Artificial intelligence,bureaucratic form,and discretion in public service

1. Introduction

The diffusion of information and communication technologies (ICTs) over the past several decades has profoundly changed administrative behavior in public organizations (Fountain, 2001). As their capacity for data generation, storage, and processing continued to grow exponentially over the late 20 th and early 21 st centuries, ICTs became increasingly central to organizational decision-making processes (Gil-Garcia et al., 2018). This increased emphasis on technologically-mediated decision-making directly affects the quantity and quality of discretion – the latitude afforded individuals with delegated responsibilities to use their judgement when making a decision – afforded public administrators (Bullock, 2019).

From a top-down, managerial perspective, one of the principle benefits of ICT-mediated discretion is the standardization of bureaucratic behavior. Standardization reduces the risk of moral hazard endemic to the delegation of responsibilities and the corresponding degree of discretion afforded both street-level and mid-level agents. This standardization by ICT may also improve decision making and outcomes (Greer & Bullock, 2018). A bottom-up perspective, by contrast, views these changes as the erosion of professionalization, particularly among street-level bureaucrats who interface directly with the public and have considerable discretion to determine how to provide services and enforce policies (Lipsky, 2010). As street-level bureaucrats’ work processes transition from person-to-person interactions to interfacing with ICTs, service-providing public organizations transition from street- to screen-level, and ultimately system-level, bureaucracies (Bovens & Zouridis, 2002). More recent theoretical and empirical work adds further nuance to this theory of change, calling attention to how task, technology, and organizational characteristics contextualize the effect of “digital discretion” in practice (Hannah-Moffat, 2019; Peeters & Schuilenburg, 2018; van Eijk, 2020).

With respect to technological context, recent advances in Artificial Intelligence (AI) have led to its increasing use in both private and public organizations (Rahwan et al., 2019). In particular, these uses involve a particular form of AI: machine learning-based applications that adjust their decision criteria dynamically using stochastic optimization processes. The exponential growth in AI’s use, capabilities, and risks has not gone unnoticed (Bostrom, 2014; Dietvorst et al., 2018; Zou & Schiebinger, 2018). Much of this work has focused on AI’s potential impact on the economy, the labor market, and national defense (Frank et al., 2019; Frey & Osborne, 2017; Korinek & Stiglitz, 2017; McClure, 2018). Another vein of related work focuses on the use of machine learning-based decision support systems in criminal justice for bail setting and sentencing (Binns, 2019; Hannah-Moffat, 2013, 2019; Hannah-Moffat et al., 2009). Less attention, however, has been paid to AI’s implications for public organizations and the governance of the public sector more broadly. Moreover, much of the related work focusing on public organizations is framed around the concepts of “big data” (Brayne, 2017; Ferguson, 2017) or “smart/algorithmic governance” (Eubanks, 2018; Gil-Garcia et al., 2014). Missing from this discussion is explicit attention to the implications for the public sector that arise both from allocating discretion away from humans to AI, and the corresponding changes to the form of bureaucratic organizations, often towards systems level bureaucracies.

This is now beginning to change. Recent scholarship has begun to develop frameworks for describing, understanding, and evaluating the use of AI to perform tasks in public sector and organizational contexts (Bullock, 2019; Drexler, 2019; Young et al., 2019). These scholars argue that, unlike other forms of digital discretion, AI is best understood as an agent embedded in an organizational context that executes tasks using stochastic, or non-deterministic, approaches. These approaches require learning, sensing, and probabilistic reasoning, and allow AI to perform tasks that are ill-suited for deterministic expert systems algorithms, which are built using logic-based rules. These complex and contingent tasks, and the institutional framework necessary to enable (and constrain) their execution, were previously the exclusive purview of human agents. However, for a growing set of increasingly complex tasks involving decision-making, this is no longer the case – the man in the Mechanical Turk has been made redundant. Exactly how the use of AI affects the nature of discretion in public organizations, and how this use is conditioned in turn by organizational context, however, is largely unknown.

We address this lacuna by identifying uses of AI in different organizational contexts in the public sector and analyzing them through the theoretical lenses of systems bureaucracies, digital discretion, and artificial discretion. Two research questions motivate this work: How is the use of AI both changing and changed by the bureaucratic form of public organizations, and what effect does this have on the use of discretion within these organizations? In answering these questions, we examine both the use of AI and the form of bureaucracy for predictive policing, facial recognition, and criminal pattern recognition, and how the United States Centers for Medicare & Medicaid Services (CMS) implements AI in its antifraud and improper payment reduction efforts.

