The first AI feature often fails because it tries to prove too many theses at once. Teams chase generality, plug in a large model everywhere, and discover that evaluation, latency, and ownership were never defined.
A better start is a single user job with an observable success metric. Prefer features where partial automation still helps, and where a human can complete the task when the model is uncertain.
Document the non-goals early. Explicitly exclude unbounded chat surfaces or open-ended generation if your data controls and support capacity are not ready.
Treat the first release as instrumentation for learning: capture inputs that matter, label a small gold set, and decide who reviews failures each week. That operational habit matters more than model brand names.