Skip to content
Nerdy beta

AI Cost-Savings Scorecard

Add to Space
 AI

Your AI bill isn't high because people use AI too much. It's high because of a handful of build decisions nobody revisits, and each one bills you again every month.

Overview

The AI Cost-Savings Scorecard estimates how much of your AI spend is structurally recoverable. Sixteen questions, about four minutes. You get a percentage range and a dollar range per year, plus the top five levers that would recover the money, ranked.

“Structurally recoverable” is a deliberate limit. Nothing here asks you to ration AI or slow down adoption. The cost it hunts is baked into how the system runs: context you resend uncached on every call, routine work pointed at a frontier model, jobs running live that could run overnight in a batch, and calls that never needed a model at all. One lever does tell you to skip the model, but only for work that follows fixed rules and returns the same answer every time, which was never a judgment call to begin with. These are properties of the design, not the usage, so fixing one pays every month instead of once. Compounding engineering makes your AI setups worth more over time. This is the same idea aimed at the invoice.

Your usage dashboard won’t find any of it. It shows that people are logging in and tokens are moving, which is the question you already know the answer to.

No account, no cookies, no trackers. Email yourself the results if you want something to bring to a budget conversation.

The nine levers

  • Model right-sizing — route routine work to cheaper models.
  • Prompt caching — stop paying full price to resend the same context.
  • Batch discounts — take the roughly 50% discount on work that can wait.
  • Context engineering — send the excerpt the task needs, not the whole file.
  • Output discipline — cap the expensive half of every call.
  • Agentic guardrails — keep loops and retries from burning the budget.
  • Deterministic harness — give agents scripts for the mechanical steps.
  • When to skip AI — move rule-following work to plain code.
  • Price sensitivity — break the bill down so nothing expensive hides.

You won’t need all nine. The scorecard’s job is naming the ones charging you rent today.

Resources