These two domains are chosen as initial, illustrative explorations into the use of artificial intelligence and form of bureaucracy for several reasons. Perhaps most importantly, both involve the current, active use of AI in core organizational functions, as well as past and potential future expansion of its use over time. Both cases are also highly salient in terms of the cost of service provision and the impact on individual citizens and society. Finally, the two domains provide reasonable variation across where the organizations are located along a continuum from street-level form to system-level form. These factors, taken together, make our cases important early explorations into the advance of both artificial discretion and the system-level bureaucracies in the public sector.

We proceed by first summarizing our theoretical framework, which draws upon Bovens & Zouridis’ (2002) theory of system-level bureaucracy, its extension in Busch and Henriksen’s (2018) digital discretion, and its relationship to the AI-specific focus of Young et al.’s (2019) theory of artificial discretion. We then turn to our cases, first providing descriptive analysis of how AI is currently being used in situ, and then applying our framework to test its explicative power. We conclude with a distillation of our findings coupled with suggestions for future research.

2. Theoretical framework

Scholarship on how ICT implementation affects discretion in public organizations is multifaceted, with theoretical and empirical work alternatively focusing on the national level (Fountain, 2001), local level (Busch & Eikebrokk, 2019; Tummers & Bekkers, 2014), and different policy domains including welfare (Hetling et al., 2012; Houston, 2015; Pors, 2015), criminal justice (Hannah-Moffat, 2019; Marks et al., 2015; van Eijk, 2017; Završnik, 2019), and others. A unifying feature across empirical contexts is the recognition that as public organizations increase their use of ICT, and as the tools themselves increase in their scope and capacities, the result is a reshaping of administrative work itself, leading to new forms of bureaucracy (Bovens & Zouridis, 2002; Danaher, 2016; Dunleavy et al., 2006; Peeters & Schuilenburg, 2018; Peeters & Widlak, 2018).

One early and highly influential work is Bovens and Zouridis’ (2002) theory of ICT-driven organizational change from street- to system-level bureaucracies. The theory explains that as ICT become increasingly central to the execution of routine tasks, the nature of this work changes from person-to-person interactions to person-to-computer interactions, a stage termed “screen-level bureaucracy.” The theory calls explicit attention to the impact of this shift to screen-level bureaucracies for street-level bureaucrats – those whose responsibilities required direct interaction with the public – and the relatively high degree of discretion they are afforded absent these technologies. Bovens and Zouridis go on to argue that ICT may eventually become not only the primary but also the sole structure for communication, flow of information, and organization of work tasks; at this stage of development the organization is said to be a system-level bureaucracy with little to no place for street-level bureaucrats as traditionally defined (Bovens & Zouridis, 2002).

Importantly, the transition from traditional to screen- and ultimately system-level bureaucracy is not conceived of as an irresistible force. Rather, the propensity for this change is explicitly contingent on the nature of the work tasks required, and by extension the initial degree of discretion afforded to its agents and other organizational characteristics. The examples provided in the original theory are education and policing; the authors acknowledge that these domains of public service provision are extremely complex and contingent, and require empowering street-level bureaucrats – teachers and officers, respectively – with a high degree of discretion (Bovens & Zouridis, 2002). This conditioning of expectations was later enriched in Busch and Henriksen’s (2018) review and extension of the literature on ICT and discretion that followed the publication of the theory of system-level bureaucracies. Their analysis calls attention to both the technological characteristic of the ICT and the importance of the unit of analysis – micro, meso, or macro – within the organization where the ICT-mediated task takes place. It also reveals a public sector still staffed with street-level bureaucrats, but with virtually all of them embedded in varying forms of screen-level bureaucracy, with some tasks mediated by ICT and some automated entirely.

Other studies of the relationship between ICT and discretion also call attention to the importance of the technology’s characteristics. For example, technology does not just curtail discretion; it can enable it as well by affording new capabilities – whether intentional or not (Buffat, 2015; Hupe & Buffat, 2014; Thunman et al., 2020). In describing and explaining what they term “digital discretion,” Busch and Henriksen note that technological advances in ICT – most notably AI – pose unique questions for the structure of public organizations and the use of discretion within them vis-a-vis automation.

