Governing Open Large Language Models in public administrative decision-making: a framework for accountability, justification, and institutional control

1 Introduction

The accelerated incorporation of artificial intelligence (AI) into public and institutional settings has transformed the conditions under which administrative decisions are produced, justified, and implemented. In recent years, generative systems and Large Language Models have moved from experimental tools to operational infrastructures capable of drafting reports, summarizing records, classifying information, and supporting complex decision processes in legally sensitive environments (Taeihagh, 2025). This shift has expanded the managerial and analytical capacity of public organizations, but it has also intensified concerns about opacity, accountability, and institutional control when algorithmic outputs influence administrative action (Jarrahi, 2018). In this context, the challenge is no longer whether AI can assist administrative decision-making, but under what governance conditions such assistance remains lawful, reviewable, and publicly legitimate.

For readers outside the technical field, Open Large Language Models (OLLMs) can be defined as large-scale, general-purpose artificial-intelligence systems trained on extensive text corpora to generate human-like language, and whose model weights and architecture are released openly under licences that allow institutions to download, inspect, fine-tune and self-host them. They differ from closed, proprietary models—accessible only through a vendor’s interface—precisely because this openness lets public organisations adapt and operate the models on their own infrastructure, which is what makes their institutional adoption both attractive and governance-sensitive.

Within public administrative decision-making, LLMs reach institutions through at least two distinct delivery modes, and the framework proposed here applies to both. The first is the embedded LLM: a model integrated into general-purpose administrative software that civil servants already use—for example, assistants such as Microsoft Copilot built into office suites—where the model operates as one feature among many and its use may be diffuse and difficult to trace. The second is the standalone or purpose-specific model, including self-hosted Open Large Language Models that an institution deploys deliberately for defined tasks. The two modes differ in visibility, configurability and procurement, but accountability, justification and institutional control are equally indispensable in both: an embedded assistant that silently shapes the wording or substance of a decision can raise the same administrative-law concerns as a dedicated system, while sometimes being harder to govern precisely because its presence is less explicit. For this reason the framework is formulated at the level of the decision process rather than of a particular product, so that it applies whether the model is embedded in everyday software or deployed as a distinct institutional tool. It should be stressed that these embedded assistants are discussed here as a contrasting and harder-to-govern case: the primary object of the framework remains open models that an institution can host and adapt, while embedded proprietary assistants illustrate why decision-level governance is needed even where the model itself is closed.

Openness, moreover, should not be read too kindly. Openness is layered and gradient rather than binary: openly released weights and architecture do not, by themselves, provide access to training data, data provenance, evaluation history, labelling practices, post-training procedures, deployment logs, or the infrastructural conditions under which a model is actually run. This matters directly for the present argument, because the framework’s claims about auditability and institutional control depend on exactly those layers. Recent work shows that ‘open’ AI can sit on top of concentrated power in compute, data and licensing, so that openness may be more marketing than meaningful transparency (Widder et al., 2024), and that accountability is distributed across algorithmic supply chains of many actors rather than residing in a single institution (Cobbe et al., 2023). A realistic governance framework must therefore treat openness as a set of verifiable disclosures to be required and checked—not as a property that can be assumed—and recognise that transparency is necessary but not sufficient for accountability (Ananny and Crawford, 2018).

From the perspective of public administration and legal governance, the use of Open Large Language Models (OLLMs) introduces a particularly significant problem because their openness, adaptability, and low barriers to deployment increase the probability of institutional adoption across heterogeneous settings. Unlike closed proprietary systems, OLLMs can be customized, audited, and integrated into internal workflows of public institutions, universities, and regulatory agencies, which makes them attractive for organizations seeking cost-effective and flexible AI solutions (Hildebrandt, 2018). At the same time, this very openness raises difficult questions regarding traceability, responsibility allocation, quality control, and the possibility of reproducing institutional bias under the appearance of technical neutrality (Grimmelikhuijsen and Meijer, 2022). These tensions place OLLMs at the intersection of administrative law, governance theory, and organizational decision-making.

Existing scholarship has addressed parts of this problem through adjacent but still fragmented lenses. Research on AI in organizations has emphasized augmentation, efficiency, and changing managerial roles, while public-sector AI studies have focused on adoption, service delivery, and state capacity (Valle-Cruz et al., 2020). In parallel, legal and ethical analyses have highlighted transparency, explainability, due process, and accountability as necessary conditions for high-stakes AI use (Busuioc, 2021; Grimmelikhuijsen and Meijer, 2022). However, much of this literature treats these dimensions separately, leaving unresolved how OLLMs should be governed when they become embedded in administrative processes that require lawful justification, procedural fairness, and institutional oversight (Fink and Finck, 2022; Hildebrandt, 2018). This fragmentation reveals a conceptual and operational gap between AI capability and administratively valid decision support.

