Product · AWS Advanced Tier Services Partner
The FinOps copilot answers questions about your AWS spend in plain English — and every number it gives you has been verified against your own data before you see it. On every plan, including Free.
FinOps Copilot
Figures verifiedWhy did RDS go up last month?
RDS rose against the prior month, driven by two db.r5.large instances that ran the full period rather than the partial month before it.
Open RDS in Cost Explorer
And the month before that?
Illustrative. Grounded in your own months; the link is built by the app from a returned spec, never written by the model.
Why this is hard
Ask a general-purpose model about your cloud bill and you get a fluent paragraph containing numbers. Some will be right. You will not be able to tell which, and you will take it into a meeting. That is the failure mode this copilot is built around — not answer quality, but the fact that a wrong financial number reads exactly like a right one.
So the design question was never “how do we make it accurate?” It was “how do we make an unverified number impossible to ship?”
None of these is a prompt instruction. Each is a check the answer has to pass on the way out.
Before an answer leaves the server, a guard recomputes the period deltas and percent changes from your own data, then checks every dollar and percentage in the prose against them — and each returned figure against the month it claims to describe.
Not the sentence — the answer. If a single figure cannot be verified, nothing is shown. A partially-correct financial answer is worse than none, because you cannot tell which half to trust.
Answers come back carrying link specs — a service, a period, a dimension. The app turns those into Cost Explorer URLs itself. A hallucinated link is not something we filter out; it is something that cannot be expressed.
The context the copilot reasons over is assembled in your browser from data your session already holds. There is no cross-tenant query for it to make, because there is no query — the data arrives with the question.
Every figure the copilot returns is verified against your own data before the answer is released. If one number cannot be verified, you get no answer rather than a plausible one.
The copilot opens with a few suggestions drawn from your own account — the months you have, your largest service, whether you currently have anomalies. It does not suggest a question your data cannot answer.
“Why did RDS go up in August?”
Month-over-month attribution, with the delta and the percentage both verified against your own months.
“What were my biggest costs last month?”
The copilot offers this one itself — starter questions are generated from the periods and services your account actually has.
“Which services changed the most month over month?”
Offered only when you have two or more months loaded, because otherwise it is a question with no answer.
“Summarise my current cost anomalies.”
Offered when there are anomalies to summarise. Your anomaly set is part of the context, so the summary is of yours.
There is a daily question budget per account, so one runaway integration cannot spend the whole month’s allowance in an afternoon. It resets at midnight UTC.
The copilot is a drawer on your cost pages, not a separate destination. It already knows which account and which months you are looking at, so you never start by re-explaining your own context to it — and the deep links it hands back land in the Cost Explorer view that proves the answer.
Connect AWS read-only and the copilot is there on your first report. Free plan, free forever.
Refine is built and supported by HabileLabs, an AWS Advanced Tier Services Partner.
