Agentic reliability and evaluations : Enterprises that got burned by a bad eval are the most likely to remove humans from the loop, not the least

There's a cruel irony in this week's agentic AI data: the enterprises most traumatized by broken evals aren't the ones crawling toward caution. They're the ones yanking humans out of the loop fastest. Across 108 enterprises, trust in automated evaluation nearly *tripled* in July (5% → 13%), and the "evals don't match reality" grumble dropped by 10 points — yet exactly the same share as last month, just under half, shipped an agent that passed evals and face-planted a customer. That's not a reliability leap. That's a coping mechanism. Why does this matter? Because trust in your evaluation tool is quietly becoming a proxy for *emotional exhaustion*, not evidence. A bad eval used to be a red flag. Now it's a broken blood pressure monitor that says "fine." The decision to remove human oversight after getting burned signals that teams aren't optimizing for accuracy — they're optimizing for closure. The dashboards look trustworthy, the failure rate just doesn't budge. The signal? It charts in "agentic reliability" is the same game. Blue both. Closer: The eval wasn't wrong because the model was wrong. It was wrong because we refused to read the receipt.