5.1 General discussion
The results show that Job relevance has a positive direct effect on Perceived usefulness, supporting H1. The results also indicate that Output quality, although less strong, has a positive direct effect on Perceived usefulness, supporting H2. The third and last construct of Result demonstrability showed a positive direct effect on the Perceived usefulness of AI systems as well, supporting H3. The strongest effect found was the positive direct effect from Perceived usefulness on the Behavioural intention to use AI systems, supporting H4.
An interesting observation within this research is the positive attitude from employees towards AI systems, where on a 7 point scale BI (𝑥̅ = 5.45, SD = 1.00) and PU (𝑥̅ = 5.24, SD
= 1.12) reported the highest sample means, followed by RES (𝑥̅ = 4.92, SD = 1.11), OUT (𝑥̅ = 4.75, SD = 1.20) and REL (𝑥̅ = 4.45, SD = 1.43).
The effect of Age was minor with a continuous effect of only -.08 or -09 on PU and BI, which indicates that employees with a higher age are slightly less positive on the perceived usefulness and usage behaviour of AI systems. However, this negligibly effect was not statistically significant (p > .05). The same applies for the moderating effect of Age which was between -.09 and .05 and also statistically not significant amongst the tested constructs (p >
.05). This is in contrast to prior research where a clear moderating effect of age was found on the acceptance on new technologies.
The study has shown that AI knowledge has a minor positive effect of .10 (p < .05) on BI. This indicates that employees with more knowledge about AI are slightly more willing to accept or use AI systems. At the same time, the moderating effect of KAI was calculated between -.02 and .09 and statistically not significant amongst the tested constructs (p > .05).
- 44 - The analysis in Appendix IV shows that employees with a high degree of customer interaction have a more positive attitude towards AI systems compared to employees with limited or no customer interaction, as the group mean is higher amongst all constructs, except for result demonstrability. However, the moderating effect of customer proximity was not found amongst the tested hypotheses (p > .05). This indicates that the acceptance of AI systems is expected to be similar amongst the various roles within banks. Whether an employee has a high degree of customer interaction (e.g. as a client advisor or within a customer contact center) or hardly no customer interaction (e.g. as a IT developer or controller within Finance), the results suggest that all employees will embrace the AI technology equally. This could indicate that a role-specific approach is not needed during the development and implementation phase to increase the employee acceptance of AI systems.
5.2 Theoretical and managerial contribution
The results of this study deliver both a theoretical and managerial contribution. From a theoretical point of view, there is a bundle of research that applied the Technology Acceptance Model on the acceptance of new technologies by employees. However, as far as can be ascertained, this is the first research to apply the Technology Acceptance Model on AI systems amongst employees and in particular within the Dutch banking sector. This research gap has been addressed and the findings show that the fundamental determinants of employee acceptance of AI systems, tested in H1 until H4, are supported.
Another gap that has been explored is the role of customer proximity as a moderating variable in the TAM. The results, as tested in H8, indicate that there is not a significant difference between the different levels of customer proximity.
A third contribution is the conclusion that AI knowledge only has a minor positive effect on usage behaviour and is not a moderating variable in the TAM. This is an interesting
- 45 - finding, because it is in contrast to prior research (e.g. Thong, 1999; Fountaine et al, 2019;
Eastwood & Luther, 2016; Yeomans et al, 2018).
The fourth and final theoretical contribution is the creation of a new item by the author (OUT2) within the construct of Output quality. This item validated the statement that AI systems perform tasks better compared to humans (Finlay, 2018; Burgess, 2018; Kaplan &
Haenlein, 2019) and resulted in a high reliability (α = .798) and inter-tem correlation (.669).
As addressed in the introduction of this study it is expected that the development of AI systems will increase in the future and this technology will have impact on employees and the way they will perform their tasks. From a managerial point of view this study delivers a few contributions. First, it has shown that employees in the banking sector have a positive attitude towards AI systems. The technology is considered to be useful for performing their tasks and when given the opportunity they are willing to use it.
Second, the finding that customer proximity does not have an effect on user acceptance indicates that the acceptance and usage behaviour is expected to be similar amongst the various roles within banks. This could indicate that a role-specific approach is not needed during the development and implementation phase to increase the employee acceptance of AI systems.
Third and final, this study has shown that AI knowledge only has a minor positive effect on the acceptance of AI systems. Therefore it does not seem mandatory for banks to develop large scale training and communication programs to increase AI knowledge among their employees, or at least it would not result in a significant return on organization investments.
- 46 -
5.3 Limitations
There are a few limitations to this study. First of all, this study is based on evidence from the Dutch banking sector only, which limits the generalisability of the findings to other sectors, industries and countries.
Second, the respondents are divided very unequal among the three groups of Customer Proximity. Although additional tests indicated that this was only statistically significant for Perceived Usefulness, the robustness of the statistical tests would have been improved if the groups were divided more equally.
Third is the limitation of knowledge self-assessment, as respondents indicate whattheir own perception of their knowledge is, but it doesn’t mean they actually have that level of knowledge. The Dunning-Kruger effect, named after the research of Dunning and Kruger (1999) is a type of cognitive bias that causes incompetent people to overestimate their own knowledge or ability. Similarly, competent people have a tendency to underestimate their own knowledge. It is uncertain if this effect is applicable in this research, but it is interesting to observe that only 2.4 percent of the survey respondents rated themselves a 9 or 10 as being very knowledgeable about AI.
Fourth limitation is that this research addressed the AI technology in general. As described in section 2.2, there is a wide range of AI applications already available within the banking sector and the control variable has learned that about a third of respondents (34.2 percent) has worked with (the output of) an AI system before. However, it is uncertain if respondents had a particular AI system in mind when answering the questions and if so, what kind of AI system (e.g. an internal chat bot from IT or an image recognition system).
This leads to the fifth and final limitation which is practically similar to all quantitative studies; it does not clarify why respondents give certain answers, neither does it captures their
- 47 - emotions and motivations. For example, a respondent could answer questions about usefulness and output quality negatively due to a disappointing experience with an internal chat bot, or positively due to a positive experience with an AI based invoice recognition system.
5.4 Suggestions for future research
Future research can address some of the limitations from this study. To test the generalisability, a similar research can be performed within the financial sector (e.g. insurance, pension funds) or other industries (e.g. retail, logistics) and a quota sampling technique can be used to ensure the different levels of customer proximity are represented sufficiently within the sample.
Further research can address the limitation of a quantitative study by conducting research with a more experimental design. This could discover more in-depth insights in the motivations, perceptions and possible barriers of AI acceptance by employees, especially if the sample includes experienced and unexperienced users of AI systems.
Future research can also include the level of end user involvement in the development and implementation of AI systems, as this will likely have a positive effect on the acceptance of AI systems (Reim et al., 2020). Software development is typically performed in the four stages of Design, Build, Accept and Production. The stage of Accept via a User Acceptance Test (UAT) means that an end user of the new system will test the developed system. Any defects or deviations from user requirement or expectations (ideally addressed in the Design phase) can be discovered and resolved before going into production.
- 48 -