Introducing Artificial Intelligence in Government Organizations:
Employee perceptions and professional work engagement
Joey Stevens
Student number: s2031671 Master Thesis
MSc. Public Administration – Public Management Faculty of Governance and Global Affairs
Leiden University The Hague, The Netherlands
11 June 2019
2
Abstract
Considered a key driver of Industry 4.0, Artificial Intelligence (AI) is seen as a form of self-managing expertise that affects public and private organizations in many facets.
Governments increasingly turn to AI to boost efficiency, enhance performance and maintain professionalism in an era of increasing demands and complex public-sector problems. The introduction of AI in government organizations brings a series of consequences for
government workforces as the nature of work radically changes. It is argued that these consequences can be grouped in two main processes. The first process outlines that AI enables public professionals to achieve their tasks faster and with better results, resulting in higher work engagement. The second process captures the much-feared notion that AI automates tasks and that AI thereby - albeit to varying extents - replaces professionals, resulting in lower work engagement. The thesis builds on Bakker and Demerouti’s (2007) state-of-the-art Job Demands-Resources model, work engagement theory, as well as job characteristics and public service motivation (PSM) theory. The theoretical framework is applied to a multiple-case study of six government organizations at the local, regional and national government levels in The Netherlands. It is found that the way in which AI is introduced has considerable impact on how government professionals perceive AI and on how their motivation and work engagement is affected. In this process, job characteristics and PSM prove to play a key role. Therefore, this study provides useful insights and
recommendations for public management and the consulting sector when introducing AI in government organizations.
3
Foreword
This thesis is the final proof of competence for obtaining the Master’s degree in Public Administration, with a specialization in Public Management, from Leiden University, Faculty of Governance and Global Affairs in The Hague, The Netherlands. The executed research focuses on the influence of Artificial Intelligence in Government Organizations on
Professional Work Engagement. The main reason for choosing the thesis’ topic centres on the notion that in the next decades, the nature of government administration will
fundamentally change with the introduction of AI. By writing on AI at this moment in time, one acquires solid expertise on how future governments will operate, which is likely to be particularly beneficial for one’s future career in government, consulting, public affairs or any other (AI-) affiliated area.
The research for this thesis has been conducted in the second semester of the 2018-2019 academic year under supervision of Dr. J. Christensen, Assistant Professor at the Institute of Public Administration at Leiden University (The Hague Campus). I would like to take this opportunity to thank Dr. J. Christensen for his guidance and critical insight of the paper’s theoretical framework. Above all, I would like to express my gratitude to my parents and my friends for their unconditional support and trust. Furthermore, I would like to thank the ECORYS Group for offering me the possibility to gain valuable insight and expertise in public policy consulting in the course of my job roles as an assistant to the CEO, Assistant Financial Control and Assistant to the Board of Management.
The Hague, The Netherlands, (11-06-2019)
4
Table of Contents
Chapter 1: Introduction ... 6
Chapter 2: Theoretical Framework ... 12
Literature Review ... 12
Artificial Intelligence ... 12
The introduction of AI in Government Organizations ... 14
AI and public professionals’ work engagement ... 16
Theoretical Framework ... 23
Process 1 ... 24
Process 2 ... 24
Model including JD-R theory, Job Characteristics Theory, PSM Theory and Work Engagement theory ... 27
Expectations ... 28
Conceptualization into variables ... 29
Chapter 3: Research Design ... 31
Research methods ... 31
Threats to inference ... 34
Level of analysis, relevant populations and case selection ... 35
Operationalization of concepts ... 36
Chapter 4: Empirical Analysis ... 42
4.1 Empirical findings ... 42
4.1.1 The Introduction of AI in Government Organizations ... 42
4.1.2 Employee satisfaction ... 47
5
4.1.4 Motivation and Job Strain ... 55
4.1.5 Work Engagement ... 61
4.1.6 Moderating effect of Job Characteristics ... 68
4.1.7 Moderating effect of PSM ... 69
4.1.8 Scenario without AI ... 70
4.2 Empirical Analysis ... 73
Truth Table... 74
Expectation 1: The introduction of AI as a job resource is perceived as an abundance of job resources ... 76
Expectation 2: The introduction of AI as a job resource is perceived as a reduction of job demands ... 76
Expectation 3: The abundance of job resources causes increased motivation, characterized by low cynicism and excellent performance ... 77
Expectation 4: The reduction of job demands causes job strain ... 77
Expectation 5: Increased motivation results in increased work engagement ... 78
Expectation 6: Job strain causes a decrease in work engagement ... 79
Expectation 7: Job characteristics mitigate the negative effect of job strain on work engagement ... 80
Expectation 8: Public Service Motivation mitigates the negative effect of job strain on work engagement ... 81
Scenario without AI ... 82
Chapter 5: Conclusions ... 84
References ... 88
6
Chapter 1: Introduction
Artificial Intelligence is, according to the World Economic Forum (2019), widely considered as a software engine that drives the Fourth Industrial Revolution. Also known as Industry 4.0, the Fourth Industrial Revolution contains a number of disruptive technologies in a variety of fields that, besides AI, also include nanotechnology, biotechnology, the internet of things (IoT) and driverless vehicles. Artificial intelligence (AI) is a key disruptor among these 21st century forces. Seen as a form of self-managing expertise, AI disrupts both public and private organizations in many facets. AI is predominantly developed and applied in the private sector, resulting in the fact that governments lack significant expertise on AI. When AI is implemented, the power in terms of change is often considered to be transformative.
AI can centrally be defined as “a system’s ability to correctly interpret external data, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation” (Kaplan and Haenlein, 2019, p. 15). To this extent, AI mirrors cognitive functions of the human mind in terms of thought, self-improvement, learning and problem-solving (McCarthy et al., 1995). More practically speaking, AI takes predominantly form in terms of algorithms tied to one or more domains, such as Natural Language Processing and Object Recognition (van Balen,Interview, May 6th, 2019).
The human-like functions of AI have potential and risk, as among others outlined by Kaplan and Haenlein (2019) in the context of universities, corporations and governments. In pointing out the potential of AI for government, Kaplan and Haenlein (2019) present examples where governments make tasks more efficient with the use of AI. At the same time however, examples inhibit scenarios where the bad side of AI becomes obvious (Kaplan and Haenlein, 2019). Although AI will allow a multitude of tasks to be conducted faster, better, and at lower cost, it should, according to Kaplan and Haenlein (2019), be noticed that more tasks are being outsourced to AI and that employees will have to adapt.
The introduction of AI in government organizations is not new however, and although its application has been researched and used in defence and intelligence after its invention, AI has also more generally been used since the 1990s to reduce burdensome tasks (Mehr,
7
2017). Moreover, AI’s most obvious and beneficial opportunities centre on reducing
administrative burdens, helping in resolving resource allocation problems and the uptake of complex tasks (Mehr, 2017). Therefore, the introduction of AI can make government
organizations “work more efficient while freeing up time for employees to build better relationships with citizens” (Mehr, 2017, p. 1).
The introduction of AI in government organizations can be classified in three evolutionary stages: reactive automators, adaptive assistants and autonomous imaginators (Tinholt et al., 2017). The first evolutionary stage - reactive automators - consists of AI solutions aimed at process automation, where AI would monitor, analyze and act upon rule-based
programming (Tinholt et al., 2017). In this stage, AI performs a process or processes to accomplish a function or wider workflow. The second stage - adaptive assistants - combines conscious aspects of the previous stage with contextual sensitivity, thereby adding
interaction, memory and anticipation to AI solutions (Tinholt et al., 2017). In this stage, meaning brought by AI (partially) depends on context. The third and final stage -
autonomous imaginators - combines the two preceding stages with emotional awareness, moral reasoning and creative imagination (Tinholt et al., 2017). This final stage makes AI systems independent AI beings instead of repetitive systems respectively prevalent in the first and second stage (Tinholt et al., 2017).
So how does the introduction of AI influence government organizations? First, it should be underlined that the application of AI in government organizations is situated in the first and second evolutionary stages (Tinholt et al., 2017). In this regard, there is, according to Tinholt et al. (2017), economic potential in terms of increased productivity, efficiency and
employment (through the creation of new professions). There is also potential for improving public service quality in terms of service personalisation, service reliability and accuracy and service equality and availability (Tinholt et al., 2017). The downside of the coin however centres on inequality biases from public professionals on the development of AI systems, employment status of public professionals as well as challenges for our understanding of accountability (Tinholt et al., 2017).
8
As the reader might have noticed, both sides of the coin stress the importance of public professionals in the introduction of AI. A report titled ‘Arbeidsmarktanalyse Rijk 2018-2025’, written by Ecorys NL in collaboration with Dialogic (2018), features research on the
implications of technological developments on public professionals’ functions and
competencies of government organizations at the Macro (state-level), Mecro (organisation-level) and Micro (individual (organisation-level). The main conclusions of the report confirm Tinholt et al.’s (2017) claims in terms of general technological reasoning, as new employees are reportedly needed to deal with the introduction of new technologies, that certain employees receive a more specific package of dossiers to control and that technologies may implement tasks that were previously performed by employees (Ecorys NL and Dialogic, 2018).
Besides the possibility for technologies to put employees out of work or to take over part of their tasks, Ecorys NL and Dialogic (2018) however also underline a possible complementary nature of AI. In scenarios where AI is complementary to one’s job, employees co-operate with technologies to make their work more productive, facilitating a move towards co-creation (Ecorys NL and Dialogic, 2018). Additionally, the report mentions the topic of a surplus of personnel in terms of time, referring to scenarios where technologies take over part of employees’ tasks, reducing their weekly 40-hour jobs (Ecorys NL and Dialogic, 2018).
However, one may raise the question as to whether it is even possible to argue on how AI influences the work of public professionals, as Tinholt et al. (2017) pointed out that AI has not fully evolutionized yet in government. Nonetheless, Eggers et al. (2017) outline four main approaches as to how the introduction of AI impacts public professionals and the way in which they perform their work. These approaches are (1) relieve, (2) split up, (3) replace and (4) augment and should be seen as not necessarily discrete but rather as allowing some overlap between them (Eggers et al., 2017).
According to Eggers et al. (2017), the relieve approach states that AI takes over mundane tasks and frees workers for more valuable work, allowing governments to focus on reducing backlogs or shifting workers to higher-value tasks. The split up approach states that AI breaks a job into pieces or steps and automates as many steps as possible and not just routine ones, thereby reducing tasks and leaving humans to do what remains (Eggers et al.,
9
2017). In the replace approach, technology is set to do an entire job whereas in the
augment approach, AI makes workers more effective by complementing their skills, enabling employees to achieve faster and better results (Eggers et al., 2017). By taking into account all four approaches, one can identify that in three out of four approaches, AI can be seen as the introduction of - albeit to different extents - self-managing expertise that can be
beneficial to public professionals.
Besides redesigning work, Fast and Horvitz (2016) outline long-term trends in the context of a rather general public perception of AI, citing engagement and optimism vs. pessimism as key measures in their research. Fast and Horvitz (2016) among others indicate that impact on work can be positive in terms of engagement in that it makes work easier and in that it helps us to make better decisions. On the negative side of engagement, Fast and Horvitz (2016) stipulate that employees may experience a loss of control, displacement of jobs and absence of appropriate ethics and lack of progress in AI.
It is exactly here where the thesis’ added value can be found. The thesis follows in the footsteps of Ecorys NL and Dialogic’s (2008) study on the influence of technological developments on functions and competencies in government organizations. From an academic and societal point of view, research has been conducted on the evolution of AI in government as most notably outlined by Tinholt et al. (2017), on general potential and risks of AI in government as o.a. outlined by Kaplan and Haenlein (2019) and on classifications on how AI influences the way in which public professionals perform their work (Eggers et al., 2017).
However, research has not previously focused on the influence of AI on public professionals’ work engagement, referring to professionals’ work-related state of mind in contrast to a general perception of AI as outlined by Fast and Horvitz (2016). In this regard, significant academic value can be captured by formulating and positioning the thesis’ research
question on the intersection of the above-mentioned literatures whilst adding HRM-theory. I therefore decided to formulate the following research question:
What is the influence of the introduction of AI in government organizations on public professionals’ work engagement?
10
In answering this question, a qualitative comparative analysis of six cases will be made. These cases consist of government organizations at the national, regional and local government levels in The Netherlands.
The goal of this research centres on retrospectively explaining the effects of introducing artificial intelligence in government organizations on public professionals’ work
engagement. In explaining these effects, I define the introduction of AI in accordance with Kaplan and Haenlein (2019) as the practical application in government organizations of “a system’s ability to correctly interpret external data, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation” (p. 15). I secondly define public professionals as those people engaging in specified professions (e.g. policy-maker, application developer, border guard, etc.) serving some aspect of public interest and the general good of society at all levels within public, and more specifically, government organizations.
Third, work engagement is partly defined in accordance with Schaufeli and Bakker (2004) as a positive, fulfilling, work-related state of mind characterized by vigour, dedication, and absorption.
Next to academic relevance, the thesis’ research question provides for societal relevance, as government organizations may, on the basis of the thesis’ results, streamline their HR-processes at the individual and organizational level when introducing AI in order to improve organizational outcomes that directly affect society. General organizational outcomes are, among others, human capital and motivation, voluntary turnover and operational
outcomes.
The following chapters of this paper will respectively address the literature review and theoretical framework, research design, empirical analysis and conclusions. The final sections of the thesis also feature a list of references and overview of interviews.
