For a FINMA-supervised institution, buying a RAG platform is not a software purchase — it is an outsourcing arrangement that has to be assessed, documented and supervised. The question a supervisor will ask is not whether the AI is good, but whether the institution retained control, knows where client-identifying data goes, and can keep operating if the provider fails.
Is a RAG deployment “material” outsourcing?
FINMA's outsourcing expectations for banks and insurers turn on materiality: a function is material if a failure would seriously impair business continuity, compliance with regulatory duties, or the institution's risk profile. A pilot in which three analysts query public research is unlikely to be material. A platform indexing client dossiers, credit files or claims records, embedded in advisory or underwriting workflows, generally is — and once it is, the arrangement belongs in the outsourcing inventory with a named responsible owner.
Two properties push a RAG system toward materiality faster than people expect: it concentrates client-identifying data from many silos into one searchable index, and its outputs get used in decisions. Both raise the consequence of failure.
The five things to document
- Data location, per component. Inference, embeddings, vector store, document archive, backups, logs. Not one sentence about “the cloud” — six answers with countries and operators.
- The sub-outsourcing chain. Supervisors care about who your provider relies on. In a RAG stack that means the model provider, the embedding provider, any reranking or OCR service, and the infrastructure operator. Require notification before the chain changes.
- Banking secrecy and professional secrecy. Art. 47 BankA exposure attaches to client-identifying data (CID). If passages from client files travel to a foreign model API inside a prompt, that is disclosure — regardless of where the archive lives. Encryption that the operator cannot bypass materially changes this analysis.
- Audit and inspection rights. The institution, its auditors and FINMA must be able to inspect the service. Confirm this contractually, including for material sub-processors.
- Business continuity and exit. What happens if the provider fails: documented recovery objectives, a data export in a usable format, and an exit plan that does not depend on the provider's goodwill.
Why the Swiss data path simplifies all five
Each item above becomes materially easier when the entire pipeline stays in one jurisdiction. There is no cross-border transfer analysis, no CLOUD Act conflict-of-law memo, no foreign sub-processor to chase for audit rights, and the CID question narrows to a single contractual counterparty under Swiss law. This is the practical reason regulated Swiss institutions gravitate to a Swiss RAG platform rather than negotiating safeguards around a US stack — the compliance work is smaller, not just the risk.
Controls a supervisor will expect to see working
| Expectation | What to demonstrate |
|---|---|
| Data confined to Switzerland | Named operators for inference, embeddings, storage and backups; no US-incorporated service in the path |
| Confidentiality of CID | Per-tenant encryption applied before storage; operator cannot decrypt |
| Need-to-know access | Per-document access control enforced at retrieval, not only in the interface |
| Traceability | Query log and audit log: who asked what, over which documents, when — exportable |
| Output verifiability | Every answer cites the passages it used, so a reviewer can check it |
| Continuity and exit | Recovery objectives, export format, deletion certificate |
Where this lands in practice
SovraRAG is designed against exactly this list: a Swiss-only data path, operator-blind per-tenant encryption, retrieval-layer access control, complete audit and query logs, cited answers, and a dedicated single-tenant deployment option for institutions whose supervisors require infrastructure separation. Sector detail for banking and wealth management and for insurers sets out the workflows this typically supports.
This article is general information about supervisory expectations, not legal or regulatory advice. Assess materiality with your own compliance function.