Commentary: every figure in this piece comes from a cited public source — the FinOps Foundation’s State of FinOps 2026 survey — and is presented as that source’s finding, not as our measurement.
It usually arrives as a one-line message: "Can you walk me through the AI spend? Board is asking." Or it arrives inside a fundraise, when a diligence list includes a row your team has never had to fill in before. Either way, the question lands on whoever owns the API keys — and the honest first reaction, in most companies, is a quiet scramble.
Not because nobody is watching the spend. Because everybody is watching a different version of it.
If the CFO question hasn't reached you yet, the industry data says it is en route. The FinOps Foundation's State of FinOps 2026 survey — published in February 2026, drawing on 1,192 respondents representing roughly $83 billion in technology spend — reports that 98% of organizations now manage AI spend in some form. Two years earlier that figure was 31%; in 2025 it was 63%. Whatever the exact trajectory inside any single company, the direction is unambiguous: AI cost has moved from a curiosity line item to a managed spend category in roughly two budget cycles.
The same survey asked practitioners what capability they most want and got a strikingly specific answer at the top of the list: "granular monitoring of AI spend (tokens, LLM requests, GPU utilization)." Read that closely. The market's number-one ask is not a better dashboard theme or another forecast model — it is resolution: seeing the spend at the grain where it is actually generated.
We read that as practitioners telling on their tooling. You don't beg for granularity you already have.
Here is the structural problem the survey number hints at but doesn't spell out. When the CFO question arrives, most engineering teams have plenty of AI cost data — a usage dashboard from each provider, an estimate inside the observability stack, maybe a spreadsheet someone maintains. What they don't have is one answer that all of those sources agree on, and in particular one answer that agrees with the number finance can see: the invoice.
The dashboard prices usage from a public price sheet. The invoice reflects how the provider actually bills — cache multipliers, tier pricing, restated usage, mid-cycle rate changes. The two are produced by different pipelines that never check each other. So the room ends up with three figures that are each defensible in isolation and collectively embarrassing: telemetry says one thing, the estimate says another, the bill says a third.
A CFO does not experience this as a data richness problem. A CFO experiences it as nobody owns this number — which is a much worse impression than a big number that is owned, explained, and tied to the bill.
It helps to translate the question. When finance asks about the AI bill, they are almost never asking for token counts. They are asking a small set of institutional questions:
Notice what's absent: none of these questions require real-time streaming granularity. They require reconciled granularity — detail that has been tied back to the billing document, so that when it is aggregated for the board, the total is the bill, not an approximation of it.
Separate your estimated and reconciled figures, explicitly. The single highest-leverage move costs nothing: label every AI cost number that circulates internally as estimated (derived from telemetry and price sheets) or invoice-tied (checked against the bill). The label instantly reveals which category your board pack currently draws from.
Nominate an owner for the invoice relationship. Not for the dashboard — for the bill. One person whose job includes being able to explain, monthly, how the invoice total decomposes and whether it matches what usage implies. In most AI-native companies today this role exists for cloud spend and simply hasn't been extended to AI providers.
Do one manual tie-out before you're asked to. Take last month's invoice and your internal estimate for the same period and diff them, even at total level. If they match closely, you now know that — and can say so with a straight face. If they don't, you have found the gap on your own timeline instead of mid-diligence, which is the cheapest moment you will ever find it.
Build the granularity the survey respondents are asking for — but anchor it to the invoice. Token-level and request-level visibility is genuinely valuable; the survey's top ask is right. The refinement worth adding: granular data that reconciles upward to the billing document is evidence, while granular data that doesn't is just higher-resolution estimation. Resolution and reliability are separate axes, and the CFO conversation needs both.
The pattern across all four steps: the CFO question is not answered by more data. It is answered by agreement between the data and the bill — demonstrated, not presumed.
That agreement — demonstrated monthly rather than assembled under pressure — is precisely what ToTheCent produces: deterministic reconciliation of AI provider spend against the invoice itself, with anything unmatched shown rather than smoothed over.