Sinch's latest survey of 2,527 senior decision-makers across ten countries found 74% of enterprises have rolled back a live AI customer-communications agent — and 81% among those with the most mature governance frameworks. My first reaction wasn't alarm. It was recognition.
More governance does not mean fewer failures. — Sinch, The AI Production Paradox
The rollback is not the failure
In my experience, the org that never rolls back is usually the one that can't see what its agents are doing. Sinch's two biggest rollback triggers — PII leakage at 31% and hallucinations at 22% — are runtime behaviors, not design flaws. No approval board catches a live agent repeating a customer's account number mid-call. Mature-governance shops roll back more because they instrument more. A rollback is a control that fired; leaving a leaking agent in production is the actual failure.
Governance is aimed at the wrong phase
Most enterprise AI governance is pre-flight: risk registers, review committees, model cards, approval gates. The failures in this data happen post-launch, per-conversation. I once watched a team spend six weeks in review gates before pushing a support agent into production; it leaked a client's contract number in its third live call. The board had signed off on a model card, but no one had ever tested the agent against real transcripts. That's why guardrail maturity didn't lower rollbacks — it added paperwork before launch and left the runtime unprotected.
The takeaway
Stop counting approvals; measure time-to-detect and time-to-kill. Fund runtime controls — real-time PII redaction, hallucination detectors, per-agent kill-switches, and alerting that pages a human when confidence drops. Treat every rollback as a post-mortem that hardens the next launch, not a black mark on the team.