Ramp describes how they built agentic risk operations to scale payment risk management without proportionally growing headcount. Rather than letting agents reason about probabilistic decisions, they use ML models trained on historical data for autonomous risk decisions, while agents handle intake, context gathering, and routing. The architecture includes shadow mode deployment, exposure budgets to cap dollar risk, operator feedback loops to generate labeled benchmarks, and resilient async execution with fallback model providers. Each edge case feeds back into evals, enabling risk operations to become more efficient as the business scales.
593 Impressions