What Public AI can contribute to the future of AI-enabled public services

When regional administrations in the European Union start adopting AI for public services, many decisions feel local – which tool, which vendor, or do budget and talent allow for it? But the options on the table are also shaped by a market structure that is anything but local. For example, nearly every frontier AI model available is American or Chinese, with only two European models competing in the top 30. The compute is concentrated. The APIs carry pricing set elsewhere. So while Europe has built the most sophisticated regulatory framework for AI in the world, it also wants to have credible alternatives across the entire “AI stack”. What can this framework offer practitioners and decision-makers for thinking about how resilience in public services could look like?

This article is based on discussions with Felix Sieker, Project Manager at the Bertelsmann Stiftung; Piret Hirv, Head of the Data Management Competence Centre at e-Governance Academy; and Federico Plantera, Researcher in Tech and AI. Sieker is one of the authors of a White Paper on the concept of Public AI, developed by the Bertelsmann Stiftung and the Open Future Foundation, and translated into a policy brief presented at the European Parliament in January 2026. 

Written by Federico Plantera.

Concentration runs deeper than models

AI is a layered system: compute at the bottom, data on top of that. Then models, applications and software. The dominance of non-European firms runs through every layer. Sieker stresses the point: “Concentration is not only about models, but also about the underlying stack – the quality and degree of control over it. When we think of AI, we often see only the models, but actually, AI is a layered stack, and this concentration of power is mostly dominated by non-European firms.”

For public administrations, this has practical implications that are already familiar from other domains. Most leading frontier models operate through API pricing, and Sieker draws a parallel many European officials will recognise: “There’s a big debate on the reliance of many European administrations on Microsoft Office, because API prices are quite high. And if we only rely on these frontier models with API prices that are okay so far, but can increase in the future, we will definitely have a strong reliance.”

“We’ve already seen it with digital services. Once you have these lock-ins, it’s really hard to get out,” Sieker and Hirv agree.

What makes the current moment significant is that the lock-in pattern with AI has not yet fully set in – but the window for making different choices is narrowing with each procurement cycle.

The open-source landscape complicates matters further. Sieker notes that “the only leading open models European administrations can rely on are Chinese – and that’s also not necessarily an ideal situation.” Mistral remains the only European company in the frontier race, with limited uptake across European public sectors.

Sovereignty in practice

Having more options, more control over one’s fate in the stack, is what sovereignty means in practice. But Sieker is frank about the limits of the sovereignty discourse: “Sometimes it appears that sovereignty would mean having entire control of the tech stack, and I think this is not possible. It’s also not desirable. If you think about semiconductors, the supply chain is so diversified – it doesn’t make any sense to control the entire supply chain. So sovereignty should rather mean resilience, to have more of a choice.”

The White Paper proposes a practical tool for making those choices: the Gradient of Publicness, a continuum across the stack from commercial to fully public. It has a dual function: “It’s a diagnostic tool – a government can map all its AI interventions and get an overview of where the dependency on private provision is highest. And it’s strategic – once you’ve mapped them, you can make choices. For certain areas like defence or healthcare, you need public compute, because you don’t want your data hosted in another country and then sent back again,” Sieker says.

Where real alternatives exist, they remain underexplored.

– Felix Sieker

In practice, this involves trade-offs. “Many tools, from a sovereignty perspective, are more desirable and they cause less lock-in, but they are also often a bit less practical.” The relevant question is not whether to go fully public, but where along the gradient each specific intervention should sit. Where real alternatives exist, they remain underexplored. Spain’s ALIA – an open language model developed by the Barcelona Supercomputing Center – and the EU AI Factories are already operational. AI Factories are designed as “a one-stop shop for businesses, for startups, but also for public administrations,” Sieker notes. Most governments have not yet tested them.

The data mess and the AI dream

The sharpest frictions seem to come not from AI tools, but from what sits beneath them. Piret Hirv, framing the tension from eGA’s experience: “We often work with very hands-on topics – data infrastructure on municipality level, on government level – and often we can see that it’s a mess. Should we even run after fancy AI dreams without cleaning the mess first?”

In the TSI EU-supported project that framed this conversation, every regional administration confronted the same preconditions: fragmented registries, no centralised data architecture, inconsistent interoperability. And Hirv identifies a pattern from both sides of the service counter: “Companies are complaining about too complex services, that everything takes too long. And then I can see from the other side that government officials are saying, ‘But we will have this AI solution to fix all this.’ These building blocks, the data and infrastructure, are really not sufficiently in place.”

Without dismissing the concern, however, we should also resist the conclusion that governments can only wait. “If we wait until our data mess is solved, that takes a really long time. That should not be an excuse for not engaging more with AI. Maybe there’s more pressure needed to solve some of these issues,” notes Hirv. But there is a risk that needs to be named.

We can’t just have AI – it won’t make the process better. We have to work on these processes.” – Felix Sieker

Germany’s experience with digitisation offers a cautionary pattern, as Sieker explains: “What was often done is to digitise these processes. So then you have a digital process, but the process is still the same, still very bad. We can’t just have AI – it won’t make the process better. We have to work on these processes.” Without process reform, generative AI risks adding another layer to this: “We would just generate more text. The applicant will use a generative AI to generate the text, and the public official will use another generative AI to analyse the text. So we would only interact with AI, sure, but the process is still the same.”

Playing together

The idea should not be that, since there’s generative AI, we should deploy it for everything. “It kind of makes sense that public administrations are a bit slower in adopting. But public administrations play a really big role because they make many important procurement decisions. Not everything has to be fully public. But at least consider what kind of public solutions are out there too that could help, and always think about the trade-off between something that’s fully public, but maybe less convenient, and something that is fully private, but can create more dependencies,” says Sieker.

Not everything has to be fully public. But at least consider what kind of public solutions are out there too that could help, and always think about the trade-off between something that’s fully public, but maybe less convenient, and something that is fully private, but can create more dependencies,” – Felix Sieker.

India’s digital public infrastructure (DPI) stack comes up as a reference model: “There, the infrastructure is basically a set of open APIs and digital public goods, mostly around identity, payment, and data sharing, and it’s designed so that both public and private services can plug into it,” Sieker explains. So, rather than casting the private sector as an enemy, the approach creates the conditions for inclusion, playing together – and bring about digitalisation in both the economy and society.

For regional authorities, these are procurement decisions being made now, governance arrangements being designed now. Public AI offers a framework for asking where the dependencies are forming, where alternatives already exist, and where the trade-offs between convenience and control will need to be made deliberately rather than by default.

This article was produced within the project “EU Commission Project (24ES06/24DE33): Supporting regional entrepreneurship through the adoption of innovative technologies, including AI, in public services” with the financial assistance of the European Union via the Technical Support Instrument and implemented by the e-Governance Academy, in cooperation with the European Commission.