Last Black Friday, our e-commerce AI customer service system was running on a LangGraph multi-agent orchestration with GPT-4.1 as the default LLM. The system handled about 12,000 tickets per day during the peak window, with a routing layer choosing between a ClassifyAgent, a RetrievalAgent, and a ResponseAgent. Each ticket triggered 3–5 LLM calls on average. Our November invoice arrived at $14,820.71 — a brutal surprise that pushed our VP of Engineering to demand an immediate cost review.
We rebuilt the router in two days. Cheap and repetitive nodes (classifications, RAG rewrites, JSON extraction) were moved