2026 Report

Why Finance Caps What It Can’t See

State of AI ROI: How finance leaders are managing AI spend, and how they’re approaching FY2027

252 finance leaders · 79% CFOs, 21% VPs of Finance · fielded August–September 2026

Executive Summary

Finance is not debating whether to fund AI … or whether that funding is concerning. 87% of the 252 finance leaders surveyed are increasing FY2027 AI budgets, and 61% say they are already spending more than they can justify with measurable outcomes. Those are mostly the same people.Finance is not debating whether to fund AI … or whether that funding is concerning. 87% of the 252 finance leaders surveyed are increasing FY2027 AI budgets, and 61% say they are already spending more than they can justify with measurable outcomes. Those are mostly the same people.

This report traces what happens between those two facts. C-suites and aggressive boards push AI spending, and the resulting expenditures outrun finance’s ability to explain it. Allocation, the ability to break AI spend down below a single bill line, is the weakest capability finance has, and the group that owns the AI budget most often, the CFO’s office, is the group least likely to be able to see it. That’s not for want of rigor, but because allocation depends on systems finance doesn’t (or can’t) run.

When the bill lands, 86% have absorbed a concrete consequence: an initiative killed or paused, a delay in hiring, an investment ask rejected, an unplanned mid-year reallocation, or an uncomfortable board conversation. Two-thirds absorbed more than one.

The most common two consequences are subtler than any on that list. 44% approved AI spend without reliable projections, and 36% couldn’t produce a credible ROI number when the board asked.

The response is friction. Faced with a number they cannot defend, most finance teams make FY2027 spending harder rather than clearer: 27% cite new approval gates as their primary change, against 18% tying funding to demonstrated ROI. The visibility/ROI gap costs them. Teams who implemented approval gates delayed hiring at 33%, against 19% of everyone else, despite planning identical increases. They absorb more damage to control the same number.

There is a third option, and 23% of the sample are already running it. Companies that can both allocate AI spend systematically and measure AI outcomes rigorously plan the same increases as everyone else but suffer half the consequences. Measuring return on its own is not enough: companies claiming ROI without allocation are statistically indistinguishable in board-level conversations from companies claiming no return at all.

FY2027 is not the year AI spending slows down. It is the year the excuse for not measuring it runs out.

The most common two consequences are subtler than any on that list. 44% approved AI spend without reliable projections, and 36% couldn’t produce a credible ROI number when the board asked.

There is a third option, and 23% of the sample are already running it. Companies that can both allocate AI spend systematically and produce board-defensible ROI plan the same increases as everyone else but suffer half the consequences. Measuring return on its own is not enough: companies claiming ROI without allocation are statistically indistinguishable in board-level conversations from companies claiming no return at all.

FY2027 is not the year AI spending slows down. It is the year the excuse for not measuring it runs out.

Findings

AI spend is not slowing. 87% are increasing FY2027 AI budgets, and 45% of those by more than 20%, even as 61% admit they cannot justify what they already spend.
Owning the budget ≠ understanding it. CFO-led programs sit at the two weakest allocation tiers 65% of the time, against 43% of engineering-led ones, despite forecasting no more accurately.
Blind teams underspend, they don’t overspend. 66% of under-budget teams have weak allocation against 46% of over-budget ones. Poor visibility produces hesitation, not chaos.
Friction is the default lever. 27% of the market names approval gates as its primary FY2027 change. Gates work without visibility, which is exactly why they get chosen over other AI spend optimization strategies.
Gates carry a measurable cost. Teams leading with gates killed or paused initiatives at 20% compared with 7% of teams tying funding to ROI, and delayed hiring at 33% compared with 19% of everyone else.
Measuring ROI without allocation buys nothing. Companies claiming board-defensible ROI they cannot allocate report the same board proof gap as companies claiming nothing at all: 56% and 54% respectively.
The 23% who have visibility-based ROI spend the same and absorb half the damage. 3% failed to secure an AI investment against 19% of everyone else. The difference is not discipline. It is knowing which spend is producing something.