The question of AI’s impact is addressed directly in Young et al.’s (2019) framework for understanding and evaluating the use of AI in public organizations, which they term “artificial discretion.” Central to their argument is the distinction between machine learning-based AI and traditional, expert systems-based algorithmic automation. Unlike expert systems, most modern AI applications – and all of the more powerful, complex forms, e.g. neural networks – are stochastic. As a consequence, it is not possible to reverse engineer a decision reached by AI with perfect certainty despite having complete, perfect knowledge of all available input variables/features. This is reflected in how computer scientists evaluate AI decision performance: AI is evaluated probabilistically (e.g., a 93% success rate at making the correct classification decision), rather than in terms of absolute fidelity. Moreover, the final architecture of any AI system after it has been trained is fundamentally unique – even when the trainer uses identical initial architectures and training data (Russell & Norvig, 2009).

This machine learning characteristic of AI is what makes it capable of performing tasks that were previously considered solely the domain of human agents. Furthermore, AI’s stochastic behavior introduces the same information asymmetries between manager-or-organization-as-principal and AI-as-agent endogenous to administrative discretion as traditionally understood. For example, consider the use of AI for facial recognition. The outcome of a discrete task in this context is to determine whether an image of a face can be reliably paired to a separate image with associated metadata about the individual (e.g., name, age, known residence, etc.). The AI-as-agent ultimately produces a decision: it reports either a match/a set of possible matches, or that it failed to find a credible match. On a sub-task level, this process involves comparing n pre-known images of faces against the new image; at a level below this the AI agent is performing k subroutines, such as evaluating the geometries, shade gradients, and other constituent elements of the images against each other. Thus, from an organizational perspective, AI agents possess the same afforded discretion for this task as any human agent: they can be trained, and held to performance standards, but the precise outcome of any discrete task is not knowable a priori.

Young et al. (2019) also integrate the contextual effects of task type and unit of analysis previously identified in both the theory of system-level bureaucracy and digital discretion. They then derive propositions for how AI use is likely to vary based on these contexts. For tasks with low levels of required discretion, automation by AI may be possible removing much of the need for human discretion. For tasks with medium levels of required discretion, AI is mostly likely to be used as a tool of prediction, finding patterns in multidimensional data, and generating new insights into variables that affect task execution. Finally, neither automation nor prediction may be an appropriate use of AI for tasks that require a high level of discretion. Instead, AI may be used to generate better data from unstructured inputs such as images, sensors, and text that can aid human decision makers in executing a task more effectively (Young et al. 2019).

Along with the degree of discretion, the unit of analysis at which a task is embedded within a bureaucracy is also helpful for identifying what types of tasks are likely candidates for AI. The organizational context refers in part to the scope of influence or effect a given task has – localized (micro), organizational (meso), or institutional (macro). The micro level of analysis consists of tasks that are implemented at the individual level, in general at lower levels in the organization such as street-level bureaucrats. Meso-level tasks “shape and affect the organizational environment in which individual agents are embedded” (Young et al. 2019; 5). Macro-level tasks are institutional or enterprise level tasks that include the contextual factor of formulation of rules and general policies for the organization (Busch & Henriksen, 2018).

The organizational context includes not only the level (micro, meso, and macro) within the organization at which the task is embedded, but also the location of the organization along the continuum of bureaucracy form from street to system-level, which also shapes the organizational context for decision making and task execution. The location of the organization along this continuum is a function of the type of tasks that dominate that organization and its policy domain. If the tasks, in general, require less discretion then it is more likely that ICT tools have begun playing a decisive role for the decision making of that organization (it is easier to adapt these simpler tasks to automation by machine), while tasks that require more discretion have been more likely to rely on human discretion and keep the street-level bureaucracy form.

To summarize, the theory of system-level bureaucracy is further extended and developed by both digital and artificial discretion. The theoretical expectations are summarized in Table 1. Viewed synthetically, the extension of these theories of ICT-driven organizational change in the public sector to the specific case of AI lead to empirically testable expectations for how AI’s implementation will both be shaped by organizational and task contexts and, in turn, affect the way discretion is allocated and used within these organizations and, by extension, the organization itself. Accordingly, the characteristics of the tasks – in this case, the required level of discretion – that dominate a bureaucracy influence both the location of the bureaucracy on the street to system-level continuum and the allocation of uses of artificial discretion. In street- and screen-level bureaucracies we expect to see AI used collaborative with human agents as a decision support tool and for gathering new information. In system-level bureaucracies we expect to see AI used more predominantly to automate tasks, reducing or eliminating human agents in decision processes.

