Why I Built a Risk Score Instead of a Buy Button

The obvious feature to build here is a button. "You're spending $4,200/month on m5.xlarge — buy a 1-year Savings Plan and save 30%." One click, instant discount, everybody's happy.

I almost built that button. I'm glad I didn't.

Here's the problem with the button: it's only looking at the last three months. It has no idea whether the workload it's telling you to commit to will still exist in month nine. And a Reserved Instance or Savings Plan isn't a coupon — it's a bet, paid up front or amortized monthly, that a specific slice of your infrastructure will look roughly the same for a year or three. Get that bet wrong and the "savings" tool just talked you into a liability.

So instead of a buy button, I built a Commitment Risk Score — a feature whose entire job is to sometimes tell you not to buy anything yet.

The feature that argues with itself

The Commitment Risk Score pulls four signals straight from Cost Explorer for an account: how much your top instance families churn month to month, how volatile your spend is, how long your individual resources actually live, and how well you're already using the commitments you have. It weights them (35/25/25/15) into a single 0–100 score, and that score maps to a recommendation — anywhere from "3-year Convertible RI, take the discount" down to "on-demand and Spot only, buy nothing."

That last outcome is the part that made this feature interesting to build. Most cost-optimization tools are graded on how much they tell you to save. This one is graded on how honest it is about when a "savings" purchase would actually be a mistake. It costs about $0.04/account/month to compute — four GetCostAndUsage calls, refreshed weekly — the recommendation quality has nothing to do with the compute cost, so there was no excuse to cut a signal to save pennies.

I wrote up the full math — the Jaccard-distance churn calculation, the coefficient-of-variation volatility threshold, the worked examples — in a separate deep dive, because the mechanics deserved their own space: The Commitment-Risk Score: should you buy that RI?

What almost went wrong

The instinct that almost got me was averaging. My first pass weighted all four signals close to evenly, because "why not, they're all relevant." It took building the worked examples to see the problem: a team mid-Graviton-migration with high churn but a stable dollar total would score fine on an even-weighted average, because volatility and churn partially cancel out in the wrong direction. The churn signal needed to dominate — 35%, not 25% — because a changing instance mix is structurally the most common way a commitment gets stranded, regardless of what the topline spend number is doing.

The existing-waste signal ended up smallest at 15%, which felt backwards at first — isn't "you're already wasting money" the most damning fact? It is, but it's also the least predictive one. It tells you about the past, not whether the next commitment will strand. It stayed in the score because it's a legitimate red flag, but it doesn't get to drown out the forward-looking signals.

Why this is the right shape for a cost tool

CloudWise's whole premise is read-only: we look at your AWS usage and billing data, we never touch your infrastructure, and every dollar figure we show you needs to survive you checking it against Cost Explorer yourself. A recommendation engine that only ever says "buy more" doesn't survive that scrutiny for long — eventually it recommends a commitment that strands, and you stop trusting the number.

A feature that's willing to say "not yet, here's why" is the one worth trusting the next time it says "yes, buy it." That was true before I shipped this, and building it just made me believe it more.

---

CloudWise is an AWS cost optimization tool for startups — 191 automated waste checks including commitment-risk scoring, read-only by design, starting at $19/month. Run a free scan at cloudcostwise.io.