Table Of Contents

Introduction
Act 1: The spend is going up
Act 2: Nobody can see where AI spend is going
Act 3: The reckoning
Act 4: Finance’s defense mechanism
Act 5: The way out is where, not how much
Conclusions
Methodology

Introduction

This report follows one sequence. Finance is told to spend on AI, and does. The spend outruns the ability to explain it. The bill lands, and with no visibility, the only lever left for managing the heightened spend is an imposed friction. Budgets get capped. Initiatives get killed. Hiring gets delayed. None of it addresses the underlying problem.

Meanwhile, 87% of companies are increasing AI budgets for FY2027. Yet 61% say they already cannot justify what they spend.

Hold that contradiction in mind as you read. Almost every finding below is a consequence of it. Nearly a quarter of the companies we surveyed live with that same contradiction yet suffer far fewer consequences, and we found what separates them from the rest.

What the 23% do differently

Before the evidence, we bring you the conclusion as it’s top of mind.

Fifty-nine of the 252 finance leaders surveyed, 23% of the sample, can do two things at once. They allocate AI spend systematically to a team, a business unit or finer. And they measure AI outcomes rigorously enough to say what the spend returned.

That combination is worth more than either half alone. Only 3% of them failed to secure an AI investment, compared with 19% of everyone else. 27% absorbed a growth-slowing consequence, compared with 47%. 24% came through the year with no visibility consequence at all, compared with 11%.

They did not buy that with restraint. 41% of them plan a 20%+ increase in FY2027 AI spend, compared with 38% of everyone else, a difference indistinguishable from noise. They spend like everyone else and absorb about half the damage.

Nor is measurement alone enough. Companies that measure AI outcomes rigorously but cannot allocate the spend behind it report the same board-proof gap as companies measuring nothing at all: 56% against 54%. ROI measurement tells you the program returned something. Allocation tells you which part of it did.

The rest of this report is the evidence. Five acts, thirteen findings, and one essential capability behind nearly all of it.

First of all: Where do you sit?

Before Act 1, we need something from you.

This report measures three capabilities, and one of them carries almost everything that follows: allocation. Allocation is the ability to break AI spend below a single lump-sum line on a bill. Your allocation capability affects how nearly every finding in this report reads.

Find your organization on the five rungs below, and remember that answer as you read.

Today, 54% of finance teams sit at tiers 1 or 2. They can see a total, or a rough split, but nothing underneath it. It is worth knowing what that means in money. Cross-referencing reported AI and cloud spend against the share attributable to AI, the median company in this survey is running somewhere between $1.1M and $2.7M a year in AI spend. The 135 companies at the two weakest allocation tiers are carrying at least $385M of it in combined annual AI spend as a number they cannot break down. They are not the small ones, either: their reported spend sits close to that of the companies at the the two strongest allocation tiers.

Of those 135, 56% say they’ll need to get above the two weakest tiers by FY2027. Across the whole sample, only 33% expect a rough category split or a single bill line to still be enough. Whether they get there is the question this report ends on.

We asked the same five-rung question about the other two capabilities. Forecasting: can you predict the total? Outcome measurement: can you show what it returned? Both look healthier than allocation today, and neither is where the pressure is.

Forecasting, rung by rung: (1) we don’t forecast AI spend at all, (2) we set a rough number and never revisit it, (3) we forecast and review AI spend periodically, (4) we a set cadence and hold owners to it, (5) we continuously re-forecast AI spend.

Outcome measurement, rung by rung: (1) we don’t measure return, (2) we track usage but not outcomes, (3) we have directional ROI that wouldn’t survive scrutiny, (4) we have ROI we could defend to a board, (5) we have initiative-level ROI on which we base funding reallocation.

Allocation is the only one of the three where FY2027 demands more than finance has. Forecasting is already 4 points past what teams say they will need. Outcome measurement is 20 points past it, which is not a typo: in aggregate, finance is planning to measure outcomes less closely next year than it does today. Finding 11 shows why.

Two other cuts recur in what follows. The first is who drives AI investment at your company: the CFO’s office, a CEO mandate, engineering, or a cross-functional group. That shapes how fast spend moves, and who notices when it does. The second is your board’s posture: invest aggressively, invest and prove it, or prove it first. That sets how much room you have before someone asks a hard question. Neither carries as much of this report as allocation, but both are worth knowing your answer to.

Allocation is the backbone of this report. It’s the only one of the three where FY2027 demands more than finance has today. Forecasting and outcome measurement are both already ahead of what teams say they’ll need. Keep your position on all three capabilities in mind as you read. What follows is not a description of the market. It is a read on where you stand in it.

