Every engagement opens the same way: we ask which job is being done badly, by whom, and how anyone would know it got better. The model comes later — usually much later.
The temptation is to start from capability. A team reads about a new context window and works backwards to a use case. That path produces demos, and demos are where most AI projects stop.
So week one is unglamorous. We sit with the people doing the work, count the steps, and find the point where the process actually leaks time. Only then does the question of which model, at what cost per call, become answerable.