We audited 98 staging calls and the dashboard showed nothing wrong.
A voice AI does not fail loudly. It fails silently, and the dashboard keeps showing green. The cheap place to find that out is staging, before real callers are on the line.
This month we ran two adversarial audits on an enterprise IT-desk voice agent still in staging. Not a metrics review. A line-by-line read of session logs by a reviewer whose only job was to assume the system was lying and prove it.
The first pass, 55 staged calls and 46 tickets, found that "account expired or inactive" had no intent branch at all. The agent only knows lockouts and password resets, so it guesses twice and the caller gives up. No ticket, no forward, no trace. Nothing errored. In a metrics view that call simply does not exist.
The second pass was worse. Across 98 calls and 304 minutes, the model ignored the locked-field contract after caller info was confirmed, re-collected fields it already had, and overwrote confirmed data. One ticket persisted a speech-to-text-garbled email onto the caller profile. Our working hypothesis is that locked-field enforcement fails open: when the guard is ambiguous, the write goes through. The same audit found the escalated flag reading false on all 98 sessions, including sessions that really did forward, so the escalation analytics would have been undercounting from day one.
Every one of those findings came out of a staging environment, which is the whole point. Ship that build and the corrupted record is not a test profile, it is a customer's, and the first person to notice is the customer.
Here is the uncomfortable part for any operator preparing to point AI at a live customer or IT queue: none of this presented as failure. It presented as normal traffic. A dashboard measures what you told it to measure. It cannot measure a contract violation you did not know to instrument, and a clean staging dashboard is not evidence that the agent honored its contracts. It is evidence that nothing you thought to watch for went wrong.
That is why the audit cadence is not overhead. It is the control, and it belongs before launch and then on a schedule after. Neither finding was flattering. Both are exactly why the cadence exists. The failures got caught while they were still cheap, before the corrupted records were real ones, before the silent drop-offs were real callers, before a quiet contract violation became a churned account.
AI Diagnostic | All insights