The voice A/B test that picked the worse agent, and won by 4 points

We ran a clean A/B test between two versions of a phone agent. Variant B won by 4 points on our success metric. We shipped B. Two weeks later the escalation rate to human agents had gone up, and the "