The sovereignty test — can you swap out a generalist model without losing your 'company veteran' expertise?

Nadella proposes a concrete diagnostic for whether a firm actually owns its AI capability: “A company should be able to switch out a ‘generalist’ model without losing the ‘company veteran’ expertise built into their learning system. This is the key ‘test’ of your control and sovereignty in the era ahead.” If swapping the base model wipes out your accumulated expertise, that expertise was never yours — it lived in the vendor’s weights.

This is the architectural requirement behind avoiding lock-in: it demands Open harnesses with customer-owned databases are the antidote to model-provider lock-in and is the inverse failure mode of how Closed harnesses behind APIs create memory lock-in by design. Passing the test is precisely what lets a firm Routing across the whole model market — and absorbing every migration — is a defense the labs can't copy — routing to the best generalist per task while the company-veteran layer stays put — and it is the structural underpinning of why The system of work is the moat, not the model — the model is fungible underneath.