Building a support chatbot a US business can trust: retrieval, guardrails and human handoff

The resolution-rate metric needs an explicit endpoint. A customer leaving the widget, a bot handing off successfully, and a human eventually closing the ticket are three different outcomes. Counting all three as resolved would make it hard to tell whether the chatbot actually answered the question.

I would keep abandoned sessions, bot-confirmed resolutions, and completed human handoffs separate, then connect the last category to the help-desk outcome. That makes the weekly content review more informative: unanswered questions can indicate missing content, while an accepted handoff with no eventual resolution points to a different operational problem.