In the same week that a large language model was caught giving wrong answers at scale inside a customer-service system, engineers at three different firms tried to call each other. None of them knew which number to ring. The problem was fixed within four hours, but the question remains: when an AI system misbehaves, whose job is it to pick up the phone?
We have playbooks for fires and outages. Not yet for AI.
From the outside the story looks simple. Look closer and every choice involves a trade-off: costs saved in one place reappear somewhere else, and those affected are rarely the ones making the decision.
Early data suggest the trend took shape several years ago; only now is it big enough to be noticed. The numbers are still noisy, but the direction is clear.
What comes next
The biggest question is speed. If change comes slowly, everyone has time to adapt. If it comes fast, laggards will pay. Either way, preparing today is cheaper than fixing mistakes tomorrow.
