Deep DiveOther Social SciencesSystematic Review
Public Sector AI Governance: A Three-Level Framework for Institutional Restructuring
When governments adopt AI, the real transformation is not technological—it is institutional. Criado, Sandoval-Almazán, and Gil-Garcia (2025) propose a macro-meso-micro framework that treats public sector AI adoption as a question of governance restructuring rather than software procurement.
By Sean K.S. Shin
This blog summarizes research trends based on published paper abstracts. Specific numbers or findings may contain inaccuracies. For scholarly rigor, always consult the original papers cited in each post.
Disclaimer: This post is a research trend overview for informational purposes. Specific findings, statistics, and claims should be verified against the original papers before citation in academic work.
Public Sector AI Governance: A Three-Level Framework for Institutional Restructuring
A municipal government deploys a chatbot for citizen services. The technology team celebrates a successful launch. Six months later, the system is quietly shelved—not because the technology failed, but because no one clarified who was accountable when the chatbot gave incorrect information, which department owned the training data, or how the system aligned with existing administrative law. The failure was institutional, not technical.
This pattern—technological capability outrunning governance capacity—sits at the center of a growing body of research on AI in public administration. A new framework from Criado, Sandoval-Almazán, and Gil-Garcia (2025), published in Public Policy and Administration, attempts to provide the analytical structure that the field has lacked.
The Research Landscape
The study of AI in government has expanded rapidly, but much of the literature treats AI adoption as analogous to earlier waves of digital government—e-government portals, open data initiatives, digital service platforms. Criado et al. (2025) argue that this analogy is inadequate. AI, they contend, is not merely another digital tool to be adopted; it represents a form of institutional restructuring that reshapes how decisions are made, who makes them, and how accountability is distributed.
The paper proposes an analysis framework organized across three levels:
Macro level: Policy environments, regulatory institutions, and national AI strategies. This is the domain of legislation, cross-border governance agreements, and the institutional architecture within which public sector AI operates.
Meso level: Organizations and departments that implement AI systems. Here the questions concern organizational capacity, inter-departmental coordination, procurement processes, and the management structures that determine how AI projects are governed within bureaucracies.
Micro level: Individual officials and citizens who interact with AI systems. This level addresses how street-level bureaucrats respond to algorithmic decision-support, how citizens experience AI-mediated public services, and how individual discretion is affected when algorithms enter administrative processes.
Critical Analysis
The framework's central claim—that AI adoption in the public sector is fundamentally institutional rather than technological—deserves careful examination.
<
| Claim | Source | Strength | Limitation |
|---|
| AI in public administration constitutes institutional restructuring, not mere technology adoption | Criado et al., 2025 | Reframes the research agenda away from technology-centric analysis toward governance questions | The claim is conceptual; empirical validation across diverse administrative traditions would strengthen it |
| A macro-meso-micro framework is needed to systematically analyze AI actors and governance structures | Criado et al., 2025 | Provides analytical clarity by separating levels that are often conflated in the literature | Multi-level frameworks risk becoming descriptive checklists rather than explanatory models if the interactions between levels are underspecified |
| Policy implications differ across the three levels | Criado et al., 2025 | Highlights that a national AI strategy (macro) may have limited effect without organizational capacity (meso) or individual readiness (micro) | The framework's practical utility depends on whether practitioners can operationalize the level distinctions in actual governance decisions |
The multi-level structure echoes established approaches in institutional analysis—Ostrom's multi-tier governance, for instance, or Fountain's technology enactment framework. What Criado et al. add is the explicit argument that AI's characteristics (opacity, adaptiveness, capacity for autonomous decision-making) make it qualitatively different from earlier information technologies. A database system automates record-keeping; an AI system may restructure the decision itself.
This distinction matters for accountability. When a human official denies a permit application, the decision can be appealed through established administrative channels. When an algorithmic system scores applications and officials act on those scores, the locus of the decision becomes ambiguous. Is the official responsible? The algorithm's developer? The department that procured it? The framework's micro level is where these questions become most acute, but the answers depend on macro-level legal structures and meso-level organizational design.
One tension the framework surfaces but does not fully resolve is the relationship between levels. Macro-level policies (such as national AI ethics guidelines) presumably shape meso-level implementation, which in turn affects micro-level experience. But the reverse pathways—how individual officials' resistance or adaptation feeds back into organizational practice and eventually into policy—appear less developed. Institutional change rarely flows in one direction.
Open Questions
Cross-national variation: The framework is proposed as a general analytical tool, but administrative traditions vary enormously. Does the macro-meso-micro structure work equally well in Westminster systems, Napoleonic bureaucracies, and East Asian developmental states? The governance structures at each level may differ so substantially that the same framework captures different phenomena in different contexts.
Temporal dynamics: AI systems change over time—they are updated, retrained, and repurposed. The institutional arrangements governing Version 1 of a system may be inadequate for Version 3. How does the framework account for the iterative, evolving nature of AI deployment?
Democratic legitimacy: If AI adoption is indeed institutional restructuring, then it raises questions beyond efficiency and effectiveness. Who authorizes the restructuring? Citizens have a democratic interest in how their governments make decisions. The framework identifies citizens at the micro level, but the question of democratic consent to institutional change may belong at the macro level—or may cut across all three.
Empirical operationalization: Frameworks are useful insofar as they can be applied. What would a research study look like that simultaneously examines all three levels of AI governance in a single jurisdiction? The methodological challenges of multi-level analysis are considerable.
Closing
Criado, Sandoval-Almazán, and Gil-Garcia's framework offers a needed corrective to research that treats public sector AI as primarily a technology management problem. By insisting that AI adoption is institutional restructuring, they redirect attention to the governance questions that will ultimately determine whether AI in government serves democratic purposes or undermines them.
The framework's value lies less in its novelty—multi-level analysis has a long history in public administration research—than in its specificity to AI. The characteristics that make AI distinctive (opacity, adaptiveness, decision-making capacity) create governance challenges that earlier digital government frameworks were not designed to address. Whether the macro-meso-micro structure proves sufficient to capture those challenges, or whether additional dimensions are needed, is a question that empirical research will need to answer.
What is clear is that the chatbot shelved after six months failed not because of bad code, but because of missing governance. Frameworks like this one may help prevent the next version from meeting the same fate.
Disclaimer: This post is a research trend overview for informational purposes. Specific findings, statistics, and claims should be verified against the original papers before citation in academic work.
Public Sector AI Governance: A Three-Level Framework for Institutional Restructuring
A municipal government deploys a chatbot for citizen services. The technology team celebrates a successful launch. Six months later, the system is quietly shelved—not because the technology failed, but because no one clarified who was accountable when the chatbot gave incorrect information, which department owned the training data, or how the system aligned with existing administrative law. The failure was institutional, not technical.
This pattern—technological capability outrunning governance capacity—sits at the center of a growing body of research on AI in public administration. A new framework from Criado, Sandoval-Almazán, and Gil-Garcia (2025), published in Public Policy and Administration, attempts to provide the analytical structure that the field has lacked.
The Research Landscape
The study of AI in government has expanded rapidly, but much of the literature treats AI adoption as analogous to earlier waves of digital government—e-government portals, open data initiatives, digital service platforms. Criado et al. (2025) argue that this analogy is inadequate. AI, they contend, is not merely another digital tool to be adopted; it represents a form of institutional restructuring that reshapes how decisions are made, who makes them, and how accountability is distributed.
The paper proposes an analysis framework organized across three levels:
Macro level: Policy environments, regulatory institutions, and national AI strategies. This is the domain of legislation, cross-border governance agreements, and the institutional architecture within which public sector AI operates.
Meso level: Organizations and departments that implement AI systems. Here the questions concern organizational capacity, inter-departmental coordination, procurement processes, and the management structures that determine how AI projects are governed within bureaucracies.
Micro level: Individual officials and citizens who interact with AI systems. This level addresses how street-level bureaucrats respond to algorithmic decision-support, how citizens experience AI-mediated public services, and how individual discretion is affected when algorithms enter administrative processes.
Critical Analysis
The framework's central claim—that AI adoption in the public sector is fundamentally institutional rather than technological—deserves careful examination.
<
| Claim | Source | Strength | Limitation |
|---|
| AI in public administration constitutes institutional restructuring, not mere technology adoption | Criado et al., 2025 | Reframes the research agenda away from technology-centric analysis toward governance questions | The claim is conceptual; empirical validation across diverse administrative traditions would strengthen it |
| A macro-meso-micro framework is needed to systematically analyze AI actors and governance structures | Criado et al., 2025 | Provides analytical clarity by separating levels that are often conflated in the literature | Multi-level frameworks risk becoming descriptive checklists rather than explanatory models if the interactions between levels are underspecified |
| Policy implications differ across the three levels | Criado et al., 2025 | Highlights that a national AI strategy (macro) may have limited effect without organizational capacity (meso) or individual readiness (micro) | The framework's practical utility depends on whether practitioners can operationalize the level distinctions in actual governance decisions |
The multi-level structure echoes established approaches in institutional analysis—Ostrom's multi-tier governance, for instance, or Fountain's technology enactment framework. What Criado et al. add is the explicit argument that AI's characteristics (opacity, adaptiveness, capacity for autonomous decision-making) make it qualitatively different from earlier information technologies. A database system automates record-keeping; an AI system may restructure the decision itself.
This distinction matters for accountability. When a human official denies a permit application, the decision can be appealed through established administrative channels. When an algorithmic system scores applications and officials act on those scores, the locus of the decision becomes ambiguous. Is the official responsible? The algorithm's developer? The department that procured it? The framework's micro level is where these questions become most acute, but the answers depend on macro-level legal structures and meso-level organizational design.
One tension the framework surfaces but does not fully resolve is the relationship between levels. Macro-level policies (such as national AI ethics guidelines) presumably shape meso-level implementation, which in turn affects micro-level experience. But the reverse pathways—how individual officials' resistance or adaptation feeds back into organizational practice and eventually into policy—appear less developed. Institutional change rarely flows in one direction.
Open Questions
Cross-national variation: The framework is proposed as a general analytical tool, but administrative traditions vary enormously. Does the macro-meso-micro structure work equally well in Westminster systems, Napoleonic bureaucracies, and East Asian developmental states? The governance structures at each level may differ so substantially that the same framework captures different phenomena in different contexts.
Temporal dynamics: AI systems change over time—they are updated, retrained, and repurposed. The institutional arrangements governing Version 1 of a system may be inadequate for Version 3. How does the framework account for the iterative, evolving nature of AI deployment?
Democratic legitimacy: If AI adoption is indeed institutional restructuring, then it raises questions beyond efficiency and effectiveness. Who authorizes the restructuring? Citizens have a democratic interest in how their governments make decisions. The framework identifies citizens at the micro level, but the question of democratic consent to institutional change may belong at the macro level—or may cut across all three.
Empirical operationalization: Frameworks are useful insofar as they can be applied. What would a research study look like that simultaneously examines all three levels of AI governance in a single jurisdiction? The methodological challenges of multi-level analysis are considerable.
Closing
Criado, Sandoval-Almazán, and Gil-Garcia's framework offers a needed corrective to research that treats public sector AI as primarily a technology management problem. By insisting that AI adoption is institutional restructuring, they redirect attention to the governance questions that will ultimately determine whether AI in government serves democratic purposes or undermines them.
The framework's value lies less in its novelty—multi-level analysis has a long history in public administration research—than in its specificity to AI. The characteristics that make AI distinctive (opacity, adaptiveness, decision-making capacity) create governance challenges that earlier digital government frameworks were not designed to address. Whether the macro-meso-micro structure proves sufficient to capture those challenges, or whether additional dimensions are needed, is a question that empirical research will need to answer.
What is clear is that the chatbot shelved after six months failed not because of bad code, but because of missing governance. Frameworks like this one may help prevent the next version from meeting the same fate.