Act 1

The spend is going up (and it has to)

Nothing in this survey suggests finance is debating whether to fund AI. 87% are increasing FY2027 budgets, and the pressure is arriving from above finance rather than inside it. This act is about where that pressure comes from, who grants the permission, and what is already being traded away to pay for it.

In short: AI spend keeps climbing right through the credibility problem. 91% of the finance leaders who can’t justify their current AI spend are increasing it anyway, and among everyone increasing, 45% are raising spend by more than 20%.

Finding 1

CEO-mandated AI programs run over the most. Overruns teach no one restraint.

CEO-mandated AI programs blow the budget by more than 20% at 19%, against 9% of everyone else.
Consequences from poor visibility hit 86% of companies, regardless of who drives the AI program.
When asked what they would change, 57% of over-budget teams would set a larger budget for FY2027.

CEO-mandated AI programs blow their budget by more than 20% at 19%, against 9% of everyone else. That’s the “act now, reconcile later” mandate showing up in the numbers: a directive to move, with the budget conversation scheduled for whenever finance catches up.

Finance holds the bag either way: consequences hit 86% of companies regardless of who’s driving.

Ask over-budget teams what they would change in hindsight, and the instinct is not to rein in their spending. 57% say they would have set a larger budget from the start. Stronger ROI cases came in at 35%, more contingency at 41%.

Overspending doesn’t teach restraint. It teaches finance to ask for more next time.

Q24×Q11 who’s driving; Q16; Q15 · n=128

Finding 2

The boards writing blank checks are attached to the companies least able to defend the spend.

77% of companies with invest-aggressively boards already spend more on AI than they can justify.
55% of companies at every other board posture say the same.
61% of companies with invest-aggressively boards approved AI spend without reliable projections, against 39% at every other posture.
24% of companies with invest-aggressively boards could not produce a credible ROI number when the board asked, compared with 40% at every other posture.

The mechanism is visible one step earlier. The pressure to commit arrives without a matching demand for proof.

That is a sequencing problem rather than a competence one. These companies are not asked for a projection and failing to produce one. They are asked to move before a projection exists, and the approval clears anyway. The spend is committed on the strength of the mandate, not the strength of the case.

The demand does not arrive afterward either, and the last two bullets run in opposite directions for that reason. Being caught without a credible ROI number requires a board that asked for one. Fewer of these companies were caught, and since most of them say they cannot justify the spend, the likelier reading is not that their answers were better but that the question came less often. No projection is required to release the money and none is demanded once it is gone, so nothing forces a reckoning.

Q27×Q41 board posture · n 66/186; ×Q16

Finding 3

AI is being funded out of R&D, not out of headcount.

45% of companies say AI spend displaced SaaS licenses in FY2026. For FY2027, 56% expect it to.
41% say it displaced R&D or new-product investment. For FY2027, 47%.
38% say it displaced headcount or hiring. For FY2027, 33%. That’s the only line in decline.
46% say it displaced other infrastructure. For FY2027, 45%.
12% said AI spend was net-new budget. For FY2027, 10%.

The survey asks what AI spend displaced in the original plan, and what it is expected to displace in FY2027. This is budget displacement, not net-new money: every figure below is a share of companies naming that line as something AI spend came out of.

Infrastructure and sales barely move, so the whole movement sits in two lines: tools and innovation.

Q12×Q39 · n=252, multi-select

The substitution everyone predicted, buy AI and hire fewer people, is the only trade here that is receding. What is growing is the trade against the innovation budget. Companies are funding AI out of R&D, which means the innovation budget is where AI is coming from rather than an addition alongside it.

Whether that money still buys innovation, just delivered through AI rather than conventional methods, is a step the survey does not confirm. It records where the budget came from, not what the AI was used for. Either way, the new line is harder to defend. A canceled tool leaves an invoice and a renewal date. The value of the AI that replaced it is not yet known to most finance teams.

Which brings us to Act 2: where is the AI spend going and how defensible is it?

Act 2

Nobody can see where AI spend is going

If finance could see where the money went, the rest of this report would be unnecessary. Allocation, the ability to break AI spend below a single lump sum, is the weakest capability finance has right now, and the closer a function sits to the budget, the worse it tends to be. When the spend can’t be seen, it starts to feel dangerous.

