Resources

Prompt caching savings by product shape: hit rates for chat, RAG, agents and summarizers, write vs read rates, and the break-even rule, from production.

Chargeback bills a team’s AI spend, showback just reports it. Why the order matters, six drivers that complicate AI cost allocation, and when to switch.

RICE scores reach, impact, confidence, and effort, and misses what an AI feature costs to run. A fifth term catches it before the feature ships.

Twelve usage, cost, and process signals that show an AI feature is losing money, grouped by where each one shows up first, plus the re-check rule.

The AI cost stack: three layers, not three departments arguing past each other. What each layer answers, where its number lives, and which article owns it.

Twelve cost drivers behind AI agent cost (steps per task, context growth, tool calls, retries, sub-agents) and the ranges each hit in production.

Twelve FinOps tool capabilities (allocation, anomaly detection, forecasting, unit economics) checked against inference spend on production AI products.

Why the cheapest AI API by list price loses on cost per resolved task once token split, caching, retries and re-prompts are counted. Model-swap math.

The AI business case a CFO signs: cost per request at launch and at 10x, break-even, cache-adjusted margin, variance tolerance. Plus two slides to cut.

Twelve cloud cost governance controls rebuilt for inference: pacing, driver-based forecasts, anomaly alerts, pool allocation, all run by finance alone.

Twelve AI pricing models compared on the margin curve as usage per seat grows, and the crossover where seat pricing goes negative, from production data.

Twelve FinOps KPIs tested on production AI products: cost per request, per subscriber, cache hit rate and forecast variance survived. Total spend did not.

How much AI costs per month for a chat assistant, summarizer, RAG search, or agent, from three inputs you can estimate today, with production ranges.

Nine cost drivers behind one LLM request (input, output, cached tokens, retries, tool calls) and the ranges each landed at on production AI products.

Cloud FinOps rules break once the workload is inference. Four assumptions that fail on AI products, the fix for each, and what changed once they shipped.