In the next chapter, I review literature on definitions of AI as well as literature on HRM theories in order to select an adequate combination of both factors to build the thesis’ theoretical framework. In doing so, I arrive at a sound theoretical framework,
11
conceptualizing how public professionals perceive the introduction of AI and how this influences their motivation and work engagement.
In this section, I furthermore formulate a total of eight expectations that centre on the claims that AI, as a new form of expertise, adds to job resources, creates new job demands, reduces job demands and fully replaces jobs. I expect that these elements are consequential for employee satisfaction, stress, motivation and work engagement. These expectations will be tested with the use of Bakker and Demerouti’s (2007) Job Demands-Resources theory, job characteristics theory as well as Public Service Motivation theory.
The third chapter, featuring the research design and data collection methods, outlines the operationalization of variables, case selection, method of data collection, method of analysis, a reflection on validity and reliability as well as a reason on informed consent regarding respondents’ willingness to take part in interviews. In this regard, I furthermore explicitly describe the reasons for using the observational research design and the use of a multiple-case study design whilst addressing threats to inference.
The fourth chapter features two sections: the empirical findings and the analysis. The first section presents the findings of the research, including the results of the interviews as well as additional results obtained from document analyses. In the analytical section, I apply the theoretical framework to the research topic and explain the causal mechanism formulated in the research question.
The fifth and final chapter features the thesis’ conclusions. In this section, I summarize the thesis, provide a short and concise answer to the research question, and include self-reflection and key policy recommendations for government organizations on how to introduce AI focused on public professionals’ well-being.
12
Chapter 2: Theoretical Framework
Literature Review
Artificial Intelligence
The very term Artificial Intelligence (hereafter abbreviated as AI) was first coined by McCarthy et al. (1955) in their proposal for the 1956 Dartmouth Conference, the first conference on AI. McCarthy et al. (1955) proposed to conduct a study on artificial
intelligence in the summer of 1956 at Dartmouth College in Hannover, New Hampshire. The study was intended to explore new ways to make machines that could reason like humans thereby being capable of abstract thought, self-improvement and problem-solving and, as such, self-managing their own expertise (McCarthy et al., 1955).
The study was based on “the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it” (McCarthy et al., 1955, p. 1). In this study, the objective of creating AI was made, as McCarthy et al. (1955) outlined, and as such defined AI, by attempting to find how to make machines form abstractions and concepts, use language, improve themselves and solve kinds of problems that, at the time, were preserved for humans. In this regard,
intelligence is seen as the ability to acquire and apply knowledge and skills. When this ability is given to machines together with the ability to improve themselves, then machines
artificially produce intelligence.
Artificial intelligence was then founded as an academic discipline in 1956 and further developed for a variety of reasons (such as the identification of goals, problems, tools and other social factors) in terms of subfields (Crevier, 1993 and Press, 2017). These subfields include most notably natural language generation, speech recognition, virtual agents, text/data analytics, predictive analytics, machine learning, detection, computer vision, robotic process automation and decision management (Crevier, 1993 and Press, 2017).
13
The development of AI throughout history in a variety of subfields resulted in a large number of definitions, making AI, at times, a fuzzy concept (Crevier, 1993, Kaplan and Haenlein, 2019).
One of these definitions is provided by Dobrev (2005) who attempted to give a new definition of AI by stating that “AI will be such a program which in an arbitrary world will cope not worse than a human” (p. 2). This definition is based on the assumptions that (1) programs can model every calculating device, that programs (2) process external
information and that (3) they export (i.e. produce) information (Dobrev, 2005). Although a very central definition that most notably captures human-like thinking, it should be
underlined that specifications in the definition lead to substantial conceptual limitations. In comparison to McCarthy et al.’s (1955) focus on machines in defining AI, it can be pointed out that Dobrev (2005) limits his definition to (1) programs and (2) an arbitrary world, with the latter most notably referring to a world based on random choice and absence of any reason or system.
In contrast to Dobrev’s (2005) rather specific definition of AI, Kok et al. (2002) argue that instead of looking at one definition of AI, one may focus on definitions of AI systems that can be classified in four categories. These classifications centre on systems that (1) think like humans, that (2) act like humans, that (3) think rationally and that (4) act rationally (Kok et al., 2002). To a very similar extent, Nilsson (2010) defines AI as “that activity devoted to making machines intelligent, and intelligence is that quality that enables an entity to function appropriately and with foresight in its environment” (p. 13). This way, Kok et al. (2002) and Nilsson (2010) respectively focus on AI as part of systems and machines - and not just programs - whilst including elements of human-like thinking exemplified in ratio and reason. Although these definitions capture the essence of many other definitions, it should however be pointed out that these definitions still do not say anything about AI output.
A very central definition that captures AI output contextually can, however, be found in the work of Kaplan and Haenlein (2019), who defined AI as “a system’s ability to correctly interpret external data, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation” (p. 15). This definition mirrors cognitive
14
functions of the human mind as stipulated by McCarthy et al. (1995) in terms of thought, self-improvement, learning and problem-solving. Besides, Kaplan and Haenlein (2019) analyze how AI is different from related concepts and argue that AI is not to be seen as one monolithic term but instead should be seen in a more nuanced way. This can, according to Kaplan and Haenlein (2019), be achieved by viewing AI through (1) the lens of evolutionary stages (such as artificial super intelligence, artificial narrow intelligence and artificial general intelligence) or by (2) focusing on different types of AI systems (such as analytical AI,
human-inspired AI, and humanized AI systems).
The introduction of AI in Government Organizations
As previously outlined, Mehr (2017) argued that the introduction of AI in government organizations is by far not new, pointing out research on AI and subsequent application in the decades after the 1956 Dartmouth Conference. Mehr (2017) points out AI’s early application in the areas of defence and intelligence, after outlining AI’s general application since the 1990s. Regarding the latter, Mehr (2017) argued that AI was most notably applied to reduce administrative burdens, besides helping in resolving resource allocation problems and up-taking complex tasks in order to improve client services (Mehr, 2017). More
centrally, Mehr (2017) posits that the introduction of AI can make government organizations “work more efficient while freeing up time for employees to build better relationships with citizens” (Mehr, 2017, p. 1).
In line of Mehr’s (2017) arguments that AI can make governments more efficient, Tito (2017) adds that the extent to which governments can take advantage of the benefits of introducing AI “whilst meeting the challenges it brings, largely depends on how
policymakers act now” (p. 1). In contrast to Mehr (2017), Tito (2017) more specifically outlines what the challenges for governments are “that may prevent them from capturing the benefits of AI” (p. 1) whilst describing “the risk for governments of failing to act and of getting things wrong” (p. 1). Challenges respectively centre on legal, human and technical capacity, whereas risks centre on legitimacy, declining outcomes, systematized inequity, spiraling costs and abuse of government power (Tito, 2017). Furthermore, Tito (2017)
15
outlines a number of recommendations that governments can use to place themselves in a position in which they can capture the benefits of AI.
