Cloud cost optimization, evidence first

Most cloud bills grow faster than anyone’s ability to explain them. A number goes up, a dashboard shows it went up, and then the hard part begins: which service moved, who owns it, whether the change was intentional, and what it would cost — in risk, not just dollars — to unwind. By the time that chain is reconstructed by hand, the month has closed and the next surprise is already forming.

Optimization tools that jump straight to “here is what to delete” skip the part that actually matters: the evidence. A recommendation with no source, no freshness, and no confidence is just a dare. And a tool that acts on that dare automatically is a liability, not a feature. Real cost work needs a shared, trustworthy picture first — and a human in the loop for anything that changes.

CloudJaeger’s first connection is a read-only IAM role with a policy we publish before you connect. Nothing runs on your workloads, no agent is installed, and you can delete the role at any time. From that role we sync your billing data and inventory into one place, so spend by service, region, and account is bill-backed — reconciled against the invoice rather than estimated. On our own AWS estate that reconciliation lands within about 1% of the Cost and Usage Report; the same check runs on yours.

With the bill in view, ownership becomes answerable. Every service resolves to an owner source — configured, inferred, or honestly marked unknown — and unowned spend is shown at the top rather than hidden. That is the difference between a chart that says spend rose and a working picture that says which team’s which service rose, and why.

A lot of cloud cost is decided at design time, not runtime — the instance family, the data transfer path, the redundancy that a workload may not need. CloudJaeger includes a deterministic architecture review that produces pillar findings with severities and a rough-order-of-magnitude cost figure, validated against golden scenarios at every release. It connects the shape of a system to the shape of its bill, so an optimization is a design decision with a dollar figure, not a guess.

Findings from every engine land in one ranked, advisory queue. Each item blends dollar impact, confidence, reversibility, risk, and effort — scored deterministically, with a plain-English rationale. The language model only writes the explanation; it never decides. Every row shows where the number came from, how fresh the underlying data is, and how strong the evidence is, so “safe and easy” can rank ahead of “big and scary”.

Dollar figures are projected. They stay labeled as estimates until a real invoice moves and grades them against your bill — never a promise, and never presented as money already saved. That honesty is the point: an optimization program lives or dies on whether its numbers survive contact with the next invoice.

CloudJaeger can execute, but it ships with execution turned off, and that is deliberate. Arming any change requires a written mandate with caps and a named approver on every run, and live purchases are hard-blocked today. The default posture is that CloudJaeger recommends and a human decides. You get the leverage of a system that has already done the analysis, without handing it the keys to your account.

On cost coverage beyond AWS, we stay precise: AWS today; Azure and GCP are connected and labeled by proof level. We would rather label a zero-spend estate as zero than pad a number to look broader than it is.