This article addresses that gap by proposing a governance-oriented approach to OLLM-mediated administrative decision-making grounded in three interrelated constructs: accountability, justification, and public institutional control. First, accountability is understood as the capacity to attribute responsibility for decisions supported or shaped by OLLMs (Ben-Michael et al., 2024). Second, justification refers to the requirement that decisions be supported by reasons that can be examined, challenged, and defended within an administrative and legal framework (Dokumacı, 2024). Third, public institutional control refers to the organizational mechanisms (rules, procedures, audit structures, and oversight practices) through which institutions constrain and supervise the use of these models in practice (Gasser and Almeida, 2017). Together, these constructs make it possible to analyze OLLMs not merely as technical tools, but as governance objects embedded in decision systems.

The main objective of this paper is to develop a conceptual framework for governing OLLMs in administrative decision-making, with particular attention to accountability, justification, and institutional control in public-sector and public-interest environments. The article argues that current approaches centered on efficiency, automation, or technical explainability are insufficient for contexts in which decisions must satisfy legal and administrative standards of motivation, reviewability, and responsibility (Busuioc, 2021). Its main differentiator from similar studies is that it does not examine OLLMs only as technological innovations or as abstract regulatory challenges; instead, it positions them within the concrete logic of administrative action, where decision support must remain compatible with institutional legitimacy and the rule of law (Grimmelikhuijsen and Meijer, 2022). On this basis, the following sections review the relevant literature, define the administrative justification problem, and propose a governance framework for OLLM deployment in legally sensitive decision contexts.

2 Related literature

Research on AI in organizations has consistently shown that the managerial significance of AI lies not only in automation, but in its capacity to reshape decision structures, redistribute expertise, and alter the boundaries of organizational judgment. Early organizational studies emphasized human–AI complementarity, arguing that AI systems create value when they augment rather than fully replace managerial cognition, particularly under conditions of uncertainty, complexity, and equivocality (Shrestha et al., 2019). More recent work has deepened this view by examining how AI changes organizational capabilities, innovation processes, and the allocation of authority in decision-making environments (Benbya et al., 2021). Together, these studies establish that AI adoption is not a purely technical phenomenon, but an organizational transformation with implications for management theory and institutional design (Jarrahi, 2018). This characterisation should be read as indicative rather than as a settled consensus, since it rests on a still-developing body of work; it is, however, broadly corroborated across organisational and public-sector AI scholarship (Shrestha et al., 2019; Wirtz et al., 2022).

A second strand of literature has focused specifically on the public sector, where AI is increasingly framed as a tool for improving policy design, service delivery, and administrative efficiency. This body of work highlights both the opportunities and the structural challenges of deploying AI in public institutions, including fragmented implementation, weak institutional readiness, and difficulties in aligning technological innovation with legal and democratic standards (Kolkman et al., 2024). Studies on AI in government and the public policy cycle have shown that algorithmic systems can influence multiple stages of administrative action, from agenda setting and policy evaluation to case management and resource allocation (Valle-Cruz et al., 2020). However, they also show that the public sector faces distinctive constraints related to legitimacy, transparency, and procedural fairness that are not always present in private-sector deployments.

A more recent and rapidly expanding literature has begun to address generative AI and Large Language Models in institutional settings, including higher education, public administration, and governance. Much of this work has concentrated on opportunities for productivity, communication, and knowledge management, while also identifying significant risks such as hallucination, data leakage, opacity, and the amplification of organizational bias (Sobrino-García, 2021). In higher education and public institutions, these concerns are magnified because AI systems may influence decisions that affect rights, access, evaluation, or the distribution of public resources (Taeihagh, 2025). Recent policy-oriented research on generative AI governance therefore emphasizes that institutional adoption requires more than technical deployment; it requires adaptive governance arrangements capable of managing uncertainty, safeguarding accountability, and preserving public trust.

