Algorithmic transparency and bureaucratic discretion: The case of SALER early warning system

2.1 Artificial intelligence and algorithmic governance

Implications of algorithmic governance in the public sector come up hand in hand with expected situations derived from the utilization of autonomous systems to make decisions. The implications of algorithms and AI systems is growing in different areas of the public sector. Wirtz et al. (2018) distinguish at least ten areas of AI and algorithm-mediated governance with potential impact in the future of the public sector (i.e. AI Process Automation Systems, Virtual Agents, Predictive Analytics and Data Visualization, Recommendation Systems, or Speech Analytics). All of them are fostering disruptive consequences in different industries and sectors and it is supposed to occur in the same way in the public sector (Agarwal, 2018).

At the same time, the notions of AI, algorithm and algorithmic decision-making are not completely shared in the literature. Here, following the High-Level Expert Group on AI of the European Commission we broadly define AI as “systems that display intelligent behaviour by analysing their environment and taking actions – which some degree of autonomy – to achieve specific goals” (European Commission, 2019: 1), and even more precisely “AI systems are software (and possibly also hardware) systems designed by humans that, given a complex goal, act in the physical or digital dimension by perceiving their environment through data acquisition, interpreting the collected structured or unstructured data, reasoning on the knowledge, or processing the information, derived from this data and deciding the best action(s) to take to achieve the given goal. AI systems can either use symbolic rules or learn a numeric model, and they can also adapt their behaviour by analysing how the environment is affected by their previous actions” (European Commission, 2019: 6). Accordingly, the idea of algorithm is very closely related to AI as it consists of the step-by-step processes and/or rules processing inputs into outputs in AI-based systems (Jansen & Kuk, 2016). Therefore, one of the key issues of AI and algorithms since the emergence of computing is the reasoning/information processing dimension.

Also, we present the idea of decision-making based on AI-systems and algorithms, as it is one of the critical issues regarding organizational (public administration) studies. Here, we argue that the term “decision” should be considered broadly regarding AI systems, including “any act of selecting the action to take, and does not necessarily mean that AI systems are completely autonomous. A decision can also be a selection of a recommendation to be provided to a human being, who will be the final decision maker” (European Commission, 2019: 3). At the same time, decision-making based on AI-systems implies that once the action has been decided, the AI system might be ready to perform it or leave that actuation to humans (Harris & Davenport, 2005). Therefore, decision-making based on AI-systems may, or may not, convey final actuation of humans.

Algorithmic governance encompasses deep transformation in the political and public administration realm. From the political side, some claim that AI and algorithms will reshape the global order (Wright, 2018) or change interactions among countries (Lee, 2018). From the perspective of public organizations, AI and algorithms may transform processes (Brynjolfsson & McAfee, 2017), be used in the analysis of citizens behavior to understand their needs and deliver public services better suited to those needs (Vogl et al., 2020), verification of eligibility for public services and detection of fraud, or better operational performance of service providers (Margetts & Dorobantu, 2019). At the same time, different authors have recently identified several cases in the public sector implementing AI and algorithm-mediated governance coming from different public service sectors (Desouza, 2018; Sun & Madaglia, 2018) (i.e. health, education, social services, migrations, customer care, tax, social security, police, emergencies, transportation, water management, waste recycling, air control, among others). Besides, it is likely that these service areas may be improved by AI and algorithms regarding the different stages of the policy-cycle process (Pencheva et al., 2020; Valle-Cruz et al., 2020).

At the organizational and service level, governance of algorithms and decisions based on them in public agencies raise different concerns, including data quality, social and economic biases, and opacity. First, the quality of data is a key factor to avoid noise, heterogeneity, etc. Using algorithms that are nurtured by data from different organizational areas requires that quality is controlled. Second, different potential biases derive from the design and training of data, in particular, regarding to racial, gender or economic features (Cath, 2018). Hence, public organizations could produce negative outcomes using social and economic profiling based on previous biased administrative systems and data records. Also, opacity is a potential hidden burden in public agencies using AI as a result of the black-box effect and potential lack of accountability over decision-making. Here, the limits of transparency during the design and implementation of algorithms might erode public governance and reduce control and even trust in public institutions. Also, there is another key point having an appropriate understanding of regulation of algorithms in their context of application.