In short: 54% can see AI spend only as a rough category split, or as a single line on a bill. It is also the only capability where FY2027 ambition outruns today’s state, climbing from 30% to 42%. Meanwhile, forecasting is already slightly ahead of where finance says it needs to be, at 45% today against 41% needed.

Companies can forecast and measure general outcomes. They cannot drill into the detail underneath either one.

Finding 4

The office that owns the AI budget is the one that can see it least clearly.

65% of CFO-led AI programs sit at the two weakest allocation tiers, compared with 43% of engineering-led programs.
53% of CFO-led programs have finance tracking the AI spend, compared with 16% where engineering does. Allocation depends on tagging and cost-center mapping inside the systems engineering runs, which is why whoever runs the systems ends up able to see the spend.
48% of engineering-led programs have engineering tracking the spend, compared with 24% where finance does.
52% of CFO-led programs forecast strongly, compared with 43% of engineering-led ones, and 59% measure outcomes strongly, compared with 64%. Neither difference is statistically significant. Allocation is the only capability where the two genuinely diverge.

The gap is not rigor and it is not speed. CFO-led programs see their spend data as quickly as engineering-led ones do.

Finance can own the number, forecast it, and build a board story around it from the top down. It cannot tag a workload it does not run.

What the dataset shows is that owning the initiative and being able to see what it costs are two separate purchases. Most CFO-led programs have only made the first.

What neither side has is a shared view, where finance owns the number and engineering owns the systems that could explain it. Both are looking at AI spend. Neither is looking at the same version of it.

Q24×Q19 who’s driving × allocation maturity · n=96/42; ×Q20, Q21, Q23, Q25

Finding 5

The function most likely to own AI spend is also the one that finds it hardest to defend.

62% of CFO-office-driven programs say AI spend is harder to defend than any other budget line.
40% of all other programs say the same.
Across all companies, 49% agree AI spend is the hardest line to defend.

62% of CFO-office-driven programs agree that AI spend is harder to defend than any other line in the budget, against 40% of everyone else, even though real agreement runs across the board at 49% overall.

It’s the same group Finding 4 already flagged as most likely to be stuck at the two weakest allocation tiers (65%): CFO-led programs own AI spend, can’t see it clearly, and suffer the defense burden most intensely of anyone.

Q24×Q44 who’s driving · n=96/156

Finding 6

The earlier finance gets a seat, the less it gets blindsided.

66% of finance leaders without early or formal authority get caught without ROI proof.
Only 27% of active partners and approval-gate finance face that same bind.
71% of everyone without concrete ROI proof sits in a role without early authority.

Some finance leaders are consulted occasionally. Some review only the major initiatives. Some find out after the money is gone. 66% of these people get caught by the board without ROI proof, against 27% for active partners and approval-gate finance. And 71% of everyone in that bind sits in a role without early or formal authority.

Allocation depends on systems finance doesn’t run. This one is just a question of when finance is in the room.

Q26×Q42 · n=127/125

Finding 7

Risk isn’t an attitude finance brings to AI. It’s the prevailing attitude when spend can’t be seen.

25% of companies at the two weakest allocation tiers frame AI as a risk, against 9% of companies in the three stronger tiers.
43% of respondents with strong outcome measurement classify AI spend as an opportunity; just 13% classify it as a risk.
25% of respondents with weak outcome measurement call AI a risk, and 25% call it an opportunity.
58% of risk-leaning respondents plan a 20%+ FY2027 increase, compared with 34% of opportunity-leaning ones.

Companies stuck at the two weakest allocation tiers frame AI spend as more risk than opportunity at 25%, against 9% of everyone else. They are also far less likely to call it an opportunity, 26% against 45%.

Outcome measurement shows the same pattern from the other end. Where measurement is locked in at board-defensible level or better, opportunity-leaning respondents outnumber risk-leaning ones by more than three to one, 43% against 13%. Where it is not, the two views sit dead even at 25% each.

When read that way, the risk framing looks less like temperament and more like a symptom that recedes as visibility gets built. This is a two-way cut, so it shows the two traveling together rather than one causing the other.

