Leading with Trust: Can Swiss Banking Guide Sovereign AI?
At the Swiss Banking Digital Finance Day 2025, bankers, policy makers, and innovators met to answer a direct question: Can Swiss banking guide sovereign AI? The session, moderated by Paula Reichenberg of Neur-on.ai and joined by experts including Leif-Nissen Lundbæk, CEO & Co-Founder of Noxtua, focused on trust, infrastructure, data, and collaboration.
The takeaway: Swiss banks can lead by setting clear standards for control, compliance, and quality-then executing with precision.
Why trust and sovereignty matter now
Trust sits at the core of Swiss banking. As AI moves closer to core processes, sovereignty becomes a practical requirement, not a slogan.
In financial services, sovereign AI means banks keep meaningful control across the stack-from base model to deployment-while meeting Swiss and EU legal, privacy, and ethical requirements. That control protects client confidence as institutions work with highly sensitive financial and legal data. For reference, see the EU AI Act and Switzerland's Federal Act on Data Protection (FADP).
Apertus: a building block for sovereign AI
The session highlighted Apertus-a trustworthy, neutral foundation LLM that is fully open and transparent. Banks can inspect, adapt, and fine-tune it under their own policies and regulatory constraints.
It supports 1,000+ languages, including Swiss German. It is built to meet obligations under the EU AI Act and Swiss data law, making it a practical starting point for teams that need performance without giving up control. With modular and agent-based approaches, firms can target specific needs such as handling sensitive data or automating regulatory reporting.
The initial 1.0 release sets a sovereignty and compliance baseline. Planned reasoning models and Mixture of Experts architectures are slated to reach the performance levels many banking applications demand.
Infrastructure: the hybrid reality
Most banks operate mixed environments. Sensitive workloads (fraud detection, sanctions screening, compliance checks) often stay on-premises. Less sensitive tasks (voice biometrics, document scanning, routing) can run in the cloud.
The key is crisp boundaries for data flow, with security and compliance as the guardrails. One example discussed in the session combined internal data platforms, external APIs, and cloud models with strict segmentation. This choice of architecture directly affects which use cases are viable, the iteration speed, and cost. In practice, many teams adopt widely available public-cloud models for speed-even if sovereignty is not the top priority for those specific use cases.
What actually drives output quality
Data structuring is the lever that moves results. Yes, base training data and fine-tuning matter. But once deployed, the decisive factor is how the model retrieves and reasons over the bank's own data at inference.
That makes internal data products and well-structured external sources critical. Add to that the model architecture, fine-tuning cadence, inference security settings, and retrieval policies. In finance and legal contexts, traceability-what was found, selected, and quoted-must be visible for compliance and audit. Even "small" implementation choices, like server setup or prompt enrichment, can tilt outcomes.
Collaboration without losing control
There are clear spaces where banks are not direct competitors. In those areas, collaboration speeds progress and reduces duplicated effort.
Federated learning and federated RAG let institutions share insights without sharing raw data. The proposed Hive architecture points to shared platforms where firms can combine their own and third-party data under strict controls. The model is simple: keep sensitive assets sovereign, pool non-confidential expertise, and benefit from collective intelligence-while still using off-the-shelf solutions where they fit.
A practical checklist for finance leaders
- Define sovereignty: Which layers must you control (data, model, inference, deployment)? Document this.
- Segment use cases by sensitivity: On-prem for high-risk; cloud for low-risk. Set data boundaries per segment.
- Pick a base model strategy: Open, inspectable models (e.g., Apertus) for core; hosted APIs for peripheral tasks.
- Build retrieval with citations: Every answer should show sources for audit and review.
- Treat data as a product: Clean schemas, governance, lineage, and access controls. Prioritize high-value datasets.
- Institute policy-as-code: Enforce compliance, privacy, and security at inference and deployment.
- Set evaluation loops: Track accuracy, bias, drift, and incident response. Refresh fine-tunes on a schedule.
- Design for portability: Avoid hard lock-in. Containerize components and keep model artifacts versioned.
- Collaborate where safe: Explore federated learning/RAG and shared expert datasets under clear contracts.
- Upskill teams: Risk, compliance, data, and engineering need a shared playbook and shared vocabulary.
What to watch next
Over the next 12-18 months, expect reasoning-focused models and Mixture of Experts systems to raise the bar on complex financial tasks. Vendor assessments will place more weight on auditability, data locality, and incident handling.
Regulatory expectations will mature around documentation, testing, and monitoring. Banks that pilot sovereign patterns now-alongside targeted use of public-cloud models-will move faster with fewer surprises.
Further learning
If you want a curated view of practical AI tools for finance teams, explore this collection: AI tools for finance. For role-based upskilling paths, see courses by job.
Trust and sovereignty are the edge. With the right architecture and shared standards, Swiss banking can set a higher bar-and deliver safer, faster AI in production.
Your membership also unlocks:
