An AI product can use a model to interpret a request without asking that model to do every subsequent step. Resizing an image, adding up a bill and checking a required field already have good deterministic solutions.
I want to name the jobs before choosing models. Extracting a date has different requirements from planning a video. One has an output I can validate directly; the other involves choices the user may want to revise.
For extraction, start with a narrow schema and a validator. Handle missing information explicitly. For planning, preserve the draft and let the person correct it before expensive execution begins. Using the same large model for both may be convenient initially, but it hides those differences.
A fallback also needs a reason. Retry a transient provider failure, or escalate when the smaller model fails a check. Don’t keep sending the same poorly specified task to increasingly expensive models and call that routing.
These decisions show up in the interface. Some actions can be immediate. Others need a preview, a pause or an approval. I’d rather design those expectations alongside the route than add a spinner after discovering how slow it is.