Banks' AI Agent Deployments Raise Hidden Concerns: Zero Rollback Rate May Signal an Observability Blind Spot

According to a study by Sinch released in September 2026 covering more than 2,500 enterprises, about 69% of customer-facing AI agents in banking have been rolled back due to governance failures, and that proportion has not declined as projects mature. Among the financial institutions with the most robust governance, the rollback rate was actually the highest, reaching 80%.

The study points out that this counterintuitive phenomenon stems from differences in detection capability: mature projects roll back more because they can discover more failures, not because they have more failures in the first place. Failures in non-deterministic systems cannot be completely avoided, and a rollback means a failure has been caught—it is the system working as intended.

What is truly worth watching out for is a zero rollback rate. The report argues that this usually points to weak observability rather than genuine success. Traditional monitoring focuses only on latency, error rates, and uptime, and cannot judge the accuracy of AI outputs. 16% of rollbacks could not be fully diagnosed due to a lack of audit trails, meaning many banks still know nothing about the causes of failure.

In banking, every AI failure directly hits customer trust. Incorrect account balances, account details shown to the wrong party, or fraud alerts can all shake customers' confidence in entrusting their money to a bank.

Source: https://www.bankingdive.com/spons/bankings-ai-red-flag-from-0-rollbacks-to-deploying-agents-you-can-trust/831342/