In the footsteps of Tito (2017), Desouza (2018) further delves into the challenges and opportunities of introducing AI in government. These can be categorized in (1) Technology and Data, (2) Workforce and (3) Risk Management and result, according to Desouza (2018), from “the design, building, use, and evaluation of cognitive computing and machine learning to improve the management of public agencies, the decisions leaders make in designing and implementing public policies, and associated governance mechanisms” (p. 40). In contrast to Tito (2017), Desouza (2018) presents an additional section on Risk Management, which involves securing systems, risk aversion, ethical and social considerations as well as issues in governance.
Like Desouza (2018), Whittaker et al. (2018) also capture the risks of introducing AI in government that, in the context of automated decision systems accordingly centre on a lack of due process, accountability, community engagement, and auditing. In contrast to Tito (2017) and Desouza (2018) however, Whittaker et al. (2018) present specific solutions to risks (and thus challenges) of introducing AI in government. These solutions most notably centre on bias busting, toolkits, fairness formulas and system improvements (Whittaker et al., 2018). These solutions are among others exemplified in observational fairness strategies that “diagnose and mitigate bias by considering a dataset” (p. 25) and most notably with anti-classification strategies that consider pretrial risk assessments that can omit protected attributes such as gender in delivering fair results (Whittaker et al., 2018).
What the above-mentioned authors all have in common however, is the way in which the introduction of AI in government influences the workforce and more specifically how employees engage in their work. Although Mehr (2017) noted that AI causes concerns that it can replace civil servants, Mehr (2017) also pointed out that AI works best in collaboration with civil servants and that it may free up time for employees to help them establish better relations with citizens. Tito (2017), on the other hand, pointed out that government
organizations face a lack of analytical talent needed for processing AI results, whereas Desouza (2018) stipulated that “AI and cognitive systems may result in some jobs being lost due to automation” (p. 26). On the other hand, Desouza (2018) also pointed out that AI “will
16
also increase the quality of work being conducted and open up new work opportunities” (p. 26). Additionally, Whittaker et al. (2018) presented evidence of scenarios where AI systems took over decision-making tasks from civil servants and outlined negative effects in terms of the relative lower quality of decisions made by AI and subsequent impact on citizens as such. In whatever way the introduction of AI in government organizations influences
workforces, AI will most likely change the way in which public professionals engage in work, which is theoretically more specifically outlined in the next section.
AI and public professionals’ work engagement Theories on work engagement
In measuring the effects of AI on public professionals’ work engagement, one should first notice that there is no unique theoretical framework on work engagement. Instead, there are a number of theories that emphasize different aspects of work engagement and often as part of larger theoretical claims. These are the needs-satisfying approach as proposed by Kahn (1990), social exchange theory proposed by Saks (2006), the JD-R model proposed by Bakker and Demerouti (2007), the Affective Shift Model of Work Engagement proposed by Bledow, Frese, Schmitt and Kühnel (2011). However, before we move on to what
determines work engagement, we review the literatures in order to address how work engagement is conceptualized.
Conceptualizing work engagement
Due to the large number of theories on work engagement, different aspects of work engagement have been emphasized, resulting in a wide variety of concepts that have often been used interchangeably. These concepts include employee engagement, job
engagement, organizational engagement and – more recently – a more specific concept of work engagement.
Work engagement was first conceptualized by Kahn (1990) in terms of employee
engagement and - more specifically - employees’ personal engagement towards their job role at work. In this context, personal engagement is defined “as the harnessing of organization members' selves to their work roles; in engagement, people employ and
17
express themselves physically, cognitively, and emotionally during role performances” (Kahn, 1990). It should be underlined that Kahn (1990) does not fully refer to work
engagement, but moreover refers to employee engagement at work in terms of their work roles.
Whereas Kahn (1990) moreover focuses on work roles in defining employee engagement, Saks (2006) presents a multidimensional approach, distinguishing employee engagement between job engagement and organizational engagement. Saks (2006) respectively defines the latter two concepts as engagement that is role related and as “the extent to which an individual is psychologically present in a particular organizational role” (Saks, 2006, p. 604). Although Saks (2006) presents a more elaborate conceptualization of employee
engagement at work by including organizational engagement in comparison to Kahn (1990), it should be pointed out that the difference between employee engagement and a specific use of work engagement is key.
Whereas employee engagement focuses on how engaged an employee is with the job and the organization he or she is working for (Kahn, 1990 and Saks, 2006), work engagement, according to Schaufeli and Bakker (2004), focuses on the work people do at the
organization, and the fulfillment gained from doing that work.
Schaufeli and Bakker (2004), in contrast to Kahn (1990) and Saks (2006) specifically use the concept of work engagement, defining it as “a positive, fulfilling, work-related state of mind that is characterized by vigor, dedication, and absorption” (p. 4). In this regard, work
engagement is irrespective of any momentary and specific state and instead refers “to a more persistent and pervasive affective-cognitive state that is not focused on any particular object, event, individual, or behavior” (Schaufeli and Bakker, 2004, p. 4 and 5). In this rather generalist approach, which may include both Kahn’s (1990) job factors and Saks’ (2006) organizational factors as well as other factors, Schaufeli and Bakker (2004) specifically measure work engagement through a self-reported questionnaire, as further outlined later in this thesis.
Besides Bakker and Demerouti (2007), Bledow et al. (2011) also focus their attention specifically on work engagement, defining it as “the presence of positive work-related
18
feelings such as happiness and enthusiasm while performing work tasks” (1247). Work engagement is, however, very narrowly conceptualized and measured with rather basic indicators that refer to negative moods, negative events and positive affectivity, in
comparison to Schaufeli and Bakker’s (2004) extensive self-reported questionnaire on vigor, dedication and absorption.
Determinants of work engagement
The above-mentioned academics that defined work engagement all developed theories in order to determine employees’ work engagement. Respectively, these are Kahn’s (1990) needs-satisfying approach, Saks’ (2006) social exchange theory, Bakker and Demerouti’s (2007) JD-R theory and Bledow et al.’s (2007) affective shift model. In addition, Hackman and Oldham’s (1976) Job Characteristics Theory and Wise and Perry’s (1990) PSM theory include elements that can influence motivation and, as such, work engagement.
Kahn’s (1990) engagement theory, also referred to as the needs-satisfying approach, stipulates that three dimensions need to be satisfied before employees become engaged: meaningfulness (the feeling of return on personal investment in one’s role), psychological safety (feeling safe in the work environment to bring one’s full self to the role) and
availability (referring to whether employees feel sufficiently mentally as well as physically able to engage the full self at work).