2.2 Regulating algorithms in the public sector

The current doctrinal discussion on the legal aspects related to the use of the AI in the public sector has mainly focused on the analysis of the requirements, conditions and guarantees arising from the use of this technology as regards, on the one hand, the guarantee of the citizens’ rights and, on the other hand, the adequate protection of the general interest. From the latter perspective, the fact that there is a close relationship between the use of AI in the public sector and the principles of good administration in dealing with cases of corruption has been highlighted (Ponce, 2018), especially in the area of public procurement (Aarvick, 2019). The link with such principles is an essential requirement when facing the adaptation of the legal framework to this emerging technology, particularly if we take into account the singularity of the legal risks and challenges that emerge from the technological innovation approach (Valero, 2019).

Indeed, one of the areas of major concern has been the need to adapt the guarantees provided by the general regulation on transparency and access to public sector information to the algorithms (Cerrillo, 2018), in the light of the fact that the regulatory framework is rather outdated in terms of the IA’s requirements. Therefore, the proposals based on the preventive application of general principles in the design of the applications are particularly thought-provoking (Loi et al., 2019). One of the main topics of discussion concerns the transparency of the algorithms and, in particular, when private entities contracted by public administrations participate in their design. In these cases, public procurement legislation may involve restrictions on transparency that should be proactively addressed to ensure adequate control (Capdeferro, 2020: 9–10).

On the other hand, the requirements of the principle of good government are projected onto the discretion of administrative decisions, especially with regard to their motivation when algorithms are being used. In the case of the Spanish legislation, there is no specific legal obligation in this regard despite the fact that the regulatory framework is certainly recent (2015). However, the need to reserve the adoption of discretionary decisions to human decision has been discussed (Ponce, 2018), particularly given the need to weigh up certain aspects that are difficult to assess adequately by a computer system. In this context, the discussion about whether or not the decision-makers have the discretion not to use the result of computer applications based on the use of AI or big data seems to be particularly relevant, insofar as other values beyond mere technology have to be considered (van der Voor et al., 2019; Ponce, 2019).

2.3 Algorithmic transparency and the black box problem

Transparency is one of the challenges for public agencies operating algorithms to deliver public services. Algorithms are a set of rules or procedures to solve a problem or to find a solution for an administrative procedure. Then, transparency of algorithms (or algorithmic transparency) is a key factor for the understanding of their results and potential problems derived from the decisions made using autonomous systems. To put it differently, algorithmic transparency, defined here as accessibility and explainability of public decisions made by algorithms (Grimmelikhuijsen, 2019), needs to be addressed in order to appraise the impact of algorithms in public agencies and divert black box concerns in public administrations during the time of open government.

Mostly, the black box problem of algorithms has much in common with two dimensions of transparency: accessibility and explainability. Here, Grimmelikhuijsen (2019) or Grimmelikhuijsen and Meijer (2014) argue that, in the first case (accessibility), transparency is accomplished by the availability of information in different dimensions of the policy process or managerial functions of public agencies. Making the information available and observable by external users is essential to assure the accessibility of government processes (Lepri et al., 2018). However, accessibility in the case of algorithms also means that external audit agents or agencies could access algorithms assessing their compliance with ethical, legal, and governance standards, regulations and procedures.

The second dimension of transparency (explainability) of algorithms have much in common with the capacity to explain what they do and how a decision has been made in a determined context. As an example, the European Parliament (2019: 2) states that “explicability should also be available when, on the basis of profiling, citizens are the object of a set of micro-decisions that individually are not very important but which, on the whole, may have a substantial impact on them”. Burrell (2016) insists on the complexity of AI systems (i.e. machine learning) making it difficult to decode the learning capabilities of algorithms or to explain “how an algorithm rendered a certain outcome in an individual case” (Grimmelikhuijsen, 2019: 8). This idea is close to the executability of the code or the assumption that AI systems based on algorithms can act automatically (Introna, 2016). Here, explainability has much more in common with the capacity to monitor the behavior of algorithms during their operation and ensuring human intervention afterwards (according with the General Data Protection Regulation, articles 13–15).

