The off-the-shelf support bot the team initially adopted had an end-to-end resolution rate of only about 10%, and did not expose the system prompt, tool calling, or RAG infrastructure.