Kahn (1990) measured engagement in terms of personal engagement and disengagement, integrating “the idea that people need both self-employment in their work lives as a matter of course” (p. 694). Kahn (1990) more specifically measured both variables by focussing on respective psychological experiences of rational and unconscious elements of work
contexts. These work contexts accordingly create conditions that are mediated by employee perceptions, resulting in employees that personally engage or disengage (Kahn, 1990). Kahn’s (1990) results indicate that employees make decisions based on these conditions at the intersection of individual, interpersonal, group, intergroup and organizational contexts. Another attempt to theoretically capture Kahn’s (1990) concept of employee engagement is made by Saks’ (2006), who captured job and organizational engagement through social exchange theory. In this regard, Saks (2006) stipulated that employees repay their organizations in engagement in response to the resources they receive from their
19
organization. These resources are job characteristics, perceived organizational support, perceived supervisor support, rewards and recognition, procedural justice and distributive justice (Saks, 2006). Job engagement and organizational engagement are rather simply measured with two statements and the application of a Likert-scale in which participants indicate the extent to which they agree with the statements. For job and organizational engagement, these respectively are “sometimes I am so into my job that I lose track of time” (Saks, 2006, p. 608) and “one of the most exciting things for me is getting involved with things happening in this organization” (p. 608).
In scenarios where organizations fail to provide resources, individuals will, according to Saks (2006), disengage themselves from their roles. In distinguishing employee engagement between organizational engagement and job engagement, Saks’ (2006) findings indicate that resources relate more strongly to organizational engagement than to job engagement. In contrast to Kahn’s (1990) findings however, Saks’ (2006) findings indicate that both relationships remain weak.
Similar factors can be seen with regards to Kahn’s (1990) dimension of meaningfulness and Saks’ (2006) job characteristics of recognition and rewards and support and with regards to Kahn’s (1990) dimension of safety and Saks’ (2006) dimension of safety. Although both theories stress the need for resources for employees to become engaged, the key
difference centres on the fact that Kahn (1990) argues that resources need to satisfy three fixed dimensions in order for employees to become engaged, whereas Saks’ (2006) posits that employees simply repay their organization with engagement for resources they receive on an ad-hoc basis.
Taking a rather different approach than Kahn (1990) and Saks (2006), Bakker and
Demerouti’s (2007) Job Demands-Resources (JD-R) theory presents a model in which job demands and job resources are seen as two risk categories (Bakker and Demerouti, 2007). The dimensions of job demands are “physical, psychological, social, or organizational aspects of the job” (Bakker and Demerouti, 2007, p. 312). These essentially require physical and/or psychological effort or skills and physiological and/or psychological costs (Bakker and Demerouti, 2007). On the other hand, the dimensions of job resources are “physical,
20
312). These function in achieving work goals, reduce job demands and/or stimulate learning, growth and development (Bakker and Demerouti, 2007).
Bakker and Demerouti (2007) propose that the two risk factors are responsible for
imbalance between job demands on the individual and the job resources individuals have to deal with respective demands. In response to this imbalance, employees experience job strain, which is a common form of stress at the work floor (Bakker and Demerouti, 2007). Continued job strain leads to a health impairment process and lower work engagement, whereas abundant job and/or personal resources may lead to a motivational process, higher work engagement and positive outcomes (Bakker and Demerouti, 2007).
In contrast to Kahn (1990) and Saks (2006), Bakker and Demerouti (2007) explicitly refer to work engagement, which is measured according to three measures by the Utrecht work engagement scale. It should however be noticed that work engagement is not the final stage in the model, as work engagement in turn mediates the relationship between job demands and resources and positive outcomes (Bakker and Demerouti, 2007).
When comparing Bakker and Demerouti’s (2007) JD-R theory with the theories of Kahn (1990) and Saks (2006), it is clear that the latter two focus on job resources as determining work engagement, whereas Bakker and Demerouti (2007) view job resources as given and focus on the balance or imbalance with job demands. The imbalance or balance respectively leads to either job strain and lower work engagement or motivation and higher work
engagement. As such, there are a number of (mediating) factors that result in the level of work engagement in Bakker and Demerouti’s (2007) JD-R theory in contrast to a more direct focus on job resources causing work engagement in Kahn’s (1990) and Saks’ (2006)
approaches.
Besides Bakker and Demerouti (2007), Bledow et al. (2011) also focus their attention specifically on work engagement by presenting the Affective Shift Model. This model posits that a shift from negative affect to positive affect causes the emergence of high work engagement. Positive affect refers to an emotional status in which employees “set high goals for a task and expect that engaging in a task yields positive outcomes” (Bledow et al., 2011, p. 1247). Positive affect, through the initiation of goal-directed action, supports employees’ mindset distinctive of work engagement (Bledow et al., 2011). Negative affect,
21
on the other hand, refers to a negative state of mind that is not compatible with absorption, dedication and vigour in the context of achieving one’s tasks (Bledow et al., 2011). Negative affect, however, signals that things are not going well and that action needs to be taken, paving the way for a shift towards higher work engagement (Bledow et al., 2011).
Action substantiates the affective shift as employees notice that things are not going well and decide to act upon problems encountered (Bledow et al., 2011). Results indeed show that a shift from a situation of negative events and a negative mood to a situation of high-positive mood is associated with high work engagement (Bledow et al., 2011). Results also show that, when employees experience low positive affectivity, the more depending they are “on positively stimulating external events in order to become engaged” (Bledow et al., 2011, p. 1254). It should however be pointed out that work engagement is conceptualized in a rather narrow and abstract basis, namely as “the presence of positive work-related
feelings such as happiness and enthusiasm while performing work tasks” (Bledow et al., 2011, p. 1247). Furthermore, although work engagement is measured with the Utrecht Work Engagement Scale (UWES) at the start of the questionnaires, only 5 out of 8 items are used from the scale.
In comparing Bledow et al. ‘s (2011) Affective Shift Model, with the theories presented by Kahn (1990), Bakker and Demerouti (2007) and Bledow et al., (2011), it becomes
immediately clear that Bledow et al. (2011) do not view (low levels of) work engagement as static, pointing out the very reasons as to why low work engagement may shift to high work engagement. A similar factor can be seen in Bledow et al.’s (2011) use of the term tasks, which Bakker and Demerouti (2007) refer to as job demands. Bledow et al. (2011), however, take their model one step further, saying that in a positive affective mood, employees set high goals for tasks and even expect high outcomes for these tasks.
One may furthermore see correlation in terms of a positive affective mood and Kahn’s (1990) use of personal engagement in a context where employees perceive a positive environment. Both factors, however, differ regarding the fact that Kahn (1990) stresses the necessity for three static dimensions to be satisfied before employees engage, whereas Bledow et al. (2011) simply ask employees to what extent they identify with a number of
22
emotions and work events in order to determine the affective state and influence on work engagement.
