Contents
What the survey actually asked Why the claim is sincere anyway The people who have tried it rate themselves hardest What this costs when the question gets specific How to test your own rating Frequently asked questions What a rating measures

Quick Answer

48 of the 252 senior finance leaders CloudZero surveyed can see nothing below a single line on the AI bill. Of those 48, 21 rate their outcome measurement as board-defensible or better, and 10 claim initiative-level ROI, which the survey defines as tracing return to a specific initiative and reallocating budget on it. For the 10, that is arithmetically impossible without allocation. For the other 11 it is merely shallow, which is the more interesting problem. The gap worth looking at runs between what finance believes it can prove and what it can take apart.

NOTE: This is a deeper look at one finding from Why Finance Caps What It Can’t See, CloudZero’s survey of 252 senior finance leaders, fielded between August and September 2026.

Self-assessment is a problem in every maturity survey, including this one. People are asked to rate a capability, and they rate the version of it they can picture.

In CloudZero’s survey of 252 senior finance leaders, two of those ratings were collected separately. One asked how finely AI spend could be attributed. The other asked how deeply return could be tied to that spend. Put side by side, they produce a group that should not exist.

Half the finance teams at the two weakest allocation tiers, 68 of 135, rate their outcome measurement as board-defensible or better.

What the survey actually asked

The outcome measurement question offered five rungs. The fourth was board-defensible ROI. The fifth was initiative-level ROI, described as tying return to specific initiatives closely enough to reallocate budget between them.

The allocation question offered its own five rungs, from seeing only a total on the bill up to attributing spend to a customer, product, or cost center and using that inside the quarter.

Allocation is the practice of apportioning costs to whoever is responsible for them. The FinOps Foundation treats it as a capability in its own right, with its own maturity stages, which is the same shape the survey’s ladder takes.

Both ladders in full, as respondents saw them.

RungAI cost allocationOutcome measurement
1We only know our total AI spend on the billWe don’t measure the return on AI spend
2We can split it by a rough category, but no furtherWe track usage or activity, but not business outcomes
3We allocate to teams or business units, mostly by handWe estimate ROI directionally, not defensibly
4We allocate to teams or business units systematically, and trust itWe measure ROI and could defend most of it to a board
5We allocate to customer, product or cost-center level, in-quarterWe measure ROI at the initiative level and reallocate based on it

Nothing in the survey stopped a respondent from picking the top rung on one and the bottom rung on the other. And ten of those 48 did exactly that.

Among the 48 respondents who see only a total on the AI bill, 44% rate their outcome measurement as board-defensible or better and 21% claim initiative-level ROI.

Among the 48 who see only a total, 44% rate their outcome measurement as board-defensible or better, and 21% claim the top rung.

Let’s look at that for a moment. Reallocating budget between initiatives requires knowing what each initiative costs. A respondent who can see one number for all of AI does not have that.

Board-defensible ROI is a different matter, and this is where the finding gets interesting. One bill line can produce an ROI number. Total AI spend in, an estimate of what the program returned out, and the arithmetic works. Nothing in the survey’s fourth rung requires allocation, so the 11 respondents who stopped there are not claiming anything impossible.

They are claiming something shallow. A single-line ROI holds until the first follow-up question, and boards ask follow-up questions: which initiative, what would we save by stopping it, what does the next dollar buy. A number that cannot be taken apart answers none of them. The survey measures the consequence directly. Among companies claiming board-defensible ROI they cannot allocate, 56% say their board asks for proof they cannot provide, against 54% of weak-allocation companies claiming no return at all. The claim doesn’t even buy them anything in the boardroom.

People overstate. There is a more useful reading than that.

A finance leader who says the AI program returned value is likely right. The support queue moved, the engineers shipped faster, a license got retired. The survey bears that out: weak-allocation teams claiming board-defensible ROI killed or paused initiatives at 9%, against 21% of their weak-allocation peers claiming nothing.

What that leader cannot do is say which part produced it. And the survey question did not ask them to prove it, only to rate it, so the rating reflects confidence in the outcome rather than confidence in the attribution.

That distinction stays invisible right up until a board asks a follow-up question.

The people who have tried it rate themselves hardest

Teams with strong allocation rate their own outcome measurement lower than teams with none.

Self-rated outcome measurement against allocation maturity: the teams with the strongest allocation rate themselves lowest.

Teams with strong allocation are more conservative about what they claim. They have attempted the decomposition, watched it fail on a shared model endpoint or an untagged pipeline, and adjusted their self-assessment down.

Teams that have never attempted it have nothing to adjust against. Their confidence is intact because it has not been tested.

This is a two-way cut of self-reported capability against self-reported capability, so it cannot prove which way the causation runs. What it does show is that the divergence is concentrated among teams that have not done the allocation work.

What this costs when the question gets specific

A board rarely asks whether AI produced value. It asks which part of it to fund again.

In the full sample, surveyed between August and September 2026, 47% of finance leaders report a board asking for ROI proof they cannot produce. That number runs straight through the group that rates itself well.

The pattern is not confined to this survey. Deloitte’s survey of 3,500 director-to-C-suite leaders, fielded between May and June 2026, found 71% of companies tracking total AI inference and API spend, 51% tracking token volume, and only 44% tracking per-unit inference cost. That is a broader population than finance and a different question, but it points the same way: the bill is visible, and what drives it is not.

The practical consequence is budget inertia. A finance team that cannot name the strongest initiative cannot name the weakest one either, so the whole line gets treated as a single decision: raise it, hold it, or cut it. AI cost allocation is what turns one decision into several.

How to test your own rating

Three questions will settle it faster than a maturity model will.

Name the single most expensive AI initiative you ran last quarter, in dollars. Not the vendor, the initiative.

Name the one you would stop. Then say what stopping it would save, to the nearest ten thousand.

Say which customer segment is cheapest to serve on AI, and by how much.

If those take research rather than recall, the rating is measuring the program. That’s a normal place to be. It becomes a problem when it goes to a board unlabeled.

Frequently asked questions

What a rating measures

The teams in this survey who claim the most are, on average, the ones who have decomposed the least. That’s what happens when a capability is rated by people who have never had to use it under pressure.

In practice that means naming the initiative, the dollars, and what stopping it would save, before someone else asks you to.

Before the next board meeting, would your own rating survive someone asking for the arithmetic behind it?

This is one of thirteen findings in the full report, which publishes the methodology and every response distribution alongside them.

Read the full report: Why Finance Caps What It Can’t See