This theoretical framework integrates insights from traditional ICT literature, advances in our understanding of digital discretion, and recent work highlighting the unique characteristics and impacts of AI and its use by public organizations. Building from previous literature we have argued that AI presents unique challenges to public administration. For policy domains that are characterized predominantly by high discretion tasks we expect more traditional street-level and screen-level bureaucracies to dominate and for a less reliance on artificial discretion and when artificial discretion is employed the uses are to gather more information and to provide decision support. However, for policy domains that are characterized by low discretion tasks we expect system-level bureaucracies to emerge and for a stronger reliance of artificial discretion to develop and for artificial discretion to take the form of automation.

To test the explanatory power of this framework, we selected two important policy domains: policing and public health insurance. These domains were chosen to provide cases with strongly differing organizational contexts and associated requirements for discretion (Seawright & Gerring, 2008). Both policy domains have long histories to observe any shifts along the continuum from street-level bureaucracy to system-level bureaucracy, and the impacts of artificial intelligence. We collected government reports and reviewed related prior research for our analysis.

In our two illustrative cases, we have described some of the uses of AI that can be found in both the policing and public health insurance domains. Through these cases we demonstrate how the use of AI affects the exercise of discretion and how both organizational and task characteristics influence the uses of AI in turn. We further show that the use of AI itself may pressure bureaucracies to transition from street- and screen-level to system-level bureaucracies. These effects carry significant potential consequences for organizational transparency and accountability: the complexity and automaticity of system-level bureaucracies makes it difficult to provide satisfactory ex post explanations of their decisions. On the other hand, these effects may yield significant gains in efficiency and effectiveness for many of the outcomes these organizations are held accountable for providing to the public. Table 2 summarizes our findings, which we discuss in more detail above.

The use of AI affects the exercise of discretion. CMS is a system-level bureaucracy that employs AI to expand both the scope of its work efforts and its efficiency, or the rate at which it can perform its required tasks. While humans remain integral to the overall structure, many of CMS tasks no longer require human discretion or with the use of AI and digital discretion have expanded into tasks that would not be feasible for humans to complete. This use of AI, then, constitutes a process improvement that does not require the organization to adapt its system-level behavior by having to create new positions for human agents. Here it may seem that human discretion is often completely absent, as system level bureaucracy theory would suggest. Indeed, one use of automation by CMS is the identification of improperly paid public health insurance claims and automated prepayment edits to help prevent fraud. AI can also be used to predict claimants’ risk levels or where patterns of fraud may be occurring by utilizing predictive analytics. Finally, with the widespread digitization of personal and event data, AI is used to gather new types of data on claimants and claims. AI can be used in this way to help gather new types of data to be considered as part of the claims process.

Street-level bureaucrats in law enforcement settings, however, are still entrusted with a great degree of discretion, both explicit and tacit, in performing their duties because of both the high degree of uncertainty associated with many of their tasks and the political importance of their outcomes. This discretion can be problematic. For example, information asymmetries and resulting moral hazard between street-level officers and their managers, and social and political harm when officers make mistakes or abuse their discretion to needlessly punish individuals and socioeconomic groups (Pierson et al., 2020; Selbst, 2017). Additionally, COMPSTAT’s effectiveness is as much about its ability to regulate and make for predictable, measurable behavior on the part of street-level officers as it is about reducing crime rates (Benbouzid, 2019). Studies on the applied use of predictive policing have shown that it is recognized both by management and by street-level employees as a way for management to exert more control and have more certainty about how the street-level officers are spending their time (Brayne, 2017).

All three of AI’s identified uses in law enforcement encroach on tasks that have historically been the purview of the officer on the street or an entire professional class of law enforcement employees called crime analysts. Historically, crime forecasting would either involve an ad-hoc, experience-based decision by veteran officers, or crime analysts using tools like Geographic Information Systems (GIS) to produce hotspot mapping and other forecast outputs. Similarly, finding criminal patterns across multiple cases was either the job of a detective to notice patterns based on experience, or a professional crime analyst. Facial recognition in policing was classically part of the community policing model of foot patrolmen getting to know the residents in a small area and knowing the individuals on the street by virtue of familiarity, or stopping unknown individuals and asking for identification.