2.4 Algorithms and discretion of public employees’ decisions

Decision-making based on algorithms and discretion of public managers and employees is the other dimension that we explore in this article. Up to now, there is limited empirical research about decisions of bureaucrats based on algorithms (Meijer & Wessels, 2019). AI and algorithms in public agencies seem to evoke the Weberian iron cage, now transformed into a digital fortress, where rules are not readily understood or even available for study (algorithmic opacity) and automated decisions might lead to unintended effects (algorithmic biases) (Young et al., 2019). Besides, it is also expected that algorithms will change the way knowledge-based public employees work and take decisions, so that all aspects of individuals’ performance are quantified, compared to others, and managed against algorithm models, that will carry on important changes requiring deeper understanding (Orlikowski & Scott, 2016). Then, we focus on decisions where AI may either augment or replace human discretion. Following Young et al. (2019), we argue that the relationship of AI to discretion is unique because of three AI’ design features: (a) it is built for automating learning and decision-making processes through abstract mathematical representation of problems; (b) it can utilize input data with speed and dimensionality that vastly outstrip human cognition; and (c) as more data become available, it can “learn” and adjust its behavior by updating its decision heuristics’ (Young et al., 2019: 2).

Therefore, discretion in decision-making is one of the emerging issues regarding algorithms and public managers and employees. At first sight, algorithms can copy the ways in which tacit knowledge is acquired by public employees in knowledge-based organizations, including public administrations, and this may affect personal discretion of the worker to take his/her own decisions (Faraj et al., 2018). Discretion of street-level bureaucrats is one of the traditional topics in public administration (Lipsky, 1971), and ICTs developed the idea of an emerging new system-level bureaucracy (Bovens & Zouridis, 2002). Challenges to discretion go bottom-up affecting street-level bureaucrats (i.e. social workers, doctors, judges, etc.) and public managers (from middle to top), in all cases, working in intensive knowledge-based organizations and units surrounded by algorithms (Barth & Arnold, 1999). Then, decisions based on AI systems seem to encompass rules outperforming traditional limitations of human bounded rationality (Simon, 1991). However, this is not yet clear in all policy and service domains.

Consequently, decision-making based on the use of algorithms might impact the capacity of public employees not only to organize their jobs and autonomy, but also to get fair results or to be accountable for the decisions made by those autonomous systems. At this stage, the study of implications for public officials’ discretion of algorithmic decision-making is critical to discern to what extent their work environment will be challenged by the generalization of algorithms and the potential side effects of their use (Meijer & Wessels, 2019). At the same time, studying algorithmic decision-making in the public sector is important to assess the interaction of public employees with AI systems and whether they understand how algorithms operate and the consequences of their opacity/openness. Also, this might suggest that we are facing the advent of an algorithmic bureaucracy (Vogl et al., 2020), linked to the imbrication of algorithms with traditional forms of public sector organizations and the socio-technical relationships between workers and their work tools.

3.1 SALER. Algorithms to prevent irregularities and promote good government

SALER emerged from the opportunity to take advantage of big data and AI systems to fight fraud and corruption in public administrations. Cases involving corruption practices hit several Spanish regional and local public administrations in the past decade, and particularly in the Valencian region. This required public officials to take measures to prevent these events from happening again (Puncel, 2019), as a new government took office in 2015. This environment included the attention to fiscal consolidation policies and the growing social preoccupation against governmental corruption (Parrado et al., 2018; Ruvalcaba-Gomez et al., 2017; Villoria et al., 2014). Therefore, one could say that in the mid 2010s, corruption arose as a key issue in the Spanish political landscape.

