Before agentic AI scales in government: the democratic authorization gap

Review

Mohammed Salah, Fadi Abdelfattah, Aisha Al Araimi and Alhamzah Al Sayed Noor identify a prospective failure in agentic public administration: a break in the demonstrable connection between legally and democratically grounded authority and machine-selected action. Their September 2026 Perspective separates accountability after an outcome from authority before execution. An agency can remain answerable, name a responsible official and configure valid technical permissions while still lacking authority to delegate the action performed.

Sections 2 and 3 give that distinction analytical substance. Mandate translation turns broad objectives into choices about records, priorities and interventions. Recursive delegation can transfer consequential discretion beyond permissible boundaries. Action diffusion distributes the authority trail across actors and time. Contestability lag lets administrative consequences accumulate before officials or citizens can challenge them. The authors expressly distinguish routine tool invocation from legal delegation; the relevant boundary is consequential discretion or action attributable to public authority.

Section 4 maps these mechanisms to five pilot conditions: bounded authorization, permission inheritance, action-level traceability, named institutional responsibility, and interruption with reversible redress. Authorization attaches to an action space rather than merely to model procurement. Vendors and technical complexity cannot displace the institution as accountable actor. Procurement becomes the point at which legal competence, technical permissions and remedy duties must be connected before pilot routines harden into infrastructure.

The method is conceptual synthesis, not an evaluation of deployments. Section 1 uses procurement and pilot reports to motivate timing, while Section 5 calls for comparative research and limits general claims about efficiency, fairness or harm. Its evidence supports an anticipatory governance agenda, not the prevalence of authorization failures or the effectiveness of its five conditions. The contribution is a runtime continuity question grounded in delegation scholarship, rather than the discovery that permission differs from legitimacy. It complements the archive's Aegis review, which questions policy authorship behind runtime enforcement, and Agents of Chaos, which examines failures of delegated control.

The proposal leaves a decisive translation problem open. A demonstrable legal chain does not settle disputed statutory interpretation, lawful discretion or democratic legitimacy. Nor does a record of tool use prove that the operational rule correctly expresses the institution's mandate. Permission inheritance prevents expansion only where permissions preserve purpose and meaning across components. The paper does not specify how changes in legal authority invalidate pending actions or reach downstream agents. These are requirements for implementing its conditions, not demonstrated defects in an existing system.

A fair response is that a Perspective should define research conditions rather than supply a complete enforcement architecture. That resolves any demand for production software. It leaves the adequacy of the conditions untested: agencies must still identify who approves mandate translation, what evidence a reviewing forum accepts, and how quickly suspension takes effect. Section 4 already qualifies restoration by feasibility and allows other legal remedies. The remaining question is whether a remedy can arrive before a deadline, disclosure or enforcement step produces an effect that record correction cannot undo.

My governance judgment is that this framework should inform pilot admission, while its sufficiency remains a research question. Otherwise permission designers and vendors may acquire practical discretion over the meaning of public authority. Comparative pilots should test changed mandates, revoked delegations, contested interpretations and delayed complaints, recording prevention, reconstruction and remedy times. Independent review showing that these controls preserve lawful discretion and deliver timely remedies would justify a stronger operational claim. Until then, the paper establishes a useful condition on experimentation: public institutions must make the authority behind machine-selected action examinable throughout execution.

Key Insight

A traceable permission chain can document how an agent acted without establishing that the public institution had legitimate authority to delegate the action.