Agentic AI in Government Hits 'Hard Part': Deciding What a Machine May Decide

The modern administrative state runs on paperwork, process, and precedent. For decades, governments have attempted to digitize this machinery, replacing filing cabinets with databases and manual data entry with basic software scripts. These early efforts yielded modest efficiency gains, but they largely preserved the fundamental structure of bureaucratic work: humans made the decisions, and machines recorded them.

But a new threshold is being crossed. Artificial intelligence is shifting from a passive tool that retrieves information to an active collaborator that executes complex tasks. This is the era of "agentic AI"—systems capable of receiving a high-level goal, breaking it down into sequential steps, and autonomously navigating the workflow to completion without human intervention at every juncture.

The appeal for overburdened public sectors is immense. Across the globe, government agencies face rising caseloads, stagnant budgets, and a mandate to deliver faster, more personalized services. According to a landmark study by Stanford University and the Administrative Conference of the United States, nearly half of all federal administrative agencies are already experimenting with AI tools to manage this load.[1]

Yet, as these autonomous systems move from pilot programs to core infrastructure, they are colliding with a fundamental pillar of democratic governance: the law. The hardest part of deploying agentic AI in government is not writing the code, but deciding what a machine is legally permitted to decide.

To understand the friction, one must first understand the mechanism of agentic AI. Traditional government automation relies on Robotic Process Automation (RPA). RPA is deterministic; it follows rigid, pre-programmed rules. If a citizen submits a form with a missing field, the RPA script breaks or flags an error. It cannot reason its way around an exception.[4]

Agentic AI operates differently. Powered by large language models and contextual reasoning, an AI agent can understand intent. If tasked with analyzing public comments on a proposed regulation, it does not merely search for keywords. It can read thousands of submissions, identify substantive arguments, group identical concerns regardless of phrasing, and draft a synthesized summary for policymakers.[3]

In a procurement context, an agentic system can autonomously monitor vendor databases, cross-reference compliance records, flag potential legal vulnerabilities, and route the finalized dossier to the appropriate department. The AI manages the sequential dependencies and parallel workflows that characterize bureaucratic administration.

This capability represents a paradigm shift in efficiency, but it triggers a profound legal bottleneck. Administrative law is built on the principle of accountability. When a legislature grants an agency the power to regulate, distribute benefits, or enforce rules, it expects that power to be exercised by accountable human officials.

This is known as the non-delegation doctrine. While an agency can delegate routine administrative tasks, it cannot unlawfully delegate its discretionary power—its judgment—to an unauthorized third party. As AI agents become more sophisticated, the line between "processing information" and "exercising judgment" begins to blur.

If an AI agent reviews a disability claim, synthesizes the medical records, and drafts a recommendation for denial based on its interpretation of agency guidelines, who is actually making the decision? If a human official merely clicks "approve" on the AI's drafted response, courts may view this as a de facto delegation of administrative power to a machine.

Legal scholars and digital constitutionalists warn that this dynamic forces a structural transformation in public service. The state is shifting from a model of active human decision-making to one of human "overseeing."[2]

Under this emerging framework—codified in regulations like the European Union's AI Act, which mandates natural person oversight for high-risk public sector AI—the machine does the heavy lifting. The human bureaucrat is placed "on the loop," tasked with monitoring the system's functioning and intervening when necessary.[2]

The goal is to preserve legal accountability. The AI can autonomously manage workflow routing, document tracking, and content generation, but human experts must retain the final authority for substantive policy questions and regulatory interpretations. The legal validity of the decision rests entirely on the human's signature.[5]

However, this resolution introduces its own set of risks, chief among them being "automation bias." This is the well-documented psychological tendency for humans to defer to the outputs of automated systems, particularly when those systems are highly complex and the human is under time pressure.[2]

If a civil servant is required to oversee hundreds of AI-generated decisions a day, the oversight can easily become a rubber stamp. Meaningful human oversight requires the official to fully understand the analytic process through which the AI arrived at its conclusion—a daunting task when dealing with advanced neural networks that operate as "black boxes."

Furthermore, the shift to an overseeing state alters the epistemic relationship between the government and the citizen. When automated systems handle routine cases, public servants become insulated from the day-to-day realities of the people they serve, stepping in only to manage exceptions and edge cases.

The challenge for the next decade of administrative law will be closing the gap between technical capability and democratic control. Governments must design hybrid workflows that play to the comparative advantages of both humans and machines.

AI excels at pattern recognition, scale, and routine analysis. Humans remain essential for empathy, navigating ambiguity, and bearing the moral and legal weight of state power. As agentic AI continues to evolve, the true test of government modernization will be ensuring that the machinery of the state remains tethered to human accountability.