The risk framing doesn’t come with restraint. Respondents who see AI as a risk plan a 20%+ FY2027 increase at 58% vs. 34% of opportunity-leaning ones, the widest spending gap of any segment in this survey.

Q19×Q37 allocation maturity · n=135/117; Q21×Q37 · n=143/109; Q37×Q31 · n=45/88

Now, let’s look at how finance responds when the bill comes.

Act 3

The reckoning

This is the moment where the bill shows up on finance’s desk. There are two prevailing responses: 1. Teams quietly spend less than they planned; 2. Finance teams start tightening the leash on FY2027 spending, not by cutting the AI budget but by making it harder to draw down. Both are reactions to the same AI spend blind spot.

In short: Nearly nine in ten finance leaders got burned by poor AI visibility this year, and 38% hit the ceiling of what this survey let them report, in other words, the question allowed three answers and they selected all three. The damage isn’t evenly spread, either. Companies at the two weakest allocation tiers cite a stalled or rejected investment ask at 21%, more than double the 9% at the three stronger tiers.

Finding 8

Poor visibility tightens the purse strings.

66% of teams that came in under budget have weak allocation, compared with 57% of teams on plan and 46% of teams over budget.
Companies weak on all three maturity dimensions (allocation, forecasting, outcomes) represent just 6% of over-budget teams but 17% of under-budget ones.
Weak-across-the-board companies are 3 times more common among under-spenders than over-spenders.

Companies with the weakest allocation maturity are the least likely to be over budget, not the most: 46% of over-budget teams have weak allocation, rising to 57% on-plan and 66% under-budget.

It’s not just allocation: companies sitting in the two weakest tiers on all three maturity dimensions at once (allocation, forecasting, and outcome measurement) make up just 6% of over-budget teams but 17% of under-budget ones.

Q19×Q11 allocation maturity; Q19×Q20×Q21×Q11 · n=128/63/59

Finding 9

The bet stays the same size. The controls around it get tighter.

93% of those who can’t justify current AI spend are restructuring FY2027 budgeting, against 78% of those who can.
4% of companies that can’t justify their current spend are keeping the same process and changing only the number, compared with 12% of those who can.
3% of that same group are making no change at all, compared with 10% of those who can justify their spend.
Big bets look similar: 10% plan a 50%+ FY2027 jump, against 8% of everyone else.

The 61% who say they are already spending more on AI than they can justify are markedly more likely to be rebuilding how they budget it: 93% are making some structural change to the FY2027 process, against 78% of everyone else.

Only 4% of that group plan to carry their existing process into FY2027 and change nothing but the number, against 12% of everyone else. Only 3% plan no change at all, against 10%.

The size of the bet is not what shifts. 10% of the group plan a 50%+ FY2027 jump, against 8% of everyone else. Companies with strong outcome measurement plan big increases only slightly more often than companies with weak measurement, 41% against 36%.

That is neither recklessness nor confidence. It is two rational moves in one budget cycle: keep the number where anyone else would put it, and build more machinery around it.

Q41×Q40 · n=154/98; Q41×Q31; Q21×Q31 · n=143/109

What does that ‘control’ look like? For that, we move to Act 4.

Act 4

Finance’s defense mechanism

Faced with a number they can’t defend, most finance teams reach for the one lever that works without visibility: friction. Approval gates don’t require knowing where the money goes. They only require making it harder to spend without clear justification for it. This act is about what that choice costs, and why the teams making it don’t expect to stop.

In short: FY2027 budgeting is getting real teeth. 87% of finance teams are making a structural change to how they budget AI next year, split across five distinct moves: new approval gates (27%), a formal budget line (19%), tying funding to demonstrated ROI (18%), shifting the spending mix (16%), and a dedicated contingency fund (6%). Whatever finance is doing next year, it isn’t just adjusting the budget itself.

Finding 10

Teams that can’t see the spend reach for control. Teams that can reach for proof.

New approval gates are the market’s most-chosen FY2027 change at 27%, ahead of a formal budget line (19%) and tying funding to demonstrated ROI (18%).
Among teams reaching for either gates or ROI, 71% of weak-allocation teams pick the gate. Everyone else splits roughly evenly, 48% to 52%.
Weak-allocation teams tie funding to demonstrated ROI at 13%, against 24% of everyone else.
Weak-allocation teams’ planned FY2027 increases run no lower than anyone else’s, so this is not restraint.

