The case
Why this is worth a budget line.
Written to be forwarded. If you are building the internal argument, this is the page to send to the person who was not on the call.
One
You are spending more on AI every month and cannot say on what.
Almost every company running AI in production can produce one number: the invoice. Almost none can break it down by customer, by feature, or by the task the business actually asked for.
That is not negligence. The tools were built for a world where one request meant one answer. An agentic system makes a single outcome cost tens or hundreds of calls, and a per-call view of that is not a smaller version of the truth — it is a different number. We drew that out on the front page: fifteen ordinary requests, one task, nineteen times the cost of another that looks identical per call.
The practical consequence is that you cannot price a feature, cannot tell which customer is expensive, and cannot tell whether an optimisation worked.
Two
From 2027, some of this stops being optional.
California’s SB 253 requires large companies to report Scope 3 emissions from 2027, with limited assurance — an independent third party checking the work. AI sits inside Scope 3 Category 1, and it is the part growing fastest.
Be careful with this argument, though, because the regulatory picture got weaker in 2026, not stronger: the SEC proposed full rescission of its climate rule in May, and CSRD was cut back to fewer companies with Wave 2 delayed to 2028.
So we do not sell on compliance, and you should not buy on it. Buy because you cannot attribute your own spend. The disclosure requirement is what makes the same measurement useful twice, and it is the reason to start before the year you need it — an assurance provider tests a twelve-month record, and you cannot produce one retroactively.
Three
Cost and carbon are the same measurement.
For AI inference both scale with compute. The tokens you were billed for and the energy that produced them are the same event seen twice, which means one instrument answers the finance question and the sustainability question at once.
This is why the product measures four resources and prices them, rather than choosing. It is also why the two teams that normally never share a tool end up sharing this one — and why the purchase is usually approved by both.
If the environmental side is not interesting to you, ignore it. The cost attribution stands on its own, and the carbon comes out of the same arithmetic whether you look at it or not.
Four
Why not the tools you already have.
| If you already use | What it will not do |
|---|---|
| LLM observability (Langfuse, Helicone, Portkey) | Measures per call and stores your prompts. No carbon, no water, no evidence pack, and the prompt storage is what closes the door on regulated buyers. |
| Cloud FinOps (Vantage, Finout) | Thinks in cost centres and line items. It cannot see a token, so it cannot tell you which customer or which feature spent the money. |
| An enterprise carbon platform (Watershed, Persefoni) | Excellent across the whole organisation and blind to AI specifically. No token visibility, so your fastest-growing Scope 3 line is an estimate from spend. |
| A spreadsheet and a public coefficient | Works, briefly. Coefficients go stale silently, and an assurance provider tests traceability and restatement, which a spreadsheet cannot produce after the fact. |
None of these are bad tools. They were built for different questions, and most companies that buy us keep at least one of them.
Five
Why us, specifically.
Because the number has to survive someone trying to break it. Every figure carries an uncertainty range, an evidence tier and the factors it came from; when a coefficient changes we publish a restatement rather than quietly moving history; and the engine is open source, so none of that has to be taken on trust.
And because we do not store prompts. No field in the data model can hold one, and the ingest endpoint rejects a request carrying one rather than dropping it quietly. That is a structural property, not a policy, which is what makes banks, health, legal and government addressable at all.
The honest limitation: for commercial API models your energy and carbon will report the weakest evidence tier, because no provider publishes per-model energy. We say so on the methodology page rather than letting you find out in a report.
What to do next
Look at the demo before you talk to us.
It is open without a login, and it is the fastest way to decide whether this is a real problem for you or a tidy one.
Sources for every figure on this page are on the methodology page, and the coefficients are in the repository (opens in a new tab).