The notion of minimal infrastructure requirements leading to low switching costs for API deployment is a classic failure mode in planning; it only holds until the first actual integration happens.
Developers might think they can quickly redirect API calls, but once specific custom prompt engineering is baked into workflows, that illusion of ease can snap.
It's like thinking you can easily swap out a custom-built kitchen cupboard for another when you've already tailored all your plumbing and wiring around it; the initial small effort snowballs fast.
Those so-called minimal requirements quickly become a blast radius of dependencies, making rapid reallocation a pipe dream when you factor in actual operational friction.
Try telling a team that's spent months fine-tuning an AI to understand their niche data that they can just 'switch' when the whole system is built around its unique quirks.