PulseFlow Tecnologia
← Insights
·6 min read·framework-playbook·finops·agent-cost·metrics
PulseFlow Tecnologia

A FinOps framework for AI agents

AI agent cost doesn't look like traditional software cost: it's not a fixed per-user license, it's a function of how many tool calls, how many context tokens, and how many retries a specific task consumes. LangChain's State of AI Agents survey shows this already hurts: 22.4% of small companies cite cost as an adoption barrier - a number likely to grow as agents take on longer, more complex tasks.

Usage-based pricing is already the default model, not the exception

AWS Bedrock AgentCore charges Web Search at $7 per 1,000 queries - no upfront commitment, direct price per unit of use. That's not an isolated AWS choice: it's the standard any agent FinOps needs to assume as a starting point. Unlike a software license with predictable month-to-month cost, an agent's cost varies with the actual volume of work it performs - which makes granular measurement, not aggregate estimation, the only way to forecast spend with confidence.

Why "cost per response" is the wrong metric

Measuring only the final response's cost hides where the money actually goes: how many tool calls a task required, how many context tokens got processed at each step, and how many times the agent had to retry before succeeding. A task that "cost little" on the final response but required 5 retries and 10 tool calls has a much higher real cost than it appears to.

The framework's ten metrics

  1. Cost per execution - the total spend from start to finish of a task, not just the last call
  2. Cost per tool call - every called tool has its own cost, some much higher than others (external search, for example, charges per query)
  3. Cost per PR - for coding agents, the real cost of every generated pull request, including correction iterations
  4. Cost per requirement - from the initial requirement to delivery, covering every attempt
  5. Cost per bug fixed - comparable to the cost of a human fixing the same bug, to justify the investment
  6. Tokens per step - where context is actually being spent within the flow, not just the total
  7. Retries per task - every new attempt is additional cost; high retries signal a design problem, not just bad luck
  8. Latency per flow - time is also cost, especially when it blocks a human waiting
  9. Cost per context - the price of maintaining (or not) a large, expensive context window per call
  10. Cost avoided by automation - the metric that justifies everything: how much the same task would cost without the agent

What this means for whoever approves the budget

Without these ten metrics, approving budget for AI agents becomes a decision of faith, not data. With them, you can answer concrete questions: is this agent getting more expensive per task over time (a sign of degradation or more complex tasks)? Are retries climbing (a sign the agent's design, not usage volume, is the problem)? Does cost avoided by automation still outweigh actual spend, or has the expected savings stopped holding up?

Sources