Four production pitfalls that turn RAG demos into broken chatbots

A common pattern we see: a Series A team builds a RAG assistant, runs a 50-question internal demo, ships to production, and within two weeks the support inbox is full of "the AI gave me a wrong answer