The most profound transformation introduced by advanced artificial intelligence into the public sphere does not coincide with the automation of procedures, nor with the acceleration of administrative timeframes. It coincides with the progressive transfer of institutional rationality into computational architectures capable of classifying, predicting, ordering priorities, and distributing attention. Public administration is not simply made more efficient: it is reconstructed as a predictive system, in which files, acts, and proceedings tend to become informational flows subject to models, thresholds, correlations, and probabilistic assessments.

The shift is structural because it modifies the place where judgment is formed. In classical bureaucracy, the decision was inscribed within a visible chain of competences, rules, responsibilities, and formal acts. In algorithmic bureaucracy, a growing part of the decision moves upstream, into the design of the system: into the choice of data, the definition of objectives, the configuration of models, and the weighting of acceptable errors. Discretion does not disappear, but it changes location. It no longer belongs only to the official who interprets a rule; it also belongs to those who design the technical infrastructure within which that rule is made operational.

The OECD observes that governments use AI to design better policies, make more informed decisions, improve relations with citizens, and increase the quality of services, but it also emphasizes that risks and limits must be considered an integral part of the adoption process. Its observatory on AI policies now records more than 900 national policies and initiatives, a sign that the issue no longer concerns marginal experiments, but the ordinary redefinition of state capacity.

The economic point is decisive. Every public administration is also an allocative device: it distributes time, resources, access, authorizations, rights, controls, and priorities. When AI enters this device, it does not merely produce organizational efficiency. It affects the formation of public value, because it orients which practices move forward, which anomalies emerge, which risks are made visible, which needs are anticipated, and which subjects remain statistically marginal. The administrative machine thus becomes a machine of economic selection, not only of legal selection.

The promise of administrative AI lies in the possibility of reducing inertia, duplication, opacity, and inequalities of access. An administration capable of reading weak signals, territorial needs, recurring fraud, or procedural congestion in real time can free resources and improve the quality of expenditure. Yet the same infrastructure can generate a new form of asymmetry: the asymmetry between those who understand the logic of the system and those who merely undergo its outcomes. The citizen risks no longer finding himself before a slow office, but before a rapid, opaque, and difficult-to-challenge decision.

The institutional question, therefore, is not the replacement of man by machine, but the quality of responsibility that survives the mediation of the machine. Human supervision cannot be a decorative formula. It must mean technical competence, the ability to challenge the model, the effective possibility of suspending a procedure, the obligation to justify the exception, and the right to understand the impact of the decision. UNESCO, in its Recommendation on the Ethics of Artificial Intelligence, places human rights, dignity, transparency, equity, and human supervision at the center of AI governance.

European legislation on AI confirms this direction. The European Commission’s AI Office is preparing guidelines for the application of the AI Act, with particular attention to the classification of high-risk systems, transparency obligations, the reporting of serious incidents, fundamental rights impact assessments, and responsibility along the value chain. This is not a technical compliance exercise: it is the attempt to bring computational rationality back within an institutional perimeter of traceability, control, and legitimation.

Here the most delicate connection opens between AI, institutions, and money. The administrative capacity of a state is not neutral with respect to its fiscal capacity, the quality of expenditure, trust in public money, and the credibility of policies. If AI makes it possible to reduce waste, fraud, and inefficiencies, it can indirectly strengthen the material basis of institutional trust. But if it produces decisions perceived as arbitrary, inaccessible, or discriminatory, it can erode precisely that fiduciary capital which makes fiscal obedience, the acceptance of regulation, and the stability of economic expectations possible.

The parallel with the enterprise is direct. Every organization that introduces AI into selection processes, pricing, personnel management, compliance, financial planning, or customer relations is not merely purchasing a tool. It is redefining its own operational theory of value. An algorithm that optimizes cash flows can increase financial clarity, but it can also narrow the strategic horizon to what is immediately measurable. A system that evaluates performance can make previously ignored inefficiencies visible, but it can also transform work into a sequence of indicators incapable of grasping learning, trust, experience, and tacit responsibility.

The risk is not an excess of technique, but the institutional poverty of technique. Technology becomes dangerous when it is left without a political, organizational, and economic grammar. The NIST AI Risk Management Framework proposes risk management articulated through functions of governance, mapping, measurement, and management, across the entire life cycle of systems. The release, in 2026, of a concept note on the profile of trustworthy AI in critical infrastructure confirms that the issue is moving toward the systems on which operational continuity, collective security, and economic stability depend.

Algorithmic rationality therefore introduces a new form of computational power. Not the spectacular power of visible surveillance, but the ordinary power of classification. Deciding which data count, which correlations become operational, which thresholds activate a control, and which profiles generate priority means affecting the concrete distribution of opportunities, costs, and rights. The governance of AI does not concern only protection from technical errors, but the definition of the conditions through which a society recognizes value, need, risk, and merit.

For this reason, post-AI administration cannot be conceived as a simple digital evolution of bureaucracy. It is a new form of institutional infrastructure, in which rule, data, procedure, and model become elements of the same decision-making apparatus. Its legitimacy will depend less on the rhetoric of innovation and more on the capacity to make calculation interrogable, procedure verifiable, outcome contestable, and responsibility recognizable.

The orientation that emerges is not the rejection of the machine, but the rejection of its apparent innocence. AI can make the state more capable, the enterprise more lucid, and governance more sensitive to the complexity of phenomena. But this possibility requires institutions capable of governing computational power as a new form of economic power. The future of administrative rationality will not be decided by the quantity of automated processes, but by the quality of judgment that societies will still be able to place above, within, and against calculation.

Global AI Observatory