October 1, 2026, (Inside AI) — Governments worldwide are shifting from using artificial intelligence to assist officials to deploying autonomous agents that execute administrative tasks directly, raising unresolved questions about accountability, oversight, and the limits of delegated authority. The transition from generative AI, which produces text and images, to agentic AI, which plans and acts across multiple steps, marks a structural change in how public services may be delivered.
The shift is no longer theoretical. Singapore released its Model AI Governance Framework for Agentic AI earlier this year. The United Arab Emirates announced in April 2026 that it plans to move 50 percent of its government services and operations to agentic AI within two years. Abu Dhabi also renamed its Ministerial Development Council as the Ministerial Council for Artificial Intelligence and Development, charging it with integrating AI into government work.
These moves signal a departure from AI as a support tool. Agentic systems can gather information, formulate plans, access databases, and execute tasks with varying degrees of autonomy. That capability introduces a new class of risk. The potential blast radius of a single error, meaning the number of people or transactions affected, increases sharply as more administrative stages are delegated to machines.
From Assistance to Administrative Action
Conventional AI has been embedded in governance for years. In healthcare, it enables faster diagnosis and treatment recommendations. In agriculture, India's National Pest Surveillance System uses AI and machine learning to detect pest infestations. The Kisan e-Mitra chatbot, developed by the Ministry of Agriculture and Farmers' Welfare, answers questions about schemes such as PM-KISAN, PMFBY, and KCC in 11 regional languages. It handles more than 8,000 farmer queries daily and had answered over 95 lakh queries as of March 2026.
Read: Google DeepMind Study: AI Agents Cheat and Whistleblow in Swarms
Adding an agentic component to such tools could transform their function. An AI agent could access farmer records, prepare and submit applications, and track their progress. Each step requires a new level of administrative authority. Accessing records involves handling personal data. Submitting applications constitutes an administrative action.
Similar possibilities exist in healthcare, where agents could schedule appointments, follow up on patients, document records, and coordinate between institutions. In broader administration, agents could verify documents, process grievances, and coordinate across departments, reducing routine workloads and waiting periods.
The appeal is clear. AI agents can operate around the clock and serve citizens in multiple languages, a significant advantage in a multilingual country like India. But the risks are equally significant.
First, AI systems are prone to errors and hallucinations. Generative AI can produce incorrect information with confidence. When an agent acts on such information, it can lead to erroneous administrative actions with real consequences for citizens. Singapore's framework explicitly highlights risks from agents taking unauthorised or erroneous actions.
Second, accountability becomes complex. An erroneous action could stem from the underlying model, the data it relies on, or the workflow process. An AI system cannot bear legal or administrative responsibility. Clear lines of responsibility must be established before delegation.
Third, bias and discrimination remain concerns. AI systems can reproduce patterns present in training data. Fourth, privacy and cybersecurity risks grow as agents access multiple databases and applications. Agentic systems are vulnerable to prompt injection and agent hijacking, where malicious instructions from external content influence AI behaviour.
Read: AI Agents Invent Surreal Dialect, Raising Oversight Concerns
Fifth, the digital divide poses a particular challenge in countries like India. If government services increasingly depend on AI-enabled interfaces, citizens with limited digital literacy, connectivity, or access could face new barriers. AI-enabled governance cannot replace accessible human and offline channels altogether.
Finally, the environmental costs of AI infrastructure are substantial. Data centres require significant electricity and, in many cases, clean water for cooling, aggravating resource pressure.
Governments are beginning to institutionalise AI oversight. The UAE has had a minister for AI since 2017. In July 2026, the UK appointed its first dedicated AI minister, Kanishka Narayan. In India, Kerala created a dedicated AI portfolio at the cabinet level, and Tamil Nadu appointed R Kumar as Minister for AI, IT and Digital Services. India's Ministry of Electronics and Information Technology established the AI Governance and Economic Group (AIGEG) to build an institutional framework for AI governance.
The question now is whether AI is simply a technology portfolio or is becoming embedded in the function of a state. The answer will determine which government functions can be delegated to AI, what permissions such systems can have, and where human oversight must remain.
A balanced approach would combine the benefits of agentic AI with well-defined limits on autonomy. Low-risk administrative processes might allow greater AI autonomy. Decisions affecting welfare entitlements, health, employment, and other important domains would require meaningful human oversight.
Accountability requires a clearly identifiable human or institutional authority responsible for AI actions. Government departments can maintain audit trails showing the information accessed, decisions made, and actions undertaken by an agent. Singapore's framework stresses that humans remain ultimately accountable and recommends clearly defined points for human approval in significant decisions.
The principle of least-privilege access can also apply. An AI agent should have access only to the data, permissions, and systems necessary to perform its designated task. This limits the consequences of errors or cyberattacks.
Clearly defined points at which an AI agent must defer to a human official are essential. This is particularly important in decisions involving discretion, conflicting information, or significant impact on an individual's rights or entitlements.
The evolution from generative to agentic AI is not only a change in technology but also in the nature of digital governance. For India, the task is to determine which government functions can be delegated, what permissions systems can have, and where human oversight remains necessary. The effective use of agentic AI will depend on the quality of government data and the safeguards governing AI actions.
