The engagement
Meridian’s AI support copilot handled every ticket with the same frontier model regardless of complexity, and had no real-time visibility into token spend — monthly LLM costs had grown 6x in two quarters with no corresponding growth in ticket volume.
What we built
A full FinLLMOps implementation: multi-tier model routing sent classification and simple-resolution tickets to compact fine-tuned models while reserving the frontier model for genuinely complex cases, semantic caching eliminated redundant calls on near-identical tickets, and per-workflow token budgets with automated fallback policies caught runaway loops before they hit the invoice.
Outcome
Monthly LLM spend dropped 58% and P95 response latency improved 31% — because smaller models routing simple tickets respond faster than a frontier model handling everything — with no measurable drop in resolution quality. The engagement paid for itself in six weeks.