Fintech · Customer support AI
Support automation blocked 41 unsafe customer promises in 14 days
Published April 12, 2026
Outcome
41 unsafe promises blocked · 100% audit coverage · zero compliance escalations
The problem
Customer ran an AI support bot replying to tier-1 cases. Compliance discovered the model had occasionally promised refund timelines and policy exceptions outside operator authority.
The risk
Unauthorized refund language is a CFPB exposure surface. The customer's compliance team flagged the bot as a top-3 risk in their quarterly review.
Implementation
Routed bot replies through the gate with a policy version that flagged refund-timeline patterns, money-back guarantees, and personal-data references. The compliance reviewer used the dashboard mode; engineers used webhooks for analytics.
What changed
- 41 unsafe promises blocked in 14 days.
- Zero compliance escalations during the rollout window.
- Audit log exported weekly to compliance with a single button.