"In terms of where the transition usually breaks down, a couple of breakdowns come to mind. The first one is about dealing with real-world usage, where real users that enter the inputs that an AI system might not be able to anticipate. Like a recent viral example where someone was chatting with a chatbot system for restaurants and they asked about how to solve technical problems. I think the second one is about latency and cost. You know, these don't show up until real production loads hits. So these are two symptoms of architectural decisions that were made during prototyping and might be very expensive to unwind."