Two clarifications are warranted before integrating these literatures. First, although several foundational works on algorithmic accountability and public-sector AI predate the public release of generative chatbots in late 2022, they remain relevant because the institutional problems they identify—opacity, responsibility attribution and contestability—are intensified rather than replaced by OLLMs. Second, OLLMs are not simply another instance of the predictive machine-learning systems discussed in that earlier literature: their generativity, fluency and apparent reasoning make their outputs more persuasive and harder to challenge, their general-purpose nature blurs the boundary of authorised use, and their openness distributes deployment across many actors. These features sharpen, rather than merely repeat, the governance difficulties previously associated with classificatory or predictive AI (Taeihagh, 2025).

This is also reflected in a broader policy shift. Between roughly 2017 and 2019, AI governance was dominated by non-binding ethical principles building on earlier soft-law instruments such as the OECD (2019) Recommendation on Artificial Intelligence; since then, several jurisdictions have moved toward more substantive and binding arrangements, including the European Union’s AI Act, Canada’s Directive on Automated Decision-Making and its proposed Artificial Intelligence and Data Act, and emerging statutory or strategic instruments in China, South Korea and Brazil. This evolution from light ethical guidance to enforceable, risk-based regulation provides concrete reference points for operational governance frameworks and signals that institutional adoption of AI is increasingly expected to meet legal, and not merely ethical, standards (Veale and Zuiderveen Borgesius, 2021; Wirtz et al., 2022; Taeihagh, 2025).

A further body of literature, so far underused in this debate, comes from administrative law on automated decision-making, explainability and the duty to give reasons. This scholarship clarifies why the legal status of an automated decision differs from that of a human one: administrative law generally requires that decisions affecting rights be supported by adequate, reviewable reasons, and AI-mediated decisions strain this requirement when their rationale cannot be reconstructed (Finck, 2020; Fink and Finck, 2022; Hildebrandt, 2018). For OLLMs this raises a specific question that the present framework foregrounds: whether a civil servant who uses an OLLM to draft or support a decision is, in legal terms, equivalent to a system making the decision. The difference matters because responsibility, the duty to give reasons and the possibility of contestation remain attached to the human and institutional act, even when the linguistic content originates in the model (Brand, 2020).

Despite these advances, the literature remains fragmented in a way that is highly relevant for this paper. Organizational studies explain how AI alters managerial decision structures, public administration studies analyze adoption and policy implications, and governance-oriented work identifies ethical and institutional risks (Taeihagh, 2025). Yet these streams rarely converge around a single question: how should OLLMs be governed when they become embedded in administrative decision-making processes that require accountability, justification, and institutional control? (Wirtz et al., 2022). Existing scholarship provides important insights into augmentation, public-sector implementation, and AI governance, but it does not sufficiently integrate these dimensions into a framework specifically oriented to OLLMs in administratively consequential environments (Benbya et al., 2021). This unresolved gap motivates the present study and justifies the need for a governance framework tailored to legally and institutionally sensitive decision contexts.

3 The administrative justification problem

The increasing use of AI in public and institutional decision-making has made clear that the central challenge is not only whether these systems are efficient, but whether the decisions they shape remain administratively justifiable. In administrative settings, a decision must be more than useful or technically accurate; it must also be grounded in reasons that can be communicated, reviewed, and contested within an institutional framework (Young et al., 2019). Recent legal and public-administration scholarship shows that algorithmic and AI-supported decisions can undermine this requirement when they displace or obscure the reasoning that links facts, norms, and discretionary judgment (Bovens and Zouridis, 2002; Zalnieriute et al., 2019). This concern has been emphasized both in public-law analyses of automated decision-making and in public-administration research on AI adoption in government.

This problem becomes sharper with OLLMs because they generate outputs that are often coherent and persuasive, yet not necessarily anchored in the standards that administrative action requires. OLLMs can draft reports, summarize records, classify cases, and suggest courses of action, but the model’s fluency does not guarantee that the recommendation is normatively grounded, procedurally fair, or institutionally controllable (Balo and Khmelnytskyi, 2023). Public-sector AI research has shown that when AI systems enter government workflows, they affect not only speed and consistency but also the distribution of responsibility, the visibility of reasoning, and the practical ability of administrators and affected parties to interrogate decisions (Cheong, 2024). In this sense, OLLMs intensify an already recognized tension in public administration: decisions may appear rational and efficient while becoming less transparent and less accountable.

From a legal and governance perspective, the administrative justification problem can therefore be defined as the gap between algorithmically plausible outputs and institutionally defensible decisions. In public administration, justification performs at least three functions: it constrains power, enables accountability, and protects the right of affected subjects to understand and challenge the grounds of administrative action (Khairullah et al., 2025). Studies on accountable AI and algorithmic decision-making consistently stress that information deficits, opaque model behavior, and weak auditability make it difficult to assign responsibility or to sustain meaningful review of decisions influenced by AI (Young et al., 2019). This is especially problematic in administrative environments, where legitimacy depends not only on outcomes but also on the procedural integrity of how those outcomes are reached.

