Owning AI Infrastructure Keeps Government Intelligence From Crossing Borders

Governments using AI from outside providers can send sensitive information across borders even when their files stay at home. The questions and context they share can reveal policy priorities, spending plans and gaps in public services. As AI agents take on more work, governments need to consider where that information is processed and who controls the infrastructure behind it.

Keertesh Tiwari is a Manager in Government and Public Sector Digital Government Advisory at KPMG India, where he works on digital transformation programs for government agencies and private companies. He has more than a decade of experience in system architecture and application development, including leading development teams for clients in the U.S., U.K. and Japan at Accenture. Earlier in his career, he worked with the United Nations Development Programme to build a mobile app that let citizens in the Indian state of Chhattisgarh submit complaints directly to government departments and track their status.

“It’s very important for governments to invest heavily in infrastructure so they can process data and keep their intelligence within their borders,” Tiwari says. Intelligence here means what a government learns from its data, such as where public services are falling short or which areas are growing fastest. When that analysis runs on servers in another country, the findings travel there too, along with the requests that produced them.

What requests give away

Picture a health ministry asking an AI tool to help plan hospital budgets for three regions. It never uploads a patient file, but the request alone shows where the ministry plans to spend and which regions concern it. Over hundreds of requests, those details show a lot about how a government works. Many agencies don’t see that risk when they sign up.

“Every provider nowadays says you can simply share your data, process it and get intelligence out of it through AI. But they’re not telling governments that the data will be transferred somewhere else to be processed,” Tiwari notes. Government data about citizens is sensitive, and many countries have laws that limit where it can go. Once it reaches a provider abroad, the government that sent it can’t control who sees it or how long it’s kept.

Private companies face the same exposure. A bank that asks an AI tool to review loan applications from one region, or an insurer that asks for help pricing policies in a single city, reveals its plans just as a ministry does. For U.S. agencies, the bigger question is who controls that data once a model provider has it. Katherine Hennessey, Head of U.S. Government Strategy at Everpure, previously served in professional staff roles on the U.S. Senate Committee on Appropriations. “Once an agency’s data is inside a provider’s environment, the contracting and the control questions turn getting it back out into a years-long problem,” Hennessey says.

Processing at home

A government that owns the computers its AI runs on keeps its data under its own control. Larger countries can afford that kind of infrastructure, and many already use AI for core state functions.

Many others can’t. Data centers and the staff to run them cost more than most small national budgets allow. “For small country governments, it’s going to be a bit difficult. For now, they will have to safeguard the data,” Tiwari says. Until they can stand up their own systems, these countries have to be selective about what they send to outside vendors.

Policy is starting to catch up. At India’s AI Impact Summit in New Delhi in February 2026, countries adopted the New Delhi Declaration on AI Impact, which sets out a shared vision for trusted AI rooted in respect for national sovereignty and equitable access to technology. In Washington, the White House budget office updated its rules for how federal agencies use and buy AI, and the purchasing memo urges them to favor American-made AI.

For federal agencies, independence also keeps options open as AI models improve. “An agency running a live mission can’t live with downtime while it switches platforms, so being able to hold its own data and swap the model underneath it becomes a continuity-of-operations requirement,” Hennessey notes.

When AI does the work

In a personal project, Tiwari built a small marketing agency staffed entirely by AI agents, which are programs that carry out tasks on their own. He started with seven agents, each with its own job, such as making cold calls, posting on social media, writing content or reporting results. He has since added agents for finance and risk, and the system now has 24.

“Just by engaging an AI agent, it was helping me build and start performing from day one like a professional. And it wasn’t costing me a lot of money,” Tiwari says. A new employee needs several rounds of interviews and months of training before contributing. Agents can start right away, which saves money for small businesses and government offices.

For governments, that speed comes with a catch. Every agent needs data to do its job, and when agents run on an outside provider’s servers, each task sends some of that data across the border. The quality of that data matters as much as its destination. Agents increasingly handle work that feeds into public decisions, so a flaw in what an agent takes in can carry straight through. “In the private sector a bad automated output costs you a customer or a quarter. In government, a flawed record flowing through an agent can land in a benefits determination or a tactical decision, which is why the verification bar has to be higher than anything industry would accept,” Hennessey says.

Even with clean data, people still need to review what agents produce. Models sometimes invent sources that don’t exist, so a person has to check the work before it affects policy. “Those are the things which only a human can identify and rectify,” Tiwari says.

Governments that process their data at home and have people review the results are in the best position to use AI well. “It is going to help the governments, but at this stage I don’t feel it can replace human decision-making,” Tiwari says.