From demo to delivery, the missing piece is often responsibility, not model quality
I still think this is one of the easiest traps to fall into. Many demos succeed because the path is unusually clean. Delivery systems are not clean. Real use introduces uncertainty, approvals, ownership questions, and failure recovery needs. That is why the gap between demo and deployment is often about responsibility design rather than raw model quality.
A stronger model inside a vague system can be much harder to ship than a weaker model inside a workflow with clean review and handoff points. That sounds almost obvious once written down, but teams still underestimate it all the time.