They were answering 2,100 tickets a week with a team of forty, and hiring to keep up. We measured the work for ten days, then built one agent to handle the eighty percent that repeated.
The queue was healthy on paper. Response times were inside target, satisfaction sat at 4.6, and nobody was complaining. What the audit found was the shape underneath: sixty-one percent of tickets were one of nine questions, and the answer already existed in their help centre.
Each of those took an agent four minutes on average, most of it spent finding the right article and rewriting it in their own voice. Four minutes, thirteen hundred times a week.
The second finding mattered more. The team was slowest on the tickets that actually needed thought, because those queued behind the repetitive ones.
Reads the ticket, matches it against the nine known shapes, and scores its own confidence before doing anything else.
Writes the reply from their help centre in their tone of voice, citing the article it used so the agent can check it in one glance.
Anything below the confidence threshold, or touching billing, goes to a person in Slack with the draft attached and one click to send.
Ninety historical tickets with known-correct answers, run on every change. It has caught two regressions since launch.