In a slightly dissimilar way to the above-mentioned authors and without explicitly mentioning the term work engagement, Hackman and Oldham (1976) present the Job Characteristics model. This model centres on three classes of variables that determine how positive a person responds - i.e. engages in activities – to their jobs (Hackman and Oldham, 1976). These are employees’ psychological states that need to be present for internally motivated work behavior to develop, job characteristics that create these states and individual attributes that determine positive responses (Hackman and Oldham, 1976). In order to measure the three independent variables, Hackman and Oldham (1976) developed the Motivating Potential Score, which combines scores on the following five dimensions attributed to the variables: Skill variety, Task identity, Task significance, Autonomy and Feedback. In measuring the outcome variable, employee responses, Hackman and Oldham (1976) measure work motivation, work performance, high satisfaction and low absenteeism and turnover by asking employees’ relatedness on a number of statements included in questionnaires. Results confirm the hypothesis that “people who have high need for personal growth and development will respond more positively to a job high in motivating potential than people with low growth need strength” (p. 258), although weak results remain in terms of absenteeism and autonomy (Hackman and Oldham, 1976).
As previously outlined, Hackman and Oldham (1976) do not use any specific conceptual form of work engagement like Kahn (1990), Saks (2006), Bakker and Demerouti (2007) or Bledow et al. (2011) to measure how employees respond to their jobs. Nonetheless, generic similar factors in measuring employee responses may be identified in terms of motivation, performance and satisfaction. Among others, this is exemplified by Bakker and Demerouti’s (2007) JD-R model, which includes motivation, and as such to a certain extent satisfaction, as well as positive outcomes that can be seen as a type of performance.
What is more important, however, is that Bakker et al.’ (2004), albeit in a different article, take the outcomes of job characteristics (o.a. autonomy, social support and possibilities for self-growth) a step further. Bakker et al’ (2004) namely outline that job characteristics lead
23
to extra-role performance, which is defined “as actions that go beyond what is stated in formal job descriptions and that increase organizational effectiveness” (p. 91). In this
regard, the responses elicited by job characteristics do not simply lead to work engagement, but to something of a higher level, by which employees exceed performance as purely required by their job descriptions.
In addition, Perry and Wise (1990) coined the term Public Service Motivation (PSM) to outline why people engage in public service. Perry and Wise (1990) define PSM as “an individual's predisposition to respond to motives grounded primarily or uniquely in public institutions and organizations” (p. 368). In this regard, Perry and Wise (1990) present a typology of motives for public service that include rational, norm-based and affective motives. As these motives are grounded in public organizations, it becomes clear that the typology, as proposed by Perry and Wise (1990) can function as an important variable between the introduction of (AI) technologies and work engagement.
Taking into account the above-mentioned approaches, one could namely argue that PSM might still motivate employees to engage in work even though Kahn’s (1990) needs are not satisfied, if Saks’ (2006) job resources are not given to employees, if employees experience Bakker and Demerouti’s (2007) job strain, if Bledow et al.’s (2011) affective shift does not take place or if basic job characteristics, as outlined by Hackman and Oldham (1976), are not provided for.
Theoretical Framework
Taking the literature review into account, it is in this section that I establish the core theoretical framework, model and expectations.
I expect that a main causal relationship exists between the introduction of Artificial
Intelligence and public professionals’ work engagement. The main theory used in explaining this relationship is JD-R theory as proposed by Bakker and Demerouti (2007). In this regard, I expect that the introduction of AI causes two main processes:
(1): The introduction of AI as a physical workplace resource is perceived by employees as creating an abundance of job resources, which leads to a motivational process that can be
24
characterized by low cynicism and excellent performance, which in turn leads to higher work engagement
(2): The introduction of AI as a physical workplace resource is perceived as a reduction of job demands due to the automation of tasks, which leads to job strain, resulting in lower work engagement, which in turn can be moderated by job characteristics and PSM.
Process 1
Regarding the first process, I start from the point of view that public professionals perceive the introduction of AI as the introduction of a job resource – specifically a physical
workplace resource - that they use to achieve their job demands, which I refer to as job-related tasks. AI has namely proven to be able to be a positive resource for employees in that it can augment employees’ work by complementing employees’ skills and by helping them to achieve faster and better results (Eggers et al., 2017).
Next to this, AI can relieve employees from mundane tasks allowing employees to focus on more valuable work or higher-value tasks (Eggers et al., 2017). I therefore expect that the introduction of AI is seen as a significant increase of job resources with job resources, in the context of the JD-R model, therefore seen as abundant.
I then also expect that - in the footsteps of Bakker and Demerouti’s (2007) JD-R model - the abundance of resources instigates a motivational process, with motivation being intrinsic or extrinsic or both.
As argued by Bakker and Demerouti (2007), I expect that this process can furthermore be characterized by low cynicism and excellent performance.
The JD-R model finally argues that the motivational process can cause higher work
engagement; I therefore also expect that the motivational process, as fuelled by AI, leads to higher work engagement.
Process 2
Based on the JD-R model, I also expect that public professionals perceive AI as a job resource that, due to its automotive nature, reduces job demands by taking over part of employees’ tasks or their entire job. The reduction of tasks by AI is outlined by Eggers et al. (2017) in terms of the split up approach, which states that AI breaks up jobs and automates
25
as many pieces as possible, leaving humans to do what remains. A second scenario centres on the replace approach, where AI takes over an entire job (Eggers et al., 2017).
I therefore expect that AI reduces job demands to the extent that this creates an imbalance between job resources and job demands, where there is an oversupply of job resources and a shortage of job demands to keep employees busy at work. In this scenario, employees experience job strain, which may be exemplified by a displacement of jobs, feelings of uselessness, a loss of control and absence of appropriate ethics as outlined by Fritz and Horvitz (2016). In turn, I expect that job strain causes lower engagement, the negative effect of which may be moderated by job characteristics - as pointed out by Bakker et al. (2004) - and Public Service Motivation.
Regarding job characteristics, I expect that the negative relationship between job strain and work engagement is moderated by Hackman and Oldham’s (1976) notion of job
characteristics. Hackman and Oldham (1976) outlined that the job characteristics of skill variety, task identity, task significance, autonomy and feedback cause work motivation, satisfaction, performance and attendance. Moreover, according to Bakker and Demerouti (2004), job characteristics lead to extra-role performance, where individuals act beyond what is required by their job descriptions. This suggests that job characteristics are particularly motivating for individuals to engage with higher levels of work engagement, perhaps even in the face of job strain. I therefore more specifically expect that Hackman and Oldham’s (1976) job characteristics - i.e. skill variety, task identity, task significance,
autonomy and feedback - mitigate the negative effect of job strain on work engagement. Regarding Public Service Motivation, I expect that the relationship between job strain and a decrease in work engagement is moderated by Perry and Wise’s (1990) reason of PSM. PSM namely refers to an individual's predisposition to engage in meaningful social, community and/or public service and to as such “respond to motives grounded primarily or uniquely in public institutions and organizations” (Perry and Wise, 1990, p. 360). More specifically, these motives are essentially comprised by rational motives, norm-based motives and affective motives (Perry and Wise, 1990). Moreover, as these motives are deemed to be predominantly prevalent in public organizations, it can according to Kjeldsen and Jacobsen (2012), be argued that individuals to whom these motives are relevant are attracted to
26
employment in public organizations. Taking this into account, it can be expected that PSM is particularly motivating for individuals to engage in higher levels of work engagement, perhaps even in the face of job strain. I therefore more specifically expect that Perry and Wise’s (1990) reason of PSM - consisting of rational motives, norm-based motives and affective motives - mitigate the negative effect of job strain on work engagement. In the next sections, the theoretical framework is presented in a model, featuring the variables as well as key causal mechanisms, after which all concepts are respectively defined.
