Everyone demos AI support agents that magically resolve everything. Here are our real numbers after ninety days in production. The agent fully resolves thirty eight percent of conversations. It escalates cleanly with context attached in another twenty percent. And yes, it fails awkwardly about six percent of the time, which is why the human handoff matters more than the model. What made it work was not the model. It was the knowledge base. We spent two months rewriting our help docs to be precise, and the agent got twice as good without touching the model. Garbage in, garbage out applies to support AI more than anywhere.