EP005: Inside the Cost of Running Agents - Stories from the Rehearsal Room #2

**The FinOps Daily Brief: Inside the Cost of Running Agents**Intuit's accounting agent went live to ten million small businesses last year, and the cost curve looked like a wall. That's what AI agent cost looks like in production right now, and the practitioners closest to the problem are still figuring out the levers. J.R. sat in on three FinOpsX rehearsals last week from speakers wrestling with the same question, and this episode is what stood out.Eran Lador and Balachandra from Intuit walk through the QuickBooks Accounting Agent, which broke its cost into three buckets that every team running an AI product is going to recognize. LLM cost, vector DB cost, and ordinary cloud cost. The framework that lands is the accuracy / cost / latency triangle. You optimize for two of the three, and the third one suffers, which turns AI cost into a design decision per workload rather than an infrastructure number rolled up at the end of the month. Safi Siddiqui follows with the cleanest map of AI cost from a practitioner this year. Three layers. Infrastructure and data at the bottom, models in the middle, agents at the top. Each layer has its own attribution tools, and the agent layer is where unit economics finally become possible. Paul Abruzzo from Nvidia opens the lightning talk track at FinOpsX with the GPU economics underneath all of it, and the distinction most teams haven't drawn yet between training and inference cost shapes.Stack the three together and you get the new FinOps job description. Three cost groups, three layers, the triangle, build small, instrument with OpenTelemetry, and have an answer when the CFO asks.Three takeaways close the episode. FinOpsX is June 8 in San Diego, and all three speakers will be on stage live.