27
Model including JD-R theory, Job Characteristics Theory, PSM Theory and Work Engagement theory
Introduction of AI
Employees focus attention Perceive decrease Job Demands Perceive abundance Job Resources Job Strain Motivation Work Engagement Job Characteristics
28
Expectations
Based on the two processes previously outlined, I formulate the following eight systematic expectations that will be used to analyze results from the multiple-case study:
Expectation 1:
The introduction of AI as a job resource is perceived as an abundance of job resources Expectation 2:
The introduction of AI as a job resource is perceived as a reduction of job demands Expectation 3:
The abundance of job resources causes increased motivation, characterized by low cynicism and excellent performance
Expectation 4:
The reduction of job demands causes job strain Expectation 5:
Increased motivation results in increased work engagement Expectation 6:
Job strain causes a decrease in work engagement Expectation 7:
Job characteristics mitigates the negative effect of job strain on work engagement Expectation 8:
29
Conceptualization into variables
In the above-mentioned framework, model and hypotheses, the Introduction of AI is considered to be independent variable X. Work engagement is considered to be dependent variable Y. The mediator variables M are M(1) Job Demands/Job resources and M(2) Job Strain/ Motivation. The Moderator Variables MO are MO (1) Job Characteristics and MO(2) Public Service Motivation.
All concepts used in the theoretical framework are respectively defined below (note that these include variables as well as regular concepts that are not to be measured):
The introduction of AI is defined, in accordance with the work of Kaplan and Haenlein
(2019), as the practical application in government organizations of “a system’s ability to correctly interpret external data, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation” (p. 15).
Work engagement is defined partly in accordance with the definition used for the Utrecht
Work Engagement Scale (UWES) as a positive, fulfilling, work-related state of mind characterized by vigour, dedication, and absorption (Schaufeli and Bakker, 2004).
Public professionals are defined as those people engaging in specific professions (e.g.
policy-maker, application developer, border guard etc.) serving some aspect of public interest and the general good of society at all levels within public, and more specifically, government organizations. More specifically, public professionals are employed in
government organizations and government agencies such as ministries or (tax) authorities.
Job demands are defined in accordance with Bakker and Demerouti (2007) as those
“physical, psychological, social, or organizational aspects of the job that require sustained physical and/or psychological (cognitive and emotional) effort or skills and are therefore associated with certain physiological and/or psychological costs” (p. 312).
Job resources are also defined in accordance with Bakker and Demerouti (2007), who
defined the term as “those physical, psychological, social, or organizational aspects of the job that are either/or functional in achieving work goals, reduce job demands and the associated physiological and psychological costs or stimulate personal growth, learning, and development” (p. 312).
30
Job Strain is additionally defined in accordance with Bakker and Demerouti (2007) as stress
and arousal occurred at the workplace as a result of imbalance between job resources and job demands. In this thesis, the imbalance is more specifically regarded as job resources reducing job demands, thereby causing job strain as employees feel AI is taking over their designated tasks.
Motivation is defined as either intrinsic or extrinsic motivation, with the former referring to
reasons to act based on the potential for job resources to foster employees’ growth,
learning and development (Bakker and Demerouti, 2007). The latter refers to reasons to act based on job resources being instrumental in achieving work goals (Bakker and Demerouti, 2007).
Job Characteristics, also known as job factors, are defined as aspects specific to one’s tasks
at work, such as skills and knowledge, working conditions and mental and physical demands.
Public Service Motivation is defined in accordance with Perry and Wise (1990) as “an
individual's predisposition to respond to motives grounded primarily or uniquely in public institutions and organizations” (p. 368).
31
Chapter 3: Research Design
Research methods
The type of research design used for this study is an observational research design.
Complete observer qualitative research methods will be used to interpret information based on theory and to test expectations in which interviews and document analysis are important methods.
The use of qualitative methods and especially semi-structured interviews is critical in examining the research question. Although AI was academically founded in 1956 and has been applied more extensively in the past years, it should be pointed out that, in the government administration industry, AI has only just been introduced in a minority of organizations. The vast majority of government organizations has not introduced AI or has just started to think about and develop AI. Besides, the minority of government
organizations that has introduced AI has often decided to not publish any information openly on the topic. Therefore, limited data on the use of AI in government organizations is available. Although some academics have written about the use of AI in government
organizations to certain extents, the amount of information is simply insufficient to quantify and analyze.
One may then raise the question as to why not use structured interviews, a common form of quantitative research methods. The reason why structured interviews are not viable again centres on the fact that knowledge of AI in government is limited. This makes it rather inefficient to ask closed-questions to public professionals. The respondent moreover needs to be able to, at all times, divert from questions as well as be able to bring in new ideas as a result of questions asked.
The use of qualitative research methods, on the other hand, enables me to collect sufficient data and most notably with the use of structured interviews. The use of
semi-structured interviews allows for a framework of themes to be explored and to collect data on these themes on a first-hand basis whilst allowing the interviewee to divert from questions and to bring in new ideas. The prevalence of the latter two options is crucial as
32
public professionals need to be given the opportunity to bring in new and perhaps
unsuspected information. Moreover, it may very well be that the reality of AI in government is totally different than expected due to the relative limited knowledge available of AI and professional work engagement in government organizations.
Informed consent is furthermore ensured by asking respondents whether they agree with disclosing personal information, whether they agree with formal referencing to their first and/or surnames and whether they agree with publication of the interview transcripts. The research design more specifically centres on a multiple-case study design. Data will be extracted first-hand from interviews conducted with public professionals at Dutch
government organizations. These interviews are either direct interviews or indirect interviews, depending on the amount of information the professional can provide on the influence of AI in achieving his/her own tasks or on how other professionals experience the influence of AI in achieving their tasks. This distinction is particularly relevant in scenarios where professionals at higher professional levels (e.g. at the managerial level) know more and can tell more about the influence of AI on street-level professionals than street-level professionals themselves due to differences in education and job levels.