The organizational context shapes the way AI is used. One example for policing is the institutional history around investments in information technology in American law enforcement. This can be traced back to the 1968 Omnibus Crime Control and Safe Streets Act, which created a federal resource stream for local law enforcement to fund investments in information technology. The causal theory behind the allocation of federal funds for local police to invest in ICT was that too little and too disjointed information was an impediment to solving and preventing crime. As investments in ICT continued and technological capacities grew over time, most medium- to large-sized departments found themselves with a surfeit of data stored in multiple, disjointed systems in a myriad of formats (Andrejevic, 2017). The COMPSTAT program developed by William Bratton in New York in the late 1990s/early 2000s and brought with him to Los Angeles systematically brought these – and yet more – data together in a model of ex post performance measurement and management that was rapidly emulated throughout the US and in the United Kingdom (Sherman, 2013). The COMPSTAT system is also the direct technological progenitor to modern predictive policing: as mentioned earlier both PredPol and HunchLab platforms were spun-off from previous work on COMPSTAT-based software. Given the amount of data that law enforcement agencies have, and the political pressure to put those data to use, the diffusion of AI in this context seems almost inevitable.

The current use of predictive policing represents a high-discretion task. As an organizing methodology, predictive policing has broad and pervasive impacts on organizational structure, behavior, operations, and performance. At the same time, the discretion required to forecast likely criminal activity and direct armed police officers to intervene is necessarily broad. Criminal pattern recognition is a medium-discretion, micro-level task. It replicates work done previously by individual detectives and/or crime analysts, but also requires sophisticated pattern recognition and analysis that do not currently lend themselves to complete automation. Finally, facial recognition involves a relatively low level of discretion, but the organizational context is difficult to map due to the scaling effects of AI. Whereas a patrol officer recognizing a suspect vs. an innocent passerby was previously a micro-level task, that is difficult to equate to the city-wide dragnet capabilities of AI-enabled policing.

For public health insurance one important organizational context is the abundance of numerical data. These numerical data have a history of being organized for use by quantitative analysis, and the data have already been regularly analyzed by digital computing systems. This history of data collection, organization, and analysis suggest that public health insurance tasks may be even more particularly suitable to the use of artificial discretion and the use of automation of some tasks. Tasks that involve processing health insurance claims lend themselves to task characteristics that contain clear rules, clear processes, and can be numerically described. The boundedness of many public health insurance claims tasks is also narrower and often more clearly defined than, for example, tasks with the delivery of broader health services where tasks often require dynamic interaction with human clients. This suggests that there are likely significant sets of tasks within the provision of public health insurance that have lower levels of required discretion, which also may lead to more automation and further development of the system-level bureaucracy.

We find that improper payment and fraud detection, tasks that require low levels of discretion and are embedded at the micro level of organization, are beginning to be automated. We also find tasks such as predicting hubs of fraudulent or criminal activity are being augmented and enhanced with predictive analytics tools. These tasks require more discretion and thus predictive analytics working together with human analysts, rather than complete automation of the tasks. Finally, we see that AI can also be used to gather more data for insurance claims and process vast amounts of video for police departments to help provide more detailed data to decision makers across these policy domains.

The use of AI may also affect organizational structure. We observe the role AI itself may play in moving organizations along the continuum towards system-level bureaucracies. Thus, it seems that by developing the organizational infrastructure for the use of AI furthers the organizational along its path towards a more decisive role of ICTs in the decision-making function. Police departments, which are much closer to traditional street-level bureaucracies, but also incorporate significant elements of screen-level bureaucracies, use AI for numerous types of tasks. Policing tasks generally require more discretion, but we still find that AI is used for data collection, decision-support, and automation of some tasks. This spread of AI throughout policing appears to be shifting police departments to more widespread and decisive uses of ICT and thus towards the form of system-level bureaucracy. For CMS, within this task set of reducing fraud and improper payments in the delivery of public health insurance, we find that CMS is using AI to help gather and organize insurance claims data, analyze that data for patterns, and automate large parts of the search function for identifying problematic claims. These are the types of tasks that are prevalent in a system-level bureaucracy, and where our theory expects to find AI employed in those organizations.

Jointly, this suggests that the use of AI encourages a transition into a system-level bureaucracy and that when system-level bureaucracies use AI it is often to curb and restrict human discretion, in a manner consistent with the original theory of system bureaucracies. However, our findings suggest that these pressures are neither linear nor monotonic. Instead, they are moderated by the preexisting bureaucratic form.

