Problem
Ramp processes over $200B annually on Cards and Bill Pay and keeps adding payment rails, partners and countries (35+ at the time of writing). Each addition opens a new risk surface. Risk operations teams handle policy exceptions, payment reconciliation, partner work and edge cases, and their headcount normally grows with the business. Ramp wants the opposite: operations that get cheaper per payment as volume grows. The catch is that every operational decision creates real dollar exposure, and many of these decisions are predictions, not reasoning tasks.
Example from the article: whether to release a payment depends on the probability that the debit returns, which an agent should not estimate by itself.
Solution in brief
The LLM agent only does intake, context gathering, classification and routing. Every autonomous risk decision comes from a trained model or an approved policy that the agent calls as a tool. Agent and policy are evaluated separately, and the agent gets a capped exposure budget once its metrics pass thresholds.
Questions this case answers
- How to automate a high-stakes operational workflow with an LLM agent when each decision carries direct financial exposure and must stay auditable
- How to evaluate the routing quality of an agent separately from the quality of the decision models it calls
- How to automate risk operations with LLM agents without letting the agent make the risk decision itself?
- How to evaluate an operations agent separately from the models and policies it calls?
Tools and methods
- Agent with models and policies as tools — gathers context, classifies the request, calls the model or policy, routes the result
- Operator feedback on every ticket — turns manual work into labeled cases for a benchmark
- Shadow mode — runs a new agent configuration on real production requests before it goes live
- Exposure budget — caps the total risk the agent can carry at any moment
Google or a one-time email code. No password.