It's true that custom prompt engineering can create friction when switching AI providers, but to say this directly contributes to low switching costs in API deployment feels like we're just throwing good money after bad logic.
If a company pours resources into tailoring workflows for one model, that investment actually raises their sunk cost; it doesn't magically make it cheaper to jump ship.
Would we re-enter this argument if we hadn't already spent time trying to connect these two ideas? Probably not.
For example, if a firm spends months integrating a complex AI solution for customer service, those specific integrations become a barrier to quick shifts, not an enabler.