An example that explains the relevance of the distinction in interviews is that Dutch Customs professional Maarten Veltman (Chairman of the Co-ordination Group on
Innovation) knew more about the influence of the introduction of AI on Physical Control Co-workers than the Physical Control Co-Co-workers themselves, due to the co-Co-workers’ relative limited knowledge of AI.
10 professionals from 6 different government organizations will be interviewed. These interviews comprise the following six cases and respondents:
1. Dutch Customs (Ministry of Finance):
Maarten Veltman, Chairman of the Co-ordination Group on Innovation
Tim Zandbergen, Senior Risk Analyst
2. Rijkswaterstaat (Ministry of Infrastructure and Water Management)
33 Stijn van Balen, Head of the Data Laboratory
John Steenbruggen, Innovation manager and advisor to the Chief Data Officer 3. Amsterdam-Amstelland Fire Service
Guido Legemaate, Data Scientist
4. Ministry of the Interior and Kingdom Relations
Haye Hazenberg, Senior Policy Officer Information Co-operation
Chaïm van Toledo, Advisor Business Management and Researcher Automatic Conversation Systems (supporting role)
5. Enschede City Council
Paul Geurts, Strategic Advisor Information Policy 6. Nijmegen City Council
Petra Bout, Manager Cluster Service Provision (Customer Services)
The questions in the interviews are set up by me in the context of the theoretical
framework. They are set up in the context of retrospective research methods, as I aim to uncover the influence of AI that has already been introduced. Furthermore, in order to comply with informed consent, I inform respondents at the start of the interview that the results of the interview will be used in writing the Master’s thesis. In doing so, I ask for their consent to record the interview and to refer to their names and job titles when using information from the interviews.
The methods of analysis centre on qualitative comparative analysis (QCA), as I will be conducting data analysis to determine which logical expectations are supported by the data set, i.e. the collective results from the interviews. The analysis starts with listing
combinations of variables and values from the results in a basic truth table. In this regard, the QCA is using formal logic instead of statistics to test relationships between variables. The variables, including values, are assessed on a scale and put in the table, lining them up with the outcomes. This is then followed by applying logical inference in order to determine which variables and expectations are empirically supported by the data. As such, in using
34
QCA, explanatory patterns for both process 1 and process 2 in the theoretical framework can be unraveled. Furthermore, online sources such as reports (e.g. Deloitte Government Insights) and academic journals will furthermore serve as material for document analysis and as sources to collect information on the theoretical framework.
Threats to inference
In listing combinations of variables and values on a case-by-case basis in a single truth table as part of qualitative comparative analysis, I am able to clearly draw causal inferences about the relationship between the introduction of AI and work engagement. In this regard, I am also able to make more specific inferences and determine which expectations are
empirically supported by the data. This way, I explain the impact of key variables on work engagement by making systematic comparisons across contexts and can, as such, ensure internal validity.
In terms of external validity, the QCA approach with its strong case-orientation traditionally entails the analysis of purposively selected cases, allowing for modest generalization to a broader universe of cases relevant for the research question (Thomann and Maggetti, 2017). External validity can be high when inferences about the samples “can be generalized as much as possible beyond its boundaries” (Thomann and Maggetti, 2017, p. 12). However, case sensitivity surrounding qualitative case analyses can pose challenges to this external validity and in specific regard to sampling bias (Thomann and Maggetti, 2017). Sampling bias has however been prevented as, in the process of inquiring for interviews and subsequent case selection, requests have been sent to random Dutch ministries and government organizations, all Dutch provinces, the 30 largest city councils in The Netherlands and all Dutch water boards and safety regions (veiligheidsregio’s). Out of 15 positive respondents in the period of 1 March – 31 May, the first 11 respondents were simply selected in order to conduct interviews as soon as possible. In this regard, cases have been selected non-purposively, resulting in potentially more than modest generalization of the findings. Although one may point out that the city councils have initially been selected in terms of size, it should be underlined that only larger city councils in the Netherlands introduce AI and that even the larger ones are still in a very early stage of introducing AI. Therefore, no immediate bias is prevalent in terms of sampling and, therefore, external validity can still be
35
expected to be high. Moreover, external validity can furthermore be expected to be high as the values of the variables have been selected from heavily-cited theories prevalent in the theoretical framework and from key reports authored and published by the world’s largest consultancy firms that practice public policy consulting (o.a. Deloitte and Capgemini). These sources have each studied different cases and came up with generalizable results in their respective research contexts. In applying these values in an empirical research context where expectations on random government organizations and AI domains are tested at the local, regional and national levels, high external validity may be expected.
The methods also include measures to take into account potential confounding variables. This will include measures such as identifying technologies other than AI and any other subsequent job resources that could be responsible for influencing work engagement. This will be key to take into account as part of the interviews/surveys to rule out any threats to the validity of the study. Among others, this will take shape by asking respondents a number of isolated questions on particular AI technologies. More informally, questions are also asked to find out if there are other technologies that may prove correlation with particular change factors in job resources specified to physical workplace resources. Validity and reliability of the measurement of these and other concepts are furthermore specifically outlined in the operationalization of concepts’ section.
Level of analysis, relevant populations and case selection
Government organizations are the primary level of analysis. The relevant populations more specifically are Dutch government organizations at the local, regional and national levels in the European territory of the Kingdom of the Netherlands. The unit of observation,
however, is individual public professionals engaging in professional work activities. The organizations comprising the multi-case study have essentially been non-purposively selected, although city councils were selected based on their size which is, in this period of time in The Netherlands, an antecedent for the introduction of AI, as outlined in the previous section.
36
Furthermore, although there were more than 15 respondents willing to engage in interviews, it should be noticed that this surplus did not adequately match the central definition of AI used in this thesis. Therefore, 15 ‘positive’ respondents remained that matched the definition. Due to constraints in time, the first 11 respondents were selected for interviews. As previously outlined, this led to a total of 11 interviews with public professionals from six different government organizations comprising the following six cases:
Dutch Customs (Ministry of Finance)
Rijkswaterstaat (Ministry of Infrastructure and Water Management)
Amsterdam-Amstelland Fire Service (Ministry of Justice and Security)
Ministry of the Interior and Kingdom Relations
Nijmegen City Council
Enschede City Council
This ensures that the multiple-case study design can extract sufficient data from
government organizations, enabling the identification of job demands, job resources, job characteristics and PSM and the final measurement of professional work engagement. Above all, this ensures that all expectations connected to these variables can be tested among an adequate array of respondents.
Operationalization of concepts
Preliminary note: not all concepts that are operationalized are variables that will be measured as part of the theoretical framework. Variables that are measured as part of the theoretical framework will feature additional arguments in terms of operationalization in order to ensure that measurements are well-founded, correspond properly to the real world and satisfy consistency across cases. This is more specifically outlined below in terms of validity and reliability for each single concept. Furthermore, concepts are first and foremost measured in face-to-face interviews with public professionals.