Gates are the single most common answer to managing AI spend in FY2027. 27% of finance teams name new approval gates as their primary budgeting change, ahead of a formal budget line at 19% and tying funding to demonstrated ROI at 18%.

Two of those three are the same decision seen from opposite ends. A gate controls spend you cannot evaluate. Tying funding to ROI evaluates it. Among the 115 companies that reached for approval gates or ROI, allocation decides which: 71% of weak-allocation teams took the gate, against 48% of those with stronger allocation.

The raw rates say the same thing. Weak-allocation teams tie funding to demonstrated ROI at 13%, against 24% of stronger allocators. And none of this is about holding the budget down, because their planned FY2027 increases run no lower than anyone else’s.

Oversight is not a step toward visibility. It is what you do instead.

Q40 · n=252; Q19×Q40 gate-or-ROI subgroup · n=61/54; Q19×Q40 · n=135/117

Finding 11

The companies reaching for approval gates don’t have a plan to remove the gates later.

The groups managing AI spend with approval gates versus ROI proof plan identical FY2027 spend increases, at 39% each.
62% of the approval-gate group sit at the two weakest allocation tiers, against 39% of the ROI-proof group.
Only 28% of the approval-gate group say FY2027 needs strong outcome measurement, against 61% of the ROI-proof group.
78% of gate-adders say better data would change FY2027 decisions, but only 28% expect to need better data.

Finance leaders whose primary FY2027 budgeting change is adding approval gates and those tying funding to demonstrated ROI look identical on paper. Strong forecasting today: both 48%. Strong outcome measurement: both 52%. 20%+ increase in FY2027 spend: both 39%.

The difference doesn’t stem from their appetite or general maturity. It lies in their allocation tier: 62% of those planning approval gates sit in the weaker allocation tiers versus 39% of the ROI group. Both can forecast the total and prove the return. One of them cannot break the number down.

The gate group has killed or paused an AI initiative at 20%, against 7% of those who can tie funding to ROI. And they came in under budget at 30% against 11%, which reads as hesitation rather than discipline.

Across all three, 55% of the gate-adder group experienced at least one growth-slowing consequence in 2026 against just 28% of the ROI group.

Asked what FY2027 will require, 59% of the ROI group said they want strong allocation against 36% of the gate group, and 61% want strong outcome measurement against 28%. That is the widest gap between where two groups sit and where they say they need to be.

Gates are what finance reaches for when it cannot see where money’s going. The wanting is not the problem. 78% of the gate group say better data would change specific FY2027 decisions, against 85% of the ROI group, a difference indistinguishable from noise.

Yet only 28% of the gate group expect to need strong outcome measurement, against 61% of the ROI group. Among the gate-adders who say better data would change their decisions, 72% do not expect FY2027 to require it. The gap is between appetite and expectation, not between appetite and interest. (Note: the question asking “Which best describes how your organization’s AI budgeting process is changing for FY2027?” is single-select, so this compares primary levers, not every control in place).

Q40×Q19; Q40×Q20; Q40×Q21; Q40×Q16; Q40×Q11; Q40×Q31; Q40×Q34; Q40×Q36; Q40×Q43 · n=69/46

So if you shouldn’t cap spend, what can you do? Let’s move to Act 5.

Act 5

The way out is where, not how much

The report so far describes a choice between two bad options: cap the spend, or let it run. It is a false choice, which finance leaders already sense. What separates the teams that get out of it is knowing which spending is producing something, so they can put more behind it. Not “How much are we spending?”, but “Which part of it is working?”.

In short: Almost everyone wants better data. Far fewer can produce it. 74% say that linking spend to outcomes would change specific FY2027 decisions, while only 12% can break their AI spend down finely enough to act on it inside the quarter.

That gap is the opportunity for finance leaders. Doing half the job doesn’t help: companies that measure ROI without being able to allocate the spend look no better to their board than companies measuring nothing at all. The 23% from the opening are where this lands, and Finding 13 is the full account.

Finding 12

Measuring ROI without allocating spend doesn’t survive contact with the board.

