Enterprise AI procurement drifts towards lock-in. Standardise on one provider’s API and you integrate its token format, tool-calling conventions and latency profile into every workflow; switching costs compound with each deployment. For regulated firms, that dependency is not only commercial — under operational-resilience rules such as DORA, single-provider concentration is a supervised risk.
AXA, Europe’s largest insurer, has taken a deliberate counter-position. The group has disclosed more than sixty agentic-AI use cases in testing or partial production across contact centres, underwriting and claims, alongside an in-house governed environment that cut the time to assimilate underwriting guidance from around ten minutes to under three per query. The strategic decision, though, is architectural: AXA is pursuing multi-vendor, multi-model deployment with the explicit aim of staying as independent as possible in its AI supplier relationships.
The lesson is not which model to choose
For FINMA-regulated insurers and reinsurers, the takeaway is not a model recommendation but a design principle: architect for choice. Sovereign infrastructure that can run multiple open-weight models removes single-vendor dependency at the hardware layer. The institution switches models without switching infrastructure, and the supervisor sees a resilient, auditable supply chain with no single point of foreign-jurisdiction failure.
Vendor independence is not achieved by negotiating a better contract with one supplier. It is achieved by owning the layer underneath them.