The state's new operating system

On September 29, US President Donald Trump launched America.gov, an AI-powered front door to the US federal government. Citizens ask questions in plain language and get answers drawn from more than 29,000 official websites, without needing to know which department is responsible.

An executive order signed at the launch directs agencies to connect their public-facing services to it. The administration aims to let citizens apply for benefits such as Social

Security through it from early 2027, and for most passport applications by the end of 2027.

It may prove more consequential than it looks. For three decades, governments have been digitising bureaucracy. Forms became PDFs, queues became portals and files became databases. Yet citizens still had to understand the bureaucracy: know which office to approach, which form to fill and whose jurisdiction it was. AI begins to invert that relationship. Instead of citizens learning the architecture of government, technology is being asked to learn it for them.

Washington is not alone. The United Kingdom rolled out GOV.UK Chat to the public in May, answering questions from roughly 80,000 pages of official guidance. Singapore’s VICA platform runs chatbots for more than 60 agencies. The UAE has U-Ask; Estonia has Burokratt. Governments are converging on one idea: citizens should deal with the state through an intelligent layer, not by navigating its institutional map.

That layer will not stop at answering. As agentic systems mature, AI will act. Today it explains how to renew a passport. Tomorrow it may fill in the application, verify identity, retrieve authorised records, make the payment and track the case. The chatbot is the front door. The intelligence layer sits behind it. Once that layer can authenticate, transact and act across institutions, it starts to resemble an operating system of government.

That raises questions no procurement tender answers. Who decides which sources the system trusts? Who decides what it shows, omits or recommends, and therefore which services citizens discover? Who audits its answers? What legal weight do they carry when they affect someone’s taxes, benefits, immigration status or business? Can a citizen see the source behind an answer and ask for human review?

On launch day, the Associated Press documented America.gov’s answers to politically sensitive questions changing within moments, then declining subjects it had first answered from federal sources. The specific questions matter less than the principle: once AI becomes the gateway to official information, decisions about sources, refusals and corrections become decisions about public administration. Model governance becomes part of governance itself.

The stakes rise once systems act. An agent that explains an application poses different risks from one that submits a form or authorises a payment. Its permissions must be tightly bounded, consequential actions must require the citizen’s confirmation, and every action must be traceable. The interface may become simpler, but the accountability behind it must become stronger.

Accountability also depends on who controls the technology underneath, and this is where the sovereignty debate usually goes wrong. Sovereignty cannot mean every country building its own frontier models, chips and data centres; few states could. But countries also shouldn't embed intelligent systems across critical institutions and later discover they cannot inspect them, protect sensitive data, audit decisions or switch providers. Even America.gov runs on models from two private companies: Google’s Gemini and xAI’s Grok.

The better test is strategic optionality: whether a country keeps real choice within the stack. Can it understand what runs inside its critical institutions? Can it move its data and workloads? Can it replace a provider without rebuilding the state? Can it build domestic capability where national interests demand it?

For Pakistan and the wider Global South, these questions are not abstract. We will live with frontier models whether or not we build them; our banks, universities and children already use them. Using the best models in the world should not mean surrendering control over our data, our identity systems or the rules governing how machines treat our citizens. Whether Pakistan should ‘adopt AI’ is settled. What we build around adoption is not.

Picture it working here. A shopkeeper in Multan asks what taxes apply to her business. A widow in Larkana asks whether she qualifies for BISP. A developer in Peshawar asks what he must register to export software. None of them should need to know which ministry, regulator or provincial authority holds the answer. Each could ask in Urdu, Punjabi, Sindhi, Pashto or Balochi.

A government’s users are its entire population. Translation is not enough: the system must apply local rules correctly and keep meaning intact across languages, and between voice and text. If it cannot, exclusion becomes automated. A system that repeatedly misreads a regional dialect, fails someone with a disability or stumbles over incomplete records will repeat that mistake across thousands of interactions, and strong average performance can conceal persistent failure for an entire community. Testing must therefore ask who gets wrong guidance, who abandons the process and who cannot reach a service.

All of this rests on an easily overlooked prerequisite. Before AI can understand the state, the state has to make itself understandable to machines. Government information must be authoritative, structured and current. Registries must exchange data securely. Eligibility rules must be written to be applied consistently. Identity, payments, permissions and audit trails must work across institutions. Otherwise, AI is merely a polished interface on fragmented bureaucracy.

Pakistan is not starting from zero. Nadra already provides digital identity at national scale, and the State Bank’s Raast provides instant payments. The work now is connecting registries, rules and records to those rails. The first generation of digital government put services online. The next may make the machinery of government invisible to the citizen.

Building PVARA taught us that you cannot bolt new technology onto old administrative processes. When technology changes what an institution can do, its architecture has to change too. At PVARA, our regulatory sandbox tests new products within defined limits, with reporting and safeguards, before licensing. Government AI deserves the same discipline: start with bounded services, such as finding information and completing applications; test them with the communities who will use them; expand only on evidence. As I said recently at Harvard, “The rulebook is a beginning. The public experience is the test”. For the state, that test must include the people most easily overlooked.

The next global divide will not be about access to AI, which will spread. It will be between countries that build the institutions and skills to govern the intelligence layer they depend on, and those left running on systems they cannot inspect, influence or replace.

AI is no longer just a tool that government uses. It is becoming the operating system through which government works. The question is who will control it.

The writer is a minister of state and chairman of the Pakistan Virtual Assets Regulatory Authority (PVARA).

Disclaimer: The viewpoints expressed in this piece are the writer's own and don't necessarily reflect Geo.tv's editorial policy.

Originally published in The News