50% of companies that can’t allocate their AI bill to specific projects still claim board-defensible ROI or better.
17% claim initiative-level ROI they reallocate on, which weak allocation makes impossible.
56% of the companies claiming ROI they cannot allocate say the board asks for ROI proof they cannot provide.
37% of the companies claiming ROI they cannot allocate could not produce a credible ROI number.
47% of the whole sample reports that board gap; only the allocation-plus-ROI group breaks away at 32%.

Half of the companies that can only see their AI spend as one bill or a rough category still say they can produce ROI a board would accept. One in six goes further and claims the top rung: ROI per initiative, which is precise enough to shift money between them. However, the survey’s own definitions make that impossible without allocation. How can you report return on something you can’t separate out in the first place?

The likeliest explanation is distance from the mechanics. The next paragraph supports that, since the group doing allocation by hand rates itself lowest of anyone.

Companies allocating by hand, team by team, are the first group doing any real work at it. They also rate themselves lowest of all, with just 38% of that group reporting board-defensible ROI or better, compared with 44% and 54% of the two groups doing less allocation work. At the two strongest tiers, those figures climb to 73% and 87%

Finance leaders who have actually tried allocation are hardest on themselves, and that dip is worth more than it looks. A self-rated ladder that rose cleanly from rung to rung would tell you very little. This one bends where the work starts.

Q19×Q21 · n=48/87/42/45/30

So does the claim help? Across the whole sample, 47% say their board asks for ROI proof they can’t provide. Of the companies that claim board-defensible ROI but can only see a total or a rough split, 56% report that same gap; which is no better than average. The companies who can do both run 32%. And set against companies at the same allocation tiers who claim no ROI, the claim-only group is indistinguishable: 56% to 54%. Being able to allocate is what moves the number.

Q19×Q21×Q42; Q16; Q44 · n=68/67/59, baseline n=252

The claim isn’t worthless. The companies claiming ROI they cannot allocate killed or paused fewer initiatives than their weak-allocation peers who claim nothing, 9% against 21%, and called AI a risk less often, 18% against 33%. Measuring return does steady the nerves. What it doesn’t buy is the thing they say they have. ROI measurement tells you the program returned something. Allocation tells you which part of it did.

And now we move to our final finding: the real value in having both allocation capabilities and ROI measurement together.

Finding 13

**Teams that can allocate AI spend and prove its return rarely get told no. Only 3% failed to secure an AI investment, against 19% of everyone else.**

23% of finance leaders can both allocate AI spend systematically and produce board-defensible ROI.
Only 3% of that group failed to secure an AI investment, against 19% of everyone else.
24% finished the year with no visibility consequences at all, against 11% of everyone else.
They plan the same FY2027 increases as everyone else, 41% against 38%.
Here is the group from the opening, with the twelve findings behind it. 59 companies, 23% of the sample, can allocate AI spend systematically and measure its return rigorously.

That combination is what the rest of this report keeps pointing at. Just 3% of those who can do both failed to secure an AI investment, against 19% of everyone else. Just 27% absorbed at least one growth-slowing consequence (a killed initiative, a delay in hiring, a blocked investment ask), vs. 47%. And 24% came through the year with no visibility consequence at all, against 11% of the rest.

They are not spending less on AI. 41% plan a 20%+ FY2027 increase against 38% of everyone else, a difference indistinguishable from noise. But they frame AI spend as more risk than opportunity at just 7% against 21%, and call it an opportunity at 51% against 30%, the same relationship Finding 7 traces from the other direction.

This is a two-way cut of self-reported capability against self-reported outcomes, so it shows the two traveling together, not one causing the other.

Q19×Q21 · n=59/193; ×Q16, Q31, Q37, Q40, Q42, Q43

Where this leaves FY2027

FY2027 isn’t the year AI spending slows down. It’s the year the excuse for not measuring it runs out.

87% are increasing spend. 61% can’t justify what they’re already spending. Those are mostly the same people, and when those two facts collide the only lever that works without visibility is friction. Teams whose primary FY2027 change is an approval gate killed or paused an initiative at 20% against 7%, and delayed hiring at 33% against 19% of everyone else, while planning the same increases, 39% against 39%. They absorb more damage to control the same number.

The way out isn’t restraint. The 23% who can allocate AI spend below a lump sum and measure its return rigorously plan the same FY2027 increases as everyone else, and absorb roughly half the consequences. The difference isn’t discipline. It’s knowing which spend is producing something.

