Digital sovereignty
Control is an eligibility gate. What three governments show about AI in public services
When a government puts AI into a public service, control comes first. What the model can do comes second.
That was the opening point of a session on AI for tax, jobs, and healthcare at NVIDIA GTC in Berlin. Eva-Maria Hempe, who leads NVIDIA's public sector strategy in Europe, called control an eligibility gate. A model has to pass that gate before anyone evaluates what it can do.
In this blog, we look at what that gate consists of, how three governments passed it by three different routes, and what this asks of a client who wants AI in production.
What an eligibility gate is
An eligibility gate is a requirement that is checked first. A solution that fails it is out, however strong it is on every other point.
Hempe named three things a state needs to have in place before a model is even considered.
- the data stays on the state's own infrastructure
- the organization can prove what ran, where, and under which conditions
- the model fits the national language, the national law, and the national processes
A model that can only be reached through an API, a connection to a service that runs somewhere else, is already out at that stage.
Three routes through the gate
What does a government do when the model it wants cannot pass the gate as a remote service?
According to Hempe, there are three routes.
- partnering
- building
- buying
The session showed one example of each, from three countries.
1. German customs partners with an integrator
Partnering means building together with an integrator, an IT service provider that develops the solution with the government. The result can then be reused across agencies.
Thomas Feld of Materna works with the German tax and customs authorities and with the IT service provider of the German federal government. His message was that for these authorities sovereignty is a requirement. They work with highly sensitive data, every step is bound by strict and complex rules, and the authority has to stay in control of the process.
Within that requirement, a great deal is already running.
- consumers complete the customs declaration for a small parcel in one app, with AI that helps find the right tariff
- AI reads the excise forms that small companies send in on paper or as PDF
- customs law is turned into machine-readable rules that an AI agent, a program that carries out tasks independently, can follow
- new applications are built with an AI agent that works in a fully closed environment, without access to public AI, so that data and code stay inside
Feld sees the role of people shifting from actor to supervisor. He drew three conclusions. AI is already in use in tax and customs today. Sovereign deployment already works today. And the task now is to build on the foundations that exist and to bring AI into practice faster.
2. France Travail builds and runs it itself
Building means that the agency's own engineers develop the service and run it at scale.
Maxime Fournioux works on the AI platform team of France Travail, the French public service that supports job seekers and administers unemployment benefits. He presented two AI services that France Travail has in production, in its own data center. His central lesson was that the challenge is to scale a complete service. The model is only one part of it.
The first service transcribes the interviews that job seekers have at the agencies. The projection is three million hours of audio a year. The service runs only in the morning, which leaves three hours for three thousand hours of audio. One small detail turned out to matter a great deal. The audio has to travel in a compact file format, because otherwise the internet connection of the whole agency stops working.
The second service reads scanned payslips, almost three million documents in less than six months. Moving it to faster hardware brought no improvement at first, because the code had yet to be optimized for that hardware. Once that was done, the service reached its target of ten requests per second.
In Fournioux's department, a stress test, a test that simulates the full production load, is compulsory before a service goes live. Because both services use sensitive data, the law required all processing to take place in France Travail's own data center. Every application is also assessed against the organization's ethical charter.
3. Poland buys a closed model and runs it in its own data center
Buying means purchasing a closed model, a model whose maker keeps the inner workings to itself, and running it on the institution's own hardware. Hempe called that the interesting one. How does a model with closed inner workings run on an organization's own infrastructure?
Jakub Abramczyk of ElevenLabs described a deadlock. A government wants the best available model for its most sensitive data and has to keep control of that data. Building such a model itself takes years of research and a great deal of money. For the maker of a closed model, in turn, the model is the business. If it leaks, it is open source the next day, and a model can only leak once.
The way out is confidential computing. The data is encrypted in storage, in transit, and also while it is being processed. The model runs in a protected environment on hardware in the government's own data center. The government runs the model on its own soil, and the model itself stays sealed. Before the key is released, the system verifies that the hardware really offers this protection and that the software is exactly the software that was delivered. If one of those steps fails, there is no key and nothing runs.
The Polish Ministry of Health has had this in production since 1 October. A natural Polish voice calls patients, older people in particular, to remind them of their appointment. According to Abramczyk, Polish healthcare has forty million appointments a year, and ten to twenty percent of them are missed. Every missed slot means someone else waits longer.
The price of this route was time. A lot of it went into explaining how confidential computing works, to clients, governments, security teams, and the company's own legal department. ElevenLabs also spent a few months trying to break its own setup. Known attacks on confidential computing exist, and according to Abramczyk they are out of reach in a government data center.
What helps a government that is stuck at the pilot stage
Hempe closed with one question to all three speakers. What would they tell a government that is stuck at the pilot stage?
Feld's advice was to start small and to start real. Choose an application with real users behind it, think from the first day about how it will be operated, and involve IT.
Fournioux advised designing for production during the exploration. Choose a model that can carry a production workload, and have someone on the team who already thinks about the final architecture and the stress test. At France Travail, he saw many applications stop at the production stage because that thinking was missing in the team.
Abramczyk pointed to the hardware. Governments operate at scale, and whoever wants to run a service next year should, in his words, have ordered the hardware two years ago.
Hempe added that doing this at the last minute always costs more. Her advice was to forecast the demand and to plan for it. Part of the cost of AI sits in hardware and applications. A large part sits in governance and in planning from the beginning.
What this asks of a client
Hempe ended with what the three stories have in common. Three very different institutions had very similar requirements and arrived at three different solutions. Which route is right depends on the organization itself.
In an earlier blog, we wrote that the client is responsible for the conditions that allow a project to succeed. The eligibility gate is such a condition. For a client, in our view, that comes down to four things.
- Write the gate down before the evaluation starts. Where does the data have to stay, what does the organization need to be able to prove, and which language, law, and processes does the model have to fit. A model that can only be reached through an API then drops out by itself.
- Choose the route deliberately. Partnering, building, and buying are three complete routes. Which one fits follows from the organization and the way it is governed. The choice of supplier comes after that.
- Decide on the first day who will operate the service. A pilot without an operator, a stress test, and a production architecture remains a pilot.
- Put the computing power in the plan. Hardware has lead times and usage has to be forecast. Both belong in the planning and in the budget, next to the application itself.
What we want this blog to show
With this blog, we want to show that control over AI in a public service is settled at the start, on the client side. Three governments in three countries set that requirement first, and each found a route that fits its own organization.
At Quilyx, we set the same requirement for our own systems. Our AI runs on our own equipment in the Netherlands, with local models. Only our own AI handles sensitive client and supplier data. And people keep the decision.
An organization that writes its gate down first chooses its own route.
About this article: This article is our own reading of the session Modernizing the State at NVIDIA GTC Berlin on 22 October 2026. The ideas from that session belong to the speakers and are credited to them. The translation to clients, the positions and the choices in this article are ours. We wrote and edited it together with AI, which helped with the research, the working out and the wording. Quilyx carries the editorial responsibility.