The first change in office at the government of GVA (Generalitat Valenciana, governing body of the Valencian region), after 20 years under the same ruling party, took place in 2015 elections. This situation opened a policy window for the improvement of several internal management control mechanisms. Among others aspects, it was essential to provide the Services Inspectorate (whose functions are detailed below) with new data and technological tools to improve the instruments that were in operation, and that did not allow for easy follow-up contracts or grants procedures (which are the most likely to lead to bad practices) (Puncel, 2019). Thus, SALER system (named from its Spanish acronym “Sistema de Alertas”, in English “Early Warning System”), also known as “SATAN” by Spanish news media, was developed and implemented in the GVA as a new technological framework for the Services Inspectorate. The original purpose of SALER was to enable service inspector’s data analysis on contracts, subsidies, grants, among others, in order to detect, prevent and even anticipate, conflicts of interest and corruption practices by using its potential machine-learning and descriptive-analysis capabilities (Martínez et al., 2018). Figure 1 shows a global scheme of SALER system, whose dimensions are explained below. Next paragraphs describe the technological, organizational, human, and legal dimensions of this system.

3.2 Analytical framework and methods

We have selected semi-structured interviews to gather data and information about this case study. Relying on a case study research strategy in organizational settings (Eisenhardt, 1989; Yin, 2004), using interviews helped us to develop a comprehensive assessment of the story behind actors’ experience with SALER, pursuing in-depth information around the case (Creswell, 2009). Therefore, we have used the semi-structured interview for the following reasons: a) our research design is based on a case study; b) it is an emerging phenomenon, for which we have little information and empirical data; c) we seek to maintain an exploratory logic to foster explicability based on the case study analysis, that is, to understand the experience and obtain more information about the case and the studied concepts.

The interviews were carried out as follows. First, the interviewees selection procedure was based on the chain sampling technique (Guest et al., 2006): initial interviewees recommended other relevant people inside the organization, generating a snowball effect. This technique was appropriate since we initially had few reference contacts within the organization. In total, 6 interviews were carried out (although a few more were scheduled, they could not take place due to the COVID19 crisis). The interviews targeted 3 different types of actors: 1 politician, 2 developers, and 3 service inspectors. Although the number of interviews seems low, the interviews targeted the most relevant people regarding to the development, implementation and usage of the system: the key political champion of SALER the service inspectors who use the system more often, and the IT developers who created the initial prototype. This particular selection criteria have allowed us to get different perceptions, from technical (developers), to political (policy entrepreneur), and non-technical end-users (service inspectors). All interviewees have been active participants during the process of design, adoption, implementation and use of SALER. In the case of service inspectors, they worked in the GVA and the Service Inspectorate even before the system was launched and in some cases, they have been working within this public administration for more than 20 years, keeping organizational memory of changes.

The interviews were recorded in Spanish and manually transcribed and coded by the authors. Furthermore, some interviewees provided us with additional official documentation that has also helped us to understand the case. An interview protocol listing the different questions asked to our respondents can be found at the end of the article (appendix). Each interview was manually analyzed (we did not used specialized software as they were few interviews) using the text processor “comment” function to highlight and code relevant information regarding the dimensions and operationalization. These dimensions and the operationalization of conceptual aspects are summarized in Table 1.

Our subsequent manual coding for the interview questions has been designed around the scheme presented in Table 1. Here, we denote the conceptual elements that we have identified regarding to SALER System, including the notions of algorithmic transparency and decision-making discretion. Each dimension of the interview questions/topic lists was intended to identify and link some of the operationalization indicators as specified by our research questions. Thus, our dimension 1 asked for general questions, identifying the interviewee based on their position inside the organization and his/her relationship with SALER system. Our dimension 2 was intended to gauge the impacts of SALER on the daily work of the Services Inspectorate unit requesting their components and making connections with the functions and perceived discretion of these civil servants to make their own decisions. Our dimension 3 was linked to perceptions about SALER’s level of algorithmic transparency and how they impact on discretionary power. This way of organizing the coding process is connected to the interview protocol (see Appendix). Also, we have intended to somehow reproduce this workflow in the results section.

4.1 Impacts of SALER algorithms on public managers’ work

All the interviewees agreed that one of the basic characteristics of SALER algorithms is their “decision support” nature. This conclusion implies that this system helps GVA service inspectors to use and analyze big data, but that the algorithm results should never be considered as the final decision to be made by the services inspectors. Hence, GVA service inspectors should maintain their supervisory functions. Therefore, they should review the results of the algorithm, considering them as a supplementary indicator to inform their final decisions, and not as the final result of the process itself. Thus, service inspectors have an ample sphere of freedom to take into consideration the indicators displayed by the algorithm, to interpret the findings and to take actions accordingly. To put it differently, SALER system is not making direct decisions, at this stage, it is a starting point to develop further steps into a decision. As exposed by one of the interviewees:

“In no case, an information that we obtain from SALER will be conclusive. This information will be a starting point for us to carry out an investigation. I will give you an example: if we obtain information of SALER that tells us that a certain department has had a percentage of 50% of its contracts with a final price lower than the bid price. Well, based on this information, that 50% seems to be an important number. But we are not going to directly say they are acting reckless. No, what we will do is go in to analyze, see the detail of what those contracts are, see how the algorithm works, and see what type of contract and what kind of procedure is …” (Interview 6)

Since the beginning, the fact that SALER is based on a set of algorithms oriented to support decision-making has promoted a general opinion that this system has not deeply impacted service inspectors’ ability to work autonomously. GVA service inspectors feel that SALER has provided (and will provide in the future) a new direction for the Services Inspectorate unit, and the possibility of carrying out its functions with greater efficiency and efficacy. At the same time, SALER is regarded as an additional source of data and information about detect and verify symptoms of fraud and corruption. In all cases, the involved actors highlight the human dimension behind the decisions made by the system:

“This is one more way of entry, one more way of verification. So, we do not care if this information comes from the algorithmic system, or if it has been reported by a public official, or a union, or that we have discovered it by doing an ordinary investigation. Because we always do not accept the alert until we do a check …”. (Interview 5)

Correspondingly, in the case of other GVA service managers, the impact on their discretion does seem to be perceived negatively. Different interviewees commented that other departments within the GVA have seen SALER’s logical and algorithmic system as a potential threat to their daily work, as some sort of “watchdog” or “big brother”. Despite this fact, service inspectors have strongly stated that SALER algorithmic functions are not related to intrusive monitoring and accusation, but accompaniment, prevention, and help to avoid risks in other departments. One of the interviewers put it as: “initially, part of the civil servants, including the unions, perceived this system as “I am going to persecute the official who is doing it wrong. Until we explained that this was not intended, but rather an accompaniment.” (Interview 1).

One of the most interesting parts of SALER is the risk evaluation maps and self-evaluation plans. These tools are developed with a double perspective. On the one hand, there are risks that have to do with detecting irregularities. On the other hand, there are others based on prioritization factors that have to do with management efficiency. The construction of these maps is covered by the ISO 31000 standard. Again, this tool has raised some concerns among GVA public managers, although the Services Inspectorate unit has tried to explain the role of these tools for support the decisions made by their people (humans) at a later stage of the algorithm operation:

“Because many people believe that you are going to be controlling everything to persecute them. So, of course, people say: hey, let’s see, if I am managing a contracting procedure, are you going to be chasing me with SALER? And we say: no, the other way around. With the risk map we are going to make it easier for you to have the quite position that you are not going to see any irregularity.” (Interview 5)

In addition to make public managers active participants of risks detection, service inspectors are trying to establish a self-evaluation mechanism, allowing each department of the organization to report its own risks and prioritize efficiency. To put it in other words, the purpose is once again to improve the understanding of the system as a whole and as a complementary tool. This might help other public managers to prevent the risks in their own units/departments before letting unethical practices to take place. Besides, SALER will be used to define possible lines of action to mitigate those risks.

Finally, a fundamental element for the protection of autonomy seems to be directly linked to the existence of legal support. Regarding the contents of the regional Law 22/2018, which we have already commented on previous sections, the interviewees stated that it was important for the SALER project, as it provided legal security, both internally and externally, by protecting their functions and appropriately establishing technical and functional mechanisms for SALER. Therefore, it is a law that has a very programmatically and explanatory nature, in addition to its technical-regulatory elements. As one of the interviewees commented:

“And then we passed Law 22/2018, which gives legal support to both the Service Inspection itself and to third parties. In fact, I believe that the sanctioned part is more directed to those who use the system rather than to third parties. And we realized that it was important to have a law, to know to whom the information was passed, who could use this information, how the information could be used, what the algorithms are, how the algorithms report results …” (Interview 1)

4.2 Perceptions on SALER’s algorithmic transparency and impact on discretion

This subsection presents the results of our study regarding to algorithmic transparency and discretion over decisions of public managers in the Services Inspectorate. Here, we have operationalized algorithmic transparency based on two dimensions (Grimmelikhuijsen, 2019): accessibility, that is, source code of the algorithms can be audited by external agents; and explainability, that is, algorithms results are humanly comprehensible. Regarding to the first of these dimensions (accessibility), GVA service inspectors recognize that SALER algorithms can be audited in a simple way, both by independent agents and by actors within the organization. One of these positive impacts over the service inspectors, and other GVA public managers, is based on the fact that the audits are formally recognized in Law 22/2018 itself, so that a bi-annual report has to be carried out. This adds security to both groups, service inspectors, and other civil servants within this regional administration, as the functionalities of the system could be reviewed. As indicated by one of the interviewees:

“In the law itself, it is indicated that a report on system’s audit has to be made every two years. And we can count on external help. It is not exactly defined, but we had even thought to collaborate with the anti-fraud agency, or with some other administrative bodies that have knowledge on these types of questions. To have their opinion on how the algorithms are being developed and how the system is used.” (Interview 4)

Unfortunately, at the time of writing this article no audit had been performed yet, because SALER has been recently implemented (less than two years). The greater or lesser difficulty to carry out these audits does not seem to depend on the code. Here, as SALER is based on simple algorithms, one might expect that it will not be difficult to audit them since they are based on understandable and easily explainable rules. In fact, “It would be quite easy to audit the code. Because, as I told you, the code was not very complex. The interface part, the graphics …were more complex than the warning rules. […] In the end, the tool it is just a rule application. So, it is pretty easy to audit” (Interview 2).

Finally, one of the audit concerns was the potential existence of biases in the code of this software. This fact could act as a barrier during the implementation of the algorithm, in general, and the audit process, in particular. However, the developers of SALER system highlighted that biases primarily could appear in the data, and not in the algorithm itself. Particularly, different interviewees (Interview 3) suggested that frequently data includes biases, and then you have to study it, as biases usually appear in terms of social groups, most affected groups, least affected groups …Then, algorithms are not biased, all that they do is transferring biases engendered in the data manipulated by human.

In relation to the other dimension (explainability) of SALER, interviewees shared the unanimous opinion that this tool produces humanly comprehensible responses. This is a critical aspect of the system for several reasons. The first, and probably one of the most important, is the fact that GVA service inspectors actively participated by generating the “questions” embedded in SALER’s logic system. So this algorithm really responds to their work needs and they can easily handle it and better understand the results:

“The questions we are asking the system are humanly understandable, because it is a translation into a programming language of a question like: “look for contracts made by a person in a specific period of time.” It is not applying strange formulas.” (Interview 4)

On the other hand, the nature of the algorithm itself makes it easier to understand the results and how they have been produced. SALER relies on descriptive models and it based on evidence from data. Therefore, SALER’s operations are properly documented, and there are no “black boxes” in the data crossing and analysis processes. This contrasts with other types of algorithmic models: predictive algorithms, which can be divided into intelligible models (you can trace the prediction) on unintelligible models (it is difficult to trace the prediction as algorithms learn in the process). As explained by one of the developers:

“They are simple algorithms for data analysis. In Artificial Intelligence we talk about “black boxes” or models with very low comprehensibility or intelligibility when using predictive models like neural networks: that would be non-intelligible models. And then we would also have intelligible predictive models, such as decision trees, etc., in which you can trace the reason for a decision. Our case is not like this, we do not use predictive models. They are descriptive models that can always be explained based on evidence.” (Interview 3)

Finally, SALER explainability is also achieved when the results of the algorithms are displayed responding to specific knowledge about each question. Thus, all algorithms’ questions are related to administrative processes. And, therefore, civil servants who are regularly working with contracts, grants, subsidies …will be able to interpret results and acting accordingly. However, if algorithmic complexity increases (for example, with more complicated questions), more expert knowledge on administrative processes will be required, which might imply that certain actors will no longer be able to understand the results:

“At the level we are now, probably yes. Because we are still asking very simple questions, due to the limited availability of data that we have. In the future, it will probably be more complicated. Because certain conclusions will require expert knowledge, not in computer matters, but knowledge of the processes and the rules that support these procedures …The algorithm is obviously essential, but if you do not know the (administrative) process well and the regulations that support these processes, you may think that the algorithm works in one way when it really works the other way round.” (Interview 6)

5.1 Findings

SALER algorithms are having (and probably will continue to have in the future) noticeable impacts on the work autonomy and decisions of civil servants who are actually using them. The use of SALER in the Generalitat Valenciana (GVA) regional government is still at an emerging stage, and its potential as an AI tool has not yet been fully developed. However, this case study facilitates our understanding about the impacts of this kind of algorithmic governance systems on civil servants’ work freedom to perform their duties, and discretion to make their decisions.

From this point, our article also raises some practical implications for public managers involved in the design and implementation of algorithmic governance systems. Overall, algorithmic governance will be a key issue for public agencies in the coming years. Public mangers need to understand the potential implications and consequences of algorithms and AI systems implemented in their public services and policies. Our case study supports this idea and it also highlights the potential cross-departmental nature of AI and algorithmic governance systems. Particularly, SALER system has a direct impact on corruption prevention, transparency, trust, and good governance with comprehensive results across this regional public administration.

Decision-making and discretion of public employees and managers in their span of responsibility within the public sector will be transformed in the near future. Automated decision-making based on AI systems promises to keep challenged some of the basic assumptions in public sector management, with some ideas already described since the seminal age of Simon (1973). The pervasive impacts of AI in public administration will be correctly governed if public officials and managers design adequate policies, foresee organizational impacts, and anticipate operational implications in different functional areas. Surely, how decisions are made will be part of this general reflection on how to adapt public organizations and services to the disruptive capabilities of AI. Here, better decision-making is not only about being more efficient or productive, but also how to create public value for the citizens served and for the good of the society.

At the same time, SALER system bolsters the window of opportunity for transparency and open government in contexts with high perception of corruption in public settings. As it was advanced in previous sections, this concern has been one of the motives of this system from its inception. Fiscal austerity and recovery grants, loans, or transfers will be part of the landscape of future financial mechanisms after COVID-19; hence public administrations with early warning systems and algorithms oriented to fight against corruption and provide trust to lenders, businesses, and society as a whole, will be in a more favorable position in the coming years to design and deliver better public policies and services. AI and algorithmic systems represent an opportunity to innovate public sector management, whereas they also entail challenges and risks for the future of the public sector.

Finally, algorithms like SALER can potentially change the relationships between civils servants, their tools and the way duties and tasks are organized inside public organizations. Ultimately, we may be talking about the evolution towards an algorithmic bureaucracy (Vogl et al., 2020), although in the case of SALER this statement has not been fully accomplished, regarding the emergent nature of this system. Whatever the case, the more algorithms-based systems mature, the more practitioners should encourage the identification of groups potentially impacted by their implementation, inside (and outside) public organizations.