understanding of AI capabilities. Additionally, privacy-by-design approaches and adherence to trust-building certifications like ISO 42,001 can foster organisational trust in AI technologies.

We first contribute to understanding how new external technology that remains with the owner is accepted from a trust perspective to be integrated into internal data sources and processes. In distinction to other studies, we focus on the perspective of decision makers who would introduce AI in their area of responsibility. Organisational introduction decisions are a different level with many rules, compared to individual users who might adopt available software in the company or use genAI such as ChatGPT unofficially. Second, we contribute knowledge regarding the intention to decide on using genAI in organisational contexts. Our results emphasise that factors leading to trust differ depending on the type of organisational unit, which thus should be considered.

Third, we highlight that trust is an important factor for the intention to use such AI, as it explains around 50% of the variance.

Practical implications are that organisations should be aware of different perceptions of the importance of trust factors with the adoption of AI, particularly genAI. There is a need for integrating such AI applications in sourcing decisions to gain a competitive advantage. It is, however, important to convince individual decision-makers in companies even if a company acquires the software and the use is mandatory. Top management should be aware of the different factors and accompany a general introduction and availability for organisational units emphasising trust factors. Introduction in a company not necessarily results in usage in different divisions. Our results highlight that these differ and should be considered in organisational design support and communication.

The way forward for managers thus is to establish an AI Governance structure that supports the collaborative process of establishing AI-systems within the organisation. AI Governance can be composed into three distinct parts: data, model, and system. Given these three parts the artefact that needs to be governed, by whom, and with guidelines how to govern are laid out. Thus, AI Governance would establish structures and processes that foster mutual awareness, distribute accountability to respective roles, and build collaborative structures for trust by design in AI-systems.

As with any research, our study comes with limitations. First, while we could acquire a unique dataset promptly, the measurements for benevolence and integrity had some difficulties, and we ended up with single measures. Second, given the anticipated attention span of the targeted group, it was not recommended to operationalise more items with other variables that would contribute to explaining trust formation. Hence, the research model applied could not answer more of the variability. Further, another variable from the original trust model relating to a propensity to trust and perceived risk could not be investigated. Particularly, perceived risk could not be investigated due to the sample size and industry distribution. For example, only a few observations could be collected for life science.

In a similar manner, variables (e.g. benevolence or integrity) could be covering broader aspects as formative variables. Challenges of genAI, such as explainability and bias, warrant further investigation. The theoretical trust model chosen for this study is static in nature, contrasting the dynamic formation of trust, which necessitates the adoption of an alternative trust model. Dynamic trust, although an intriguing aspect, was not the primary focus of this research. However, it paves the way for future investigations to delve into the operationalisation of dynamic trust models, building upon the foundation laid by this study. By exploring these models, researchers can gain a deeper understanding of the complexities and nuances of trust formation and its implications. Lastly, we questioned mid-level managers not top management.

One could argue that the decision for using an AI like ChatGPT would be made solely by top management. Since mid-level managers are often responsible for divisions or functional units, they can however decide on the introduction of tools like ChatGPT in their area of responsibility. Finally, although the empirical data was collected during an optimal timeframe, that is during the experiment timeframe of OpenAI and looking backwards at a peak popularity point of time, the results presented are limited to what is possible to execute as empirical design. Therefore, only a couple of explanatory variables could be operationalised for the investigation, and this is reflected in the R2 value of trust that captures between approximately 47% and 65% of variability.

This outcome reflects that there are more variables that need to be investigated for completeness and that should be taken into consideration for future research. The same applies to the R2 value of intention to use that captures between approximately 41% and 56% of variability. Also, this outcome reflects missing parts that need to be investigated. The middle part of the underlying trust concept has not been investigated in this work which is propensity for trust and perceived risk and therefore suggested for future research as well. The strength of the empirical data though is the optimal point of time of its collection. The data was collected right before the first data privacy issues of using ChatGPT became known through the media and press.

Therefore, a unique asymmetric information awareness in that regard that users did not understand or know how their private data has been used while using ChatGPT could be captured.

Future research endeavours should focus on the critical role of top management in facilitating the adoption of genAI across diverse organisational departments. By applying resource dependence theory and incorporating qualitative insights from decision-maker interviews, future studies can investigate complex dynamics and priorities that shape the successful integration of genAI, offering valuable insights for effective organisational management. While this study provides valuable insights into the divisional perspectives on trust in genAI, it is important to acknowledge the limitations with regard to cross-sectional nuances. Future research should therefore investigate cross-sectional differences through longitudinal or experimental approaches. Furthermore, cross-cultural replications of this research could offer a better understanding of the phenomenon.

For instance, comparing EU firms (or Canadian firms) versus US counterparts (or Chinese counterparts) may reveal intriguing variations in divisional perspectives on trust in genAI. By addressing these research gaps, scholars can enhance the generalisability and theoretical contributions of this body of work.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Funding

The work was supported by the Adesso SE.