Choosing a model is an architecture decision, not a vendor decision

Clients often arrive with the model chosen. Sometimes it is the right call; often it reflects which sales team got there first. We ask a different question: what does each step of this workflow actually need?

Reasoning-heavy planning steps, high-volume classification, long-document extraction and latency-sensitive voice turns have very different requirements, and the best model for each is rarely the same one. Treating the model as a component behind a stable interface lets you make those choices per step and revisit them as the providers leapfrog each other, which they do every few months.

The architecture consequence is simple: no prompt or tool definition should assume a specific provider, evals must run against every candidate model, and cost and latency must be measured per step rather than per system. Do that and a model change is an afternoon's work instead of a quarter's.