Automating half my support team's tickets surfaced the real bottleneck

I lead support at a mid-size SaaS, eleven agents handling roughly 9k tickets a month. Last quarter we wired Claude into Zendesk to triage, draft, and auto-resolve tier-1 tickets. Auto-resolve rate hit 47%. Median first-response dropped from 4h to 11m. On paper, a win. What actually happened: the remaining 53% got harder. Our agents now spend their day on escalations, refund edge cases, and angry enterprise customers whose easy tickets never reach them anymore. Two people quit within six weeks. The job they signed up for, a mix of easy wins and hard ones, no longer exists. The dopamine of closing twenty quick tickets is gone. We also lost our training pipeline. Junior hires used to learn the product by grinding password resets for three months. Now there's no on-ramp. I'm rebuilding it with synthetic ticket drills, but it's worse than the real thing. The productivity number is real. The org design assumed an easy-ticket buffer that no longer exists, and nobody modeled that. Automate the bottom of the funnel, plan for the morale and training cliff.