Diagnose allocation specifically, not “maturity” in general. Across the two cohorts compared most closely here, forecasting and outcome measurement were level: 48% and 52% strong in both. Allocation was the only capability that differed. Deliverable: your organization scored 1–5 on allocation, forecasting and outcome measurement separately, with the weakest one named.
Name the lever you are about to pull. Gates work without visibility, which is exactly why 27% of the market is reaching for them. They become a permanent tax when nothing is funded to replace them. Deliverable: a written answer to what would have to be true for you not to need these gates in FY2028, and whether anything in the budget funds it.
Stop waiting for appetite. It is already there. 74% say better data would change specific FY2027 decisions, and 61% of them don’t expect to need stronger outcome measurement next year. That is a resourcing gap, not a conviction gap. Deliverable: one named owner and a funded line for AI cost allocation, not a working group.
Aim at both capabilities, because one alone doesn’t pay. The outcome gap only appears when allocation and outcome measurement are both in place. Either one alone leaves you defending a total you can’t decompose, or seeing parts you can’t justify. Deliverable: a single view answering both questions for any AI line, where did this go and what did it return, readable by whoever owns the budget and whoever runs the systems.
Put both functions on the same numbers, not on two reports. Allocation lives inside the systems engineering runs, which is why 65% of CFO-led programs sit at the two weakest allocation tiers against 43% of engineering-led ones. Reassigning the budget doesn’t move the capability, and neither does asking each side for its own report. The gap closes when both are reading one view of the same spend.

The question the next budget cycle asks won’t be how much you spent on AI. It will be which part of it you would defend.

Methodology

Why CloudZero asked, and how to check us

CloudZero designs and provides AI spend allocation software, and this report concludes that allocation is the constraint. So here is who did what: we wrote the survey, and contracted Centiment to recruit a screened panel and collect the responses. No respondent is a CloudZero customer, and none opted in out of an existing interest in the subject. We did the analysis.

We also wrote the allocation ladder the findings lean on. It is not a framing invented in isolation: KPMG’s Global AI Pulse, 2,145 senior leaders across 20 markets, found those with strong cost visibility five times likelier to have established ROI, 15% against 3%. KPMG measured cost visibility as a single idea. This survey separates allocation from forecasting and outcome measurement, and finds that measuring return only counts when you can allocate the spend behind it.

The full instrument and every response distribution are published below. Everything you would need to disagree with us is in this document.

About this research

CloudZero surveyed 252 senior finance leaders at US companies between August and September 2026. Every respondent either owns and approves their organization’s AI budget (87%) or materially influences it (13%). No responses were excluded after collection; all percentages are calculated on the full sample of 252 unless a finding states otherwise.

Who responded

79% are CFOs and 21% are VPs of Finance. By headcount, 45% work at companies of 250–999 employees, 34% at 1,000–4,999, 15% at 5,000–9,999 and 6% at 10,000 or more. By revenue, 56% sit between $100M and $999M in ARR, with 9% above $1B and 7% below $50M. The largest industry groups are software and technology (26%), financial services and insurance (22%), manufacturing and industrial (17%) and retail and e-commerce (17%). 41% are privately held, 33% are PE- or VC-backed, and 22% are publicly traded. Reported annual cloud and AI spend runs from under $1M (8%) to $50M or more (6%), clustering between $5M and $25M.

Where this report expresses AI spend in dollars, the figure is derived, not asked. Respondents reported combined annual AI and cloud spend in bands (Q3) and the share of cloud spend attributable to AI in bands (Q28). Multiplying the two per respondent gives an estimated annual AI spend, calculated here two ways: using band midpoints, and using the bottom of every band. The conservative figure is the second. Three respondents are excluded for answering “prefer not to say” or “can’t isolate.” These are estimates from banded self-reports and should be read as an order of magnitude, not a total.

How AI maturity was measured

Three capabilities are measured separately on five-point ladders, and the distinction between them carries most of this report. Allocation is the ability to break AI spend below a single bill line, from “we only know our total AI spend on the bill” to “we allocate to the customer, product or cost-center level and use it for in-quarter decisions.” Forecasting runs from not forecasting AI spend at all to continuous re-forecasting. Outcome measurement runs from not measