Routing AI Agent Models by Step Cut Latency 86%, After Fixing Root Causes of a $48,000 Error

A follow-up post from a billing-agent team describes how it addressed upstream causes of an incident in which its agent settled a $48,000 credit that should have gone to a human. The team tested whether dumping a 350-page finance handbook into a long-context prompt could replace its retrieval pipeline, finding the cost gap real but modest and raising concerns about recall and traceability. The root cause traced to a naive chunking script that split a formula from the condition that made it inapplicable, so the agent saw the formula without its precondition. The team responded with section-aware parent-child chunking that returns full sections and version-stamps documents. The post also covers why a six-step ticket took 42.6 seconds and cost $0.184, and the latency gains from routing models by step.