Automated half my support team's tickets and discovered the other half got harder

I lead support for a B2B SaaS, team of 14 before, 7 now. We routed tier-1 to an agent pipeline last September: refunds, password resets, billing lookups, status pages. Deflection sits at 52% nine months in. The part nobody warned me about: the remaining tickets are uniformly awful. My agents used to context-switch between three easy tickets and one gnarly integration bug. Now it's gnarly all day. Average handle time went from 7 minutes to 23. Two of my best left within four months, both citing burnout, not pay. I've started padding shifts with deliberate "easy" work, manual QA of the bot's edge cases, just to give people a break from being the escalation layer for everything the model couldn't handle. The productivity math looked great on the quarterly review. The retention math is going to eat it. If you're automating a tier, budget for the morale cost of concentrating the hard stuff, or rotate people out of queue work entirely. We didn't, and I'm paying for it now.