Every six months a new foundation model resets the cost-per-token frontier. Every twelve months a new provider takes pole position. The teams who shipped an AI roadmap that bakes in one provider are the ones rewriting their stack now.
Three things keep an AI roadmap portable. One — a thin abstraction layer for the model call. Two — a separation between prompt engineering and prompt orchestration. Three — evals that run against any model so you can A/B providers without theatre.
The thin abstraction is cheap to write and saves a quarter of rewrite time when you switch. The orchestration separation matters because what your senior engineers know is how to chain calls and handle failure — what your prompt iterator knows is the model's idiosyncrasies. Keep those two in different files.
Evals are the boring part nobody wants to write until they have to. Write them first. A 200-row gold set with simple scoring beats a fancy eval framework you never run.