From ten-minute support responses to thirty seconds.

The queue was the product problem

Every support query was answered by a person, and the average wait was ten minutes. At that latency users stop treating support as something that will help them now, which changes what they do next — they abandon the task rather than wait.

What was built

A chatbot layer sitting in front of the human queue rather than replacing it. The interesting design question in this kind of system is not the answering, it is the handover: which queries the bot should refuse, and how a user gets to a person without repeating themselves.

Queries it could not resolve escalated rather than guessing, which is the difference between a support bot people use twice and one they route around.

Result

Average response time went from ten minutes to about thirty seconds, and roughly 70% of queries were closed without a human touching them.

Unlike the call counts on the other engagement, these are not my measurements. They come from the client's own support queue, which I read rather than audited — treat them as directional. The client is named now; the figures are still theirs rather than yours to check.