Diagnostic imaging is a demanding test of the claim that sovereign AI is practical rather than merely principled. Mammography data is about as sensitive as medical records get, the clinical tolerance for latency is low, and the regulatory surface spans federal data-protection law, cantonal health-data rules and professional secrecy obligations all at once. If sovereign deployment worked only in undemanding settings, this is where it would fail.
It is not failing. b-rayZ, a University Hospital Zurich spinoff, has deployed its real-time image-quality assessment and breast-density reporting platform across multiple Swiss clinical sites, including centres in Zurich and Fribourg. The reported operational results are substantial: over 25 per cent time saved in the reporting workflow, and roughly half the previous cost of quality assurance. The company raised CHF 4 million to scale the deployment further.
The architecture is the lesson
The clinical performance is the headline, but the deployment shape is the part other institutions can copy. Patient imaging is processed within the hospital’s own infrastructure. The model runs where the data already sits, inside the clinical network perimeter. There is no cross-border transfer, no external API call in the diagnostic path, and consequently no question to answer about whose jurisdiction the images passed through.
This is a meaningfully different pattern from the one most enterprises encounter when they evaluate AI tooling. The common offer is a capable model behind a vendor’s API, with residency addressed afterwards through contractual assurance - a data-processing agreement, a regional endpoint, a commitment about where the servers are. Those instruments help, but they are promises about infrastructure the customer does not control.
b-rayZ inverts that. The residency guarantee is not a clause; it is a property of where the software runs. Nothing needs to be promised about foreign lawful-access requests because there is no foreign processing to reach.
Sovereignty as a product decision
What makes this case instructive for Swiss deep-tech firms generally is that sovereignty here was a design decision taken by the vendor, not a hosting preference imposed by the customer. The product was built to run inside the client’s boundary from the start.
That distinction matters when procuring AI capability. A tool designed for cloud delivery and later fitted with an on-premises option tends to carry the assumptions of its origin: telemetry that phones home, licence checks that require connectivity, model updates that arrive through a channel nobody in the institution controls. A tool designed for in-boundary operation from the outset does not accumulate those dependencies.
The economic result is worth noting too, because it undercuts the assumption that residency is something an institution pays extra for. Halving quality-assurance costs is not a compliance expense - it is an operational return that happens to arrive in a sovereign package. The residency guarantee came free with a workflow improvement that would have justified itself regardless.
The transferable pattern
For Swiss MedTech firms, and more broadly for any institution holding data it genuinely cannot export, the pattern generalises cleanly. Clinical-grade accuracy and absolute data residency are complementary requirements rather than competing ones. The trade-off that dominated procurement conversations two years ago - capability from the cloud, or control at home - is not the trade-off currently on offer.
The remaining question for most institutions is not whether in-boundary AI can meet the clinical or engineering standard. That has been demonstrated. It is whether the infrastructure to run it is available under the same jurisdiction as the data, on terms that do not require building a data centre to find out.