Your voice AI is corrupting records right now, and your dashboard will never tell you.
A production voice AI does not fail loudly. It fails silently, and the dashboard keeps showing green.
This week Craig ran two session-log audits on our IT-desk voice AI at Norton. The July 17 pass, 55 calls and 46 tickets, found that "account expired or inactive" has no intent branch at all. The bot only knows lockouts and password resets, so it guesses twice and the caller gives up. No ticket, no forward, no trace. You would never catch that in a metrics view because nothing errored. The call simply evaporated.
The July 23 pass was worse. Across 98 calls and 304 minutes, the model ignored the locked-field contract after caller info was confirmed, re-collected fields, and overwrote confirmed data. Corrupted records actually shipped. One ticket persisted a speech-to-text-garbled email onto the caller's profile. Our working hypothesis is that locked-field enforcement fails open. Same audit found the escalated flag reading false on all 98 sessions, including real forwards, so the escalation analytics were quietly undercounting.
Here is the uncomfortable part for any operator running AI against a live customer or IT queue: none of these findings showed up as failures. They showed up as normal traffic. A dashboard measures what you told it to measure. It cannot measure a contract violation you did not know to instrument.
That is why the audit cadence is not overhead. It is the control. A voice AI in production needs an adversarial reviewer whose entire job is to assume the system is lying and prove it. Neither Norton finding was flattering. Both are exactly the point. The audit caught the contract violations before the client did, before the corrupted data compounded, before a silent failure became a churned account.
AI Value-Creation Diagnostic | All insights