AI as implemented in both policing and insurance administration reduces human discretion in some cases. But this effect is much more substantive at the margin in insurance administration because its system-level bureaucratic form eliminates potential discretion-increasing spillover effects. On the other hand, AI’s use for various forms of pattern recognition in policing has a more balanced effect on discretion: opportunities for human agents to exercise discretion are both destroyed and created. The reciprocal effect of AI on bureaucratic form, however, is much stronger in the case of policing than it is for insurance administration. While AI-augmented policing still retains features of both street- and screen-level bureaucracy, the logic of scalability and efficiency provided by AI-focused systems integration signals an acceleration towards system-level bureaucracy in organizations previously identified as being highly resistant to such changes. Meanwhile, AI’s effect on bureaucratic form is negligible in the case of insurance administration. We attribute this null effect to the inherent congruence between system-level bureaucracies and the automaticity that AI enables and enhances; they are naturally complementary systems.

5. Conclusion

In the final edition of Administrative Behavior, Simon (1997) argued that while ICTs were diffusing throughout organizations in the form of personal computers and the internet, they had not yet drastically altered the bureaucratic form or decision-making processes – including discretion – of administrative organizations. By 2002, Bovens & Zouridis argued that, in fact, for many bureaucracies, ICTs had drastically altered the bureaucratic form, giving rise to the system-level bureaucracy and reshaping the use of discretion in the process. These system-level bureaucracies take a different shape than their street and screen-level forms. Discretion is much more centralized, tasks more routinized, and few if any street-level bureaucrats are present. Since the dawn of the 21 st Century, ICTs have continued to spread throughout bureaucracies, giving rise to more system-level bureaucracies, and the ICTs themselves have continued to increase in their capabilities.

Modern AI has further increased the capabilities of ICTs to conduct the work of public organizations. Furthermore, AI’s unique machine learning characteristics allow it to operate in a non-deterministic fashion, making it much more akin to the decision-making processes of human discretion. In this article we have examined how the use of AI alters the exercise of discretion within an organization and its bureaucratic form, as well as how the form also moderates the processes for which AI is used. Drawing from extant theoretical literature on ICTs, the characteristics of AI, and its relationship to discretion and organizational context in the public sector, we argue that when AI is used within system-level bureaucracies like the Centers for Medicare and Medicaid Services it is often used for automation of tasks and the further reduction of human discretion. This stems in part from the fact that system-level bureaucracies consist of tasks that require lower levels of discretion and have already reshaped the decision-making process to make it more amendable for automation. Street- and screen-level bureaucracies such as those found in policing, on the other hand, consist of more tasks that require higher levels of discretion and thus they have been more resistant to a transition to a system-level bureaucracy. However, we find in our policing case that AI is adopted for many uses across many tasks, and thus we suggest that the use of AI may also represent a sea change in the push to transform these organizations towards a system-level bureaucratic form.

At the intersection of AI, bureaucratic form, and discretion, we find that as AI is deployed within public organizations, discretion is altered. In system-level bureaucracies discretion continues to shift away from street-level bureaucrats to software developers and machines. In bureaucracies that have not transitioned to a system-level bureaucracy, AI does reduce some forms of discretion by automating tasks. But at the same time, it also enables new uses of discretion by providing new capabilities in data gathering, decision support, and predictive analytics. Additionally, the use of AI itself requires decision making processes and bureaucratic forms that better integrate ICT infrastructure into the decision-making process. This infrastructure requirement of AI use, along with the variety of tasks for which AI may then be applied, suggest that AI implementation accelerates organizations’ transition towards a system-level bureaucracy, even when those organizations traditionally resist such efforts.

Additional theoretical and empirical work is required to further improve our understanding of how bureaucracies are evolving in response to implementing AI, and how bureaucratic forms shape the nature of AI implementations in turn. While AI can improve the execution of increasingly complex tasks, what is less known is how to carefully integrate these tools into the decision-making process and structure of public organizations. We are beginning to explore the actual uses of AI within public organizations and the consequences of that integration. However, more detailed cases studies, surveys, and interviews are needed to better understand how public servants and public organizations are implementing AI and the resulting consequences. Additionally, more theoretical work needs to be done to understand the general consequences to bureaucratic form as AI becomes more intimately integrated. This work must also revisit classic questions of public governance and administration, such as the relationship between principals and agents, agents’ motivations, and the structure of policies as human and artificial agents work together to govern.