Most companies bought AI the easy way: a premium license for everyone. Caio and Marina break down why that's become invisible waste on engineering teams, use Claude Code and Goose as opposite examples, and introduce a new discipline taking shape: AI FinOps. The question is no longer which AI is best — it's which AI is good enough for each task.
In this episode
01 Opening: the waste nobody sees
- Caio's hook: maybe your company isn't spending too much on AI — maybe it's using expensive AI for cheap problems. Marina reacts the way the audience would, a little skeptical.
- The common scenario: company buys a horizontal premium license, everyone gets the same thing, nobody measures who uses it for what.
- The episode's key turn: the next big savings in tech isn't firing devs, it's stopping paying for premium AI on tasks that don't need premium AI.
- Marina raises the practical question: okay, but how do I know I'm wasting money? Caio teases that the numbers tell the story.
02 The scary bill
- Caio drops the real numbers: Claude Pro at 20 bucks, Team Standard 20 per user, Premium 100, and the kicker — Claude Code token consumption, 150 to 250 dollars per dev per month in enterprise.
- The 50-dev simulation: ranges from 12k a year on Standard to 150k a year using Claude Code heavily. Marina reacts to the jump.
- Caio cuts off a cliché on the spot: this doesn't mean Claude is too expensive. It means AI usage needs to become cost management, just like cloud became FinOps.
- Marina builds the bridge: so it's like when everyone realized the AWS bill was blowing up and had to learn to control it.
03 The mistake of treating everyone like a power user
- Caio separates the two worlds: there are people refactoring legacy codebases, debugging hard bugs, touching architecture. And there are people asking for variable names and task summaries on the same 100-dollar license.
- Concrete example for Marina to picture: the dev using Claude Code to understand an entire codebase justifies the cost. The one only writing simple docs doesn't.
- The strong line: the problem isn't paying for Claude, it's paying for Claude on a task something else could solve.
- Marina pushes back: but doesn't giving everyone good tools boost team morale? Caio answers with the difference between purchasing simplicity and operational efficiency.
04 Goose enters as a counterpoint
- Caio introduces Goose without the hype: open source agent from Block, Apache 2.0 license, runs on your machine, has desktop, CLI and API, supports MCP and lets you pick among several model providers.
- The honest disclaimer that builds credibility: Goose is free in license, not in total cost. You still have tokens, setup, governance and maintenance.
- The line that sums it up: open source doesn't eliminate cost, it gives you back choice. You decide when to use an expensive, cheap or local model.
- Marina asks the obvious: so is it Claude or Goose? Caio knocks down the false choice, says the right answer is a usage matrix.
05 The AI stack by complexity, and governance
- Caio sketches the layers: premium for critical and complex tasks, open source for medium tasks and internal workflows, local or cheap models for simple tasks, and humans for sensitive decisions.
- The security point a lot of people forget: a code agent touches files, runs commands, accesses data. Saving on licenses without governance can get expensive.
- The minimum policy Caio defends: who can use which agent, in which repo, with which permissions, which logs, and which human review.
- Marina lands it: so savings without clear rules is a trap. Caio confirms with a real risk example.
06 Practical wrap-up: engineering AI FinOps
- Caio names the new discipline: AI FinOps — controlling tokens, agents, licenses, usage limits, consumption per user and return per task.
- The second phase of adoption: the first was giving everyone the tool, the second is figuring out who actually needs what.
- The question Caio leaves for every CTO: am I buying productivity, or am I buying hype per user?
- Action item for the listener: before renewing a contract, make a simple map of who uses it, for what, and how much it costs per task. Marina closes by challenging the audience to look at their own bill.