Neither number is new. What is new is that your regulator has started naming the failures on both sides.

The first wave of AI in compliance made your analysts faster: Summarize the alert, draft the narrative, retrieve the policy. Published benchmarks put the uplift at 15% to 20% — useful, but nowhere near a 2% detection rate.

The harder problem sits underneath your models. Every detection model in your stack is calibrated on historical alert data, which under-represents the crime nobody caught. You cannot backtest your way out of a censored record.

Written for the risk, compliance, and model governance leaders who must answer:

• Why both KYC and AML error types are now regulatory events rather than tuning metrics
  
• Why the copilot era plateaued, and what changes when the unit economics move instead of the workflow speed
  
• Why historical alert data cannot validate the thing it never caught, and where synthetic populations fit
  
• How an agentic compliance architecture stays governable — policy-bound decisions, an auditable evidence trail, and your data never leaving your own AWS environment