The issue is not simply that OLLMs “lack explanations,” but that explanation alone is insufficient for administrative law and public institutional control. A technically interpretable output may still fail to provide reasons that satisfy standards of legality, proportionality, consistency, and contestability (Agarwal and Nene, 2025). Emerging work on justifications in AI-supported decisions suggests that institutions need models capable of producing reasons that support challenge, review, and governance across the system lifecycle, rather than post hoc descriptions of how a model behaved (Kolářová and Schmude, 2026). For this reason, the next section advances a governance-oriented framework that treats accountability, justification, and institutional control as explicit requirements for the use of OLLMs in administrative decision-making.

4 Methodological approach

This article adopts a conceptual and governance-oriented methodological approach to examine Open Large Language Models (OLLMs) as political and institutional technologies embedded in administrative decision-making. The purpose of the study is not to conduct an empirical assessment or a systematic literature review, but to construct a theoretically grounded governance framework for understanding how OLLMs reshape accountability, administrative justification, and institutional control in public-interest decision processes (Jaakkola, 2020). This approach is appropriate because the article addresses a politics-of-technology problem: the use of OLLMs does not merely introduce a new technical tool into public organizations, but potentially redistributes authority, alters justificatory practices, and changes the conditions under which administrative decisions remain reviewable, contestable, and legitimate.

The methodological design combines conceptual analysis, integrative literature synthesis, and theory synthesis (Jaakkola, 2020). Conceptual analysis was used to define the three core constructs of the article: accountability, administrative justification, and public institutional control. Integrative literature synthesis was used to connect research on AI in organizations, public-sector AI adoption, algorithmic accountability, administrative legitimacy, explainability, and generative AI governance. Theory synthesis was used to translate these conceptual and policy insights into a governance cycle for OLLM-mediated administrative decision-making. This methodological combination is suitable for a conceptual article because the main contribution is a conceptual governance framework that organizes how institutions should authorize, justify, oversee, and audit the use of OLLMs in decision processes. The basis of the conceptual analysis is theoretical rather than empirical: each of the three constructs was defined by drawing on established meanings in administrative-law and public-administration scholarship—accountability from work on algorithmic accountability and answerability, administrative justification from the administrative-law duty to give reasons, and institutional control from public-sector oversight and discretion theory—and these definitions were then used as analytical lenses to read and organise the literature (Jaakkola, 2020). Accordingly, this study is a conceptual article in Jaakkola’s (2020) sense and does not adopt a design-science methodology; the earlier references to design-science research have been removed. For this genre the absence of empirical testing is appropriate rather than a deficiency: the framework is offered to be required, structured and assessed in practice, and is illustrated below through a worked example.

The literature was selected purposively rather than exhaustively. Priority was given to peer-reviewed and policy-relevant sources addressing at least one of the following themes: AI in public administration, algorithmic decision-making, administrative legitimacy, accountability, explainability, justification, institutional oversight, human review, and governance of generative AI or Large Language Models. This selection strategy was appropriate because the study does not seek to estimate prevalence or aggregate empirical effects, but to synthesize conceptually relevant contributions across public administration, AI governance, legal theory, organizational studies, and politics of technology (Jaakkola, 2020). The resulting corpus allowed the article to identify a gap between discussions of AI capability and the institutional requirements of administratively defensible decision support. In operational terms, sources were identified through Scopus, Web of Science and Google Scholar for the social-science and computing literature, and through SSRN and legal repositories for the administrative-law literature, covering the period from 2017 to 2025. Inclusion required that a source address at least one of the themes listed above and contribute conceptually to the definition of the constructs or to the design of the governance framework; purely technical benchmarking studies without governance relevance were excluded. Because the synthesis is integrative and conceptual rather than systematic, the resulting corpus is purposive and illustrative, not exhaustive, which is acknowledged as a limitation. To make this selection auditable, the search drew on Scopus, Web of Science, the ACM Digital Library, HeinOnline/SSRN (for administrative-law sources) and Google Scholar, using combinations of terms such as ‘algorithmic/automated decision-making’, ‘administrative justification/duty to give reasons’, ‘algorithmic accountability’, ‘AI governance’, ‘Large Language Models’, ‘public administration’ and ‘rule of law’, for the period 2002–2026. From an initial pool of several hundred records, approximately sixty sources were retained after screening for direct relevance to the three constructs or to the framework’s design; each retained source is associated in the analysis with the construct or framework stage it informs. Because the synthesis is integrative and theory-oriented rather than systematic, the corpus is purposive and illustrative, which is a deliberate feature of the conceptual-article genre rather than an attempt at exhaustive coverage.

The analytical procedure followed four stages. First, the study identified the institutional problem created by OLLM use in administrative decision-making: fluent and plausible model outputs may influence decisions without satisfying standards of responsibility, justification, reviewability, and public control. Second, the article derived three governance requirements from the literature. Accountability refers to the capacity to attribute responsibility for AI-supported decisions. Administrative justification refers to the production of reasons that can be documented, reviewed, and challenged. Public institutional control refers to the organizational capacity to authorize, constrain, supervise, and audit the use of OLLMs. Third, these requirements were translated into a four-stage governance cycle composed of authorization, justification design, oversight and validation, and audit and feedback. Fourth, the framework was refined through the distinction between decision support and decision substitution, since OLLMs may assist administrative reasoning but should not replace the institutional act of justification in legally or procedurally sensitive contexts.

The framework was constructed through a design logic in which the main research output is a conceptual governance architecture. In this article, the artifact is not a software system, but a process-oriented model specifying how public and public-interest institutions should govern OLLMs when their outputs enter administrative decision chains. The framework was designed to satisfy four criteria. First, it had to be institutionally grounded, meaning that it had to reflect the legal, administrative, and political conditions under which public decisions must be justified. Second, it had to be process-oriented, meaning that governance should be embedded across the decision cycle rather than applied only after a decision has been made. Third, it had to be control-sensitive, meaning that the framework should identify points at which human authorization, validation, logging, escalation, and audit are required. Fourth, it had to be adaptable, so that different institutions can calibrate the framework according to the legal sensitivity, rights impact, and administrative stakes of each use case.

The methodological contribution of this approach lies in translating broad AI governance principles into administratively actionable requirements. Rather than treating accountability, transparency, explainability, and human oversight as abstract values, the framework specifies how these principles can be operationalized through authorization rules, justificatory standards, oversight mechanisms, validation checkpoints, audit trails, and feedback procedures. This makes the framework especially relevant for politics-of-technology scholarship because it examines how a specific class of AI systems can reshape institutional authority and how public organizations can preserve democratic and administrative control over AI-assisted decisions.

The study has methodological limitations. Because it develops a conceptual governance framework, it does not provide empirical validation through case studies, interviews, experiments, simulations, or implementation data. Its contribution is therefore theoretical and normative-operational rather than empirical. However, this limitation is consistent with the article’s objective: to establish a structured framework that can guide future empirical research, institutional pilots, comparative studies, and policy evaluation. Future work should test the framework in concrete administrative settings, examine how institutions implement justification-based controls, and develop indicators to assess whether OLLM-supported decisions remain accountable, reviewable, contestable, and institutionally legitimate.

5 A governance framework for OLLM-mediated administrative decision-making

To respond to the administrative justification problem, this paper proposes a governance framework for the use of OLLMs in administrative decision-making. The framework is built on the premise that OLLMs should not be treated merely as productivity tools or conversational interfaces, but as components of institutional decision systems whose outputs can affect rights, procedures, and organizational accountability (Ananny and Crawford, 2018). Public-administration scholarship has already shown that AI changes the conditions of discretion in government and requires institutions to reconsider how authority, control, and review are distributed across human and technical actors (Eisenberg et al., 2025). In parallel, work on internal algorithmic auditing has demonstrated that accountability gaps emerge when organizations adopt AI without end-to-end governance mechanisms embedded across the lifecycle of development and deployment (Raji et al., 2020).

The proposed framework is organized around three core dimensions: accountability, administrative justification, and institutional control. The first dimension, accountability, concerns the attribution of responsibility for decisions supported or shaped by OLLMs. In administrative contexts, it must remain possible to identify who validated, adopted, or operationalized an AI-generated recommendation (Agbese et al., 2023). The second dimension, administrative justification, requires that outputs influencing decisions be linked to reasons that can be explained, documented, and challenged within an organizational and legal process (Brand, 2020). The third dimension, institutional control, refers to the procedures, oversight routines, escalation channels, and audit practices through which organizations constrain the use of OLLMs and prevent unauthorized or unreviewable decision substitution (De-Arteaga et al., 2020). These dimensions are consistent with the literature on public-sector AI, which stresses that AI systems must be governed not only through principles but through operational arrangements that preserve legitimacy and reviewability. To avoid restating the problem as its own solution, a distinction is drawn here between the administrative justification problem—the empirical gap between plausible outputs and defensible decisions diagnosed in Section 3—and administrative justification as a governance dimension, understood as the institutional capacity and obligation to produce reviewable reasons. The contribution of the framework is precisely to convert the former (a deficit) into the latter (a designed capacity) through the four-stage cycle; the dimension is therefore the response to the problem, not a renaming of it.

Operationally, the framework can be understood as a four-stage governance cycle. The first stage is authorization, in which the institution determines whether an OLLM may be used for a given administrative task and under what scope. The second stage is justification design, in which the organization specifies the minimum justificatory requirements for any OLLM-supported output, including traceable grounds, decision logs, and the conditions under which human review is mandatory (Fink and Finck, 2022; Brand, 2020). The third stage is oversight and validation, in which outputs are checked against institutional rules, legal standards, and procedural safeguards before they are relied upon in practice (Hildebrandt, 2018; Zalnieriute et al., 2019). The fourth stage is audit and feedback, in which the system’s use is retrospectively assessed to identify errors, biases, procedural failures, and governance gaps (Birkstedt et al., 2023). This cyclical structure draws on the logic of internal algorithmic auditing and also reflects public-policy research showing that AI can affect multiple stages of administrative and policy processes, not only isolated decisions.

A key contribution of this framework is that it distinguishes between decision support and decision substitution. OLLMs may appropriately assist administrators in summarizing records, drafting alternatives, identifying relevant precedents, or organizing evidence, but they should not be allowed to replace the institutional act of justification unless governance mechanisms can ensure traceability, challengeability, and compliance with administrative standards (Shrestha et al., 2019; Young et al., 2019). This distinction is especially important because research on generative AI repeatedly notes that these models can produce convincing but unreliable or unverifiable outputs, creating risks when organizational actors mistake fluency for validity (Huang et al., 2025). In public institutions, that mistake can transform AI from a support mechanism into an opaque source of discretionary power.

5.1 A worked example: a single decision through the four-stage cycle

To show how the cycle operates rather than merely listing requirements, consider one illustrative decision: a student requests a tuition-fee waiver on hardship grounds, and an administrator uses an open language model to help prepare the decision. The example is hypothetical and is used only to demonstrate what each stage catches.

Authorization. The institution has classified hardship-waiver decisions as rights-affecting and therefore high-risk; under the authorization rule the model may be used only to summarise the file and draft reasoning, not to determine the outcome, and only over documents the handling officer is entitled to see. This stage catches an attempt to let the model decide eligibility directly, which is out of scope.

Justification design. The minimum justificatory requirements specify that every factual claim in the draft must cite a document in the case file (the application, the income evidence, the relevant regulation), that the applicable rule and its criteria must be stated, and that a decision log records the model version, the inputs, and the responsible officer. When the model produces a fluent sentence asserting that the student ‘does not meet the income threshold’ without pointing to the income evidence or the regulation, this stage catches the unsupported claim: a plausible output cannot become a reason until it is grounded.

Oversight and validation. Before issuance, the grounded draft is checked against the regulation and procedural safeguards: does the cited threshold match the current regulation version, are required steps (notice, opportunity to respond) documented, is the recommendation consistent with the evidence? Here the check catches a stale regulation citation and a missing procedural step, both corrected before the decision binds anyone.

Audit and feedback. After issuance, a sample of waiver decisions is reviewed for error patterns—for example, systematic over-refusal for a particular applicant group—and disagreements between the model draft and the officer’s final decision are logged and fed back to recalibrate the justificatory requirements and the authorization scope.

Across these stages the cycle does concrete work: it prevents decision substitution (authorization), forces a plausible output to be tied to documented, challengeable reasons (justification design), tests those reasons against the law and procedure before they bind anyone (oversight), and creates the record that makes the decision contestable and the system improvable (audit). This is how the cycle is intended to convert the administrative justification gap into an institutional capacity rather than merely restate it; the same logic could be specified for advising, admissions or quality-assurance decisions.

In this sense, the framework is not only a normative proposal but also an administrative design tool. It provides a way to align OLLM adoption with institutional mandates by embedding review, justification, and control into the decision process itself (Elizabeth Walsh and Feigh, 2021). Rather than asking whether OLLMs are generically useful, the framework asks whether their institutional use is governable under public-law and administrative standards (Gursoy and Kakadiaris, 2022; Hirvonen, 2024). This shift is essential for sectors such as public administration and educational governance, where the legitimacy of decisions depends not only on efficiency or innovation, but on the continuing possibility of explanation, oversight, and contestation (Balasubramaniam et al., 2023). The next section discusses the implications of this framework for policy, institutional design, and future research on AI-mediated administrative governance.

6 Discussion and implications

The framework developed in this paper reframes OLLM adoption in administrative settings as a governance problem rather than a mere technology deployment problem. This shift is important because much of the current literature on AI in public institutions emphasizes efficiency, innovation, or service improvement, while under-specifying the conditions under which AI-assisted decisions remain institutionally defensible (De-Arteaga et al., 2020). By centering accountability, administrative justification, and institutional control, the framework responds to public-administration research that treats AI as a source of altered discretion and redistributed authority rather than as a neutral support tool (Kolkman et al., 2024). In that sense, the paper extends work on AI in government by specifying how OLLMs should be constrained when their outputs enter legally consequential decision chains.

A first implication concerns administrative law and procedural legitimacy. If OLLMs are introduced into workflows that affect rights, obligations, access, or sanctions, institutions cannot rely on technical performance alone; they must ensure that decisions can still be traced, reviewed, and challenged (Grimmelikhuijsen and Meijer, 2022). The literature on algorithmic decision-making in public law has already shown that automated systems test the adequacy of traditional administrative safeguards, especially where reasoning becomes opaque or difficult to reconstruct (Busuioc, 2021). The proposed framework contributes to that debate by making justification an operational governance requirement rather than an abstract legal aspiration (Dokumacı, 2024). This is particularly relevant for public agencies and educational institutions, where procedural fairness and reviewability are core sources of legitimacy.

A second implication concerns institutional design. The distinction introduced here between decision support and decision substitution suggests that organizations should not evaluate OLLMs only in terms of accuracy or usefulness, but also in terms of where they intervene in the administrative process and how much discretionary weight they carry (Khairullah et al., 2025). This means that governance arrangements must be task-sensitive: the use of an OLLM for drafting internal memoranda is not equivalent to its use for recommending sanctions, admissions, eligibility outcomes, or compliance determinations (Young et al., 2019). Public-sector and organizational AI studies consistently show that the effects of AI depend on workflow integration, oversight routines, and role distribution (Khairullah et al., 2025). Accordingly, the framework implies that administrative institutions need differentiated authorization rules, review thresholds, and logging practices tied to the stakes of each use case. More broadly, because OLLMs increasingly reach administrations embedded in everyday software as well as through dedicated deployments, this justification-oriented framework should be understood not as a standalone instrument but as an essential component of an institution’s overall AI-governance arrangements in the public sector.

A third implication is methodological. Current debates on trustworthy AI often move from ethical principles to broad governance recommendations, but they do not always provide a middle layer capable of translating those principles into administratively testable requirements (Gasser and Almeida, 2017). The framework proposed here offers that intermediate layer by specifying what must be present for an OLLM-assisted output to count as institutionally usable: a recognizable source of responsibility, reasons that can be documented and challenged, and oversight mechanisms that preserve control over the final decision (Kolářová and Schmude, 2026). This contributes to the literature on AI governance by complementing principles-based approaches with a process-oriented design that institutions can adapt to their own administrative settings.

To make the claimed gap explicit and to show where the contribution lies, Table 1 positions the framework against existing governance and assurance instruments by the functions each one covers. The comparison indicates that prior instruments are strong on some functions—impact assessment and auditing on documentation and review, the EU AI Act on binding control—but that none combines decision-level authorization, contestable justification, institutional control, auditability, attention to language-model-specific risks (such as fluent but ungrounded outputs), and the open-model and supply-chain considerations discussed above, while distinguishing decision support from decision substitution. That specific combination, rather than any single component, is where the framework’s contribution lies.

Table 1. Functional comparison of the proposed framework with existing governance and assurance instruments.

‘Yes’ = function explicitly addressed; ‘~’ = partially addressed; ‘No’ = not addressed. The open-model/supply-chain column is marked partial for the proposed framework because it foregrounds layered openness and supply-chain accountability as requirements to be checked (Widder et al., 2024; Cobbe et al., 2023) without resolving the underlying concentration of compute and data.

The framework can also be situated within the literature on algorithmic and AI impact assessments, which proposes structured, often ex-ante, evaluation of an AI system’s potential harms before and during deployment (Selbst, 2021; Raji et al., 2020). The governance cycle proposed here is complementary rather than redundant: whereas impact assessments are typically system-level and conducted prior to deployment, the present framework operates at the level of the individual administrative decision and across its life cycle, focusing specifically on reason-giving and reviewability. In practice, an institution could embed this justification-oriented cycle as the in-process component of a broader impact-assessment regime, using the assessment to authorise a use case and the cycle to govern each decision the system subsequently supports.

The framework also has implications for higher education governance, which is especially relevant when public universities and educational systems adopt OLLMs for administrative analytics, student classification, academic integrity processes, or internal policy support. Recent work on generative AI in education shows that institutions are already facing trade-offs between productivity gains and risks related to transparency, data governance, fairness, and academic integrity (Eisenberg et al., 2025). In those environments, the administrative justification problem becomes particularly visible because educational decisions often combine managerial, legal, and pedagogical dimensions (Kolkman et al., 2024). The present framework suggests that educational institutions should govern OLLMs not only as educational technologies, but as components of institutional decision architectures subject to accountability and review.

Finally, this discussion points toward a broader research agenda. Future work should test how justification-based governance performs in practice when compared with explanation-only approaches, particularly in settings where OLLMs are embedded in public-sector or educational administration (Sobrino-García, 2021). It should also examine how institutions can calibrate human oversight without creating symbolic review processes that merely validate AI outputs ex post (Balo and Khmelnytskyi, 2023). More broadly, the paper suggests that the next stage of research on OLLMs in public administration should move beyond adoption and ethics toward governable architectures of administrative use, where institutional legitimacy is treated as a design parameter rather than a downstream concern.

7 Conclusions and future work

This paper has argued that the deployment of OLLMs in administrative decision-making should be understood not only as a technological development, but as a problem of governance, justification, and institutional legitimacy. The central claim is that OLLMs become administratively consequential when their outputs influence decisions that affect rights, obligations, access, or organizational accountability. Under those conditions, usefulness and fluency are insufficient. What matters is whether institutions can preserve responsibility, provide reasons that remain open to scrutiny, and maintain effective control over decision processes.

The main contribution of the paper is the proposal of a governance framework structured around accountability, administrative justification, and institutional control. By introducing these dimensions as explicit requirements for OLLM-mediated administrative use, the study moves beyond approaches centered exclusively on efficiency, automation, or technical explainability. In doing so, it offers a way to translate legal and administrative concerns into a governance-oriented model that can guide institutional adoption in public-sector and public-interest environments. The framework also contributes conceptually by distinguishing decision support from decision substitution, a distinction that is critical for evaluating the legitimacy of AI-assisted administrative action.

A second contribution lies in the paper’s interdisciplinary positioning. Rather than treating public administration, legal theory, and AI governance as separate conversations, the article brings them together around a common question: how can institutions use OLLMs without eroding reviewability, contestability, and procedural fairness? This integrative perspective is especially relevant at a time when public organizations and educational institutions are under pressure to adopt generative AI while still complying with administrative standards and public expectations of transparency and responsibility.

At the same time, the paper has limitations. The proposed framework is conceptual and has not yet been empirically validated through case studies, simulations, or institutional implementation. It therefore does not claim to demonstrate effectiveness in practice, but to provide a theoretically grounded model for future application and testing. In addition, the discussion has been framed at a relatively general level across administrative settings, which means that further work is needed to adapt the framework to specific sectors, jurisdictions, and decision contexts.

Future research should therefore proceed in at least three directions. First, empirical studies are needed to examine how OLLMs are currently being used in administrative environments and whether existing governance practices are capable of preserving accountability and reviewability. Second, comparative work should explore how different legal and institutional systems define acceptable levels of human oversight, traceability, and justificatory adequacy in AI-assisted decisions. Third, methodological research should develop operational indicators and evaluative tools capable of assessing whether an OLLM-supported administrative process satisfies the requirements proposed in this framework.

Ultimately, the paper suggests that the next stage of research on generative AI in public administration should not focus only on whether these systems can improve performance, but on whether they can be institutionally governed in ways compatible with lawful and legitimate administrative action. If OLLMs are to become part of decision systems in public and educational institutions, they must be embedded within governance architectures that preserve not only efficiency, but also justification, accountability, and public trust.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Charalampos Harris Alexopoulos, University of the Aegean, Greece

Reviewed by: Theodoros Papadopoulos, University of the Aegean, Greece

Dirk Brand, Stellenbosch University, South Africa