At Rise of AI 2025 in Berlin, Georg Schlenkhoff (Mistral AI) and Oliver Stuke (Capgemini Invent) argue that Europe, despite a widespread narrative of having fallen behind in AI, is building real momentum toward sovereign generative AI, and show how that stack can be applied to unblock public-sector bureaucracy. Schlenkhoff frames sovereignty as owning infrastructure, technology and data, citing figures that only 3 percent of European R&D spending goes to technology and software and that Europe consumes 20 percent of global chips while producing only 9 percent. Stuke points to recent political momentum, including a German conference of minister presidents' sovereignty declaration, a Dutch parliamentary declaration and the EU's 200-billion-euro continent AI action plan, and presents a layered sovereign stack built with partners including StackIT, Aleph Alpha, Mistral, Fraunhofer, Demos and Atlas Systems. They walk through a concrete application in Germany's Ministry of Interior, where an AI system checks planning applications for wind turbines against legal requirements and drafts notices to applicants, while stressing that German administrative law bars fully autonomous decisions and requires a human in the loop. Audience questions cover adversarial attacks, change management and ethical data handling.
Key takeaways
- Schlenkhoff cites figures that Europe captures only 3 percent of global R&D spending in tech and software, and produces only 9 percent of the chips it consumes.
- The EU's continent AI action plan is funded with 200 billion euros, covering chip gigafactories, talent training, and public- and private-sector adoption.
- Their sovereign AI stack combines StackIT (compute), Aleph Alpha (explainability), Mistral (frontier models), Fraunhofer (compliance research) and Demos (citizen participation).
- German administrative law (article 34b) forbids fully autonomous administrative decisions, so their planning-permit AI system keeps a human clerk in the loop.
- Schlenkhoff projects a shortage of 1 million skilled public-sector workers by 2030 as a key driver for applying generative AI to reduce mundane workload.
Detailed notes
Is Europe really behind in AI?
Schlenkhoff and Stuke open by challenging the widespread thesis that Europe has lost its track in AI. Stuke argues the digital realm has only recently become a real field for European geopolitics, after years of governments focusing on regulation rather than treating technology like other strategic resources such as steel or coal. Schlenkhoff defines sovereignty as owning infrastructure, technology and data, which matters because governments handle sensitive citizen and corporate data, and cites figures that only 3 percent of European R&D spending in technology and internet software stays in Europe, while Europe consumes 20 percent of global chips but produces only 9 percent.
Political momentum toward a sovereign stack
Stuke lists recent political signals: the German conference of minister presidents adopted a declaration calling for sovereign AI and less dependence on US technology providers, the Dutch parliament passed a comparable declaration asking for a sovereign stack, and the EU declared a continent-wide AI action plan funded with 200 billion euros covering gigafactories with 100,000 chips, talent training and public- and private-sector adoption. Schlenkhoff adds that the new German coalition agreement similarly calls for European, trustworthy digital providers and reducing digital dependencies, summarized as building a "German stack" that must remain interoperable across the EU.
Building a sovereign AI stack from real components
Drawing on the Euro Stack initiative, Stuke describes a layered stack running from raw materials and chip production (citing ASML) up through AI operating systems and models to the application layer. For their own solution, the pair assembled StackIT for European GPU compute, Aleph Alpha for explainability, Mistral for frontier models, Fraunhofer for trusted-AI and compliance research (operationalized by Capgemini Invent), Demos for AI-augmented citizen participation, and Atlas Systems for end-to-end confidentiality and security as an outer layer on the StackIT infrastructure. They present this combination as a first working example of a sovereign AI stack rather than a theoretical concept.
Why generative AI, not replacing workers
Schlenkhoff identifies three drivers for applying generative AI in government: a projected shortage of 1 million skilled public-sector workers by 2030 as staff retire while workload grows; approval procedures that studies show run six months longer than legally required; and a 500-billion-euro government funding program whose bottleneck, he argues, is not money but slow planning processes. He frames generative AI as freeing staff from repetitive tasks to focus on stakeholder relationships, not as a replacement for them, citing legal, data-bias and interpretation challenges including the EU AI Act's high-risk classification rules and GDPR, plus the need to keep humans in the loop and decisions transparent and explainable.
Case study: automated planning-permit checks
The pair demonstrate a real deployment with Germany's Ministry of Interior, where a company applying to build a wind turbine submits an application that the AI system checks against the relevant environmental law (BImSchG paragraph 10), extracting a legal checklist and mapping it against the submitted document and metadata, including blacking out personal data. The system produces a to-do list for both government workers and applicants and can draft a notice, for example confirming a site plan and noise forecast are acceptable while flagging that an EIA summary and public-insurance proof are still missing, with 21 days to submit them.
Q&A highlights
Asked about adversarial attacks using fabricated but fully compliant applications, Stuke says the goal is to mirror real-world verification processes into the digital system, and notes German administrative law (article 34b) explicitly forbids full automation of these decisions, so a human clerk remains in the loop. On citizen-participation processes being flooded with AI-generated submissions, Schlenkhoff points to Demos's tooling for detecting and clustering AI-generated input. On change management in inflexible public organizations, Stuke cites an eight-year German Autobahn digitization contract where change management was built into the tender from the start, and both stress involving end users early. On ethics and data protection, Schlenkhoff argues sovereignty in deployment, running open-source models within secured, self-contained infrastructure that does not communicate externally, is how the approach keeps sensitive data from leaving secured premises.
Quotes
Sovereignty from a nation's perspective means that you are master of your own destiny.
Georg Schlenkhoff
Any technical solution cannot stand on their own feet.
Oliver Stuke
We embrace it and we say it's good because we really regulate the main parts.
Georg Schlenkhoff
Questions answered in this talk
Is Europe behind in the AI race?
According to Georg Schlenkhoff and Oliver Stuke at Rise of AI 2025, despite a common narrative that Europe has fallen behind, there is significant political and technical momentum toward sovereign European AI, including a 200-billion-euro EU continent action plan and national declarations in Germany and the Netherlands.
Can generative AI replace public-sector workers?
Schlenkhoff and Stuke state at this talk that generative AI is meant to remove repetitive, mundane tasks from public servants, not replace them; they note German administrative law (article 34b) explicitly forbids fully autonomous administrative decisions, so a human remains in the loop.
How can AI speed up government approval processes like building permits?
Stuke and Schlenkhoff describe a deployed system for Germany's Ministry of Interior that checks planning applications, such as for wind turbines, against legal requirements like BImSchG paragraph 10, generates a compliance checklist, and drafts notices to applicants about missing documents.