Leading with Trust: Can Swiss Banking Shape Sovereign AI? 

Paula Reichenbarg and Leif-Nissen Lundbæk, CEO & Co-Founder of Noxtua, at Digital Finance Day

In this interview with Paula Reichenberg, Andrea Luca Aerni, Policy Advisor Digital Finance at Swiss Banking, distills the session’s key themes and insights.

Andrea: Paula, the session opened with the question of trust and sovereignty in AI. What makes these concepts so crucial for Swiss banking today? 

Paula: Trust is the foundation of Swiss banking, and as we move into the AI era, sovereignty becomes just as important. In my view, sovereign AI in the context of financial services means that Swiss banks should be able to exert a good amount of autonomous control over the entire technology stack, from the base model to deployment, ensuring compliance with relevant Swiss and EU laws, data protection, and ethical standards. This is not just about technology but about maintaining the trust that our clients place in us, especially as we handle sensitive financial and legal data.

The session highlighted the Apertus model as a potential building block for sovereign AI. What sets Apertus apart, and why is it relevant for Swiss banks? 

Apertus is a trustworthy, neutral foundation Large Language Model that’s fully open and transparent. Banks can inspect, adapt, and fine-tune it under their own regulatory standards and requirements. It’s multilingual, supporting over 1,000 languages, including Swiss German. Most importantly, it’s compliant with the EU AI Act and Swiss data laws, making it a strong candidate for banks seeking both performance and sovereignty. The ability to customize and use modular, agentic approaches means banks can address specific needs, like sensitive data handling or regulatory reporting, with confidence. The initial 1.0 release establishes the sovereignty and compliance framework, while upcoming reasoning models and Mixture of Experts architectures will deliver the performance levels that banking applications demand. This roadmap means banks can plan their sovereign AI strategies around capabilities specifically designed for sensitive data handling and regulatory reporting.

Infrastructure was a major topic, especially the hybrid strategies Swiss banks use. What are the main considerations when designing AI infrastructure? 

According to panelists and participants at the session, banks face a complex IT-landscape. Many use a hybrid approach - combining on-premises solutions for sensitive tasks (like fraud detection or compliance) with cloud-based models for less sensitive functions (like voice biometrics or document scanning). The key is to ensure data flows securely, with clear boundaries between internal and external sources. The session’s infrastructure diagram illustrated one bank’s approach to integrating internal data, external APIs, and cloud AI models, with security and compliance as governing principles. The choice of infrastructure directly impacts what use cases are feasible and how quickly banks can innovate. This also means that banks often need to follow a pragmatic approach in which widely available models on public-cloud infrastructure are being used and sovereignty considerations are not always top priority.

The session slides listed dozens of factors that influence AI output quality - from data selection to prompt engineering. Which of these do you see as most critical for banks? 

Data structuring stands out as the most critical factor, though it operates differently at different stages. The quality of base training data and fine-tuning datasets shapes the AI's general understanding of banking-specific queries. But for deployed systems, the decisive factor is how the model interacts with a bank's own data during inference. How banks structure their internal datasets – and which structured external sources they can access – determines real-world output quality. Among the other factors we examined, model architecture, fine-tuning frequency, and security settings around inference all matter. In finance and legal contexts, the ability to track which data was found, selected, and quoted by the model is also critical: this transparency supports both compliance and auditability. Ultimately, the session made clear that every detail - from server architecture to prompt enrichment - can influence the trustworthiness and utility of AI outputs in banking.

Looking ahead, how can Swiss banks potentially collaborate to accelerate sovereign AI adoption and avoid duplicating efforts? 

The session showed that in certain specific areas where banks are not necessarily competitors, collaboration is the way forward. Thus, in the session we’ve also discussed federated learning and federated RAG (retrieval-augmented generation), which allows banks to securely share insights and to have access to structured, non-confidential expert datasets without compromising sovereignty and control. Projects like the proposed Hive architecture aim to create shared platforms where banks can interact with both their own and third-party data, benefiting from collective intelligence while maintaining strict controls. By working together, Swiss banks can set new standards for trustworthy, sovereign AI in addition to using run of the mill solutions that are readily available. 

Thank you, Paula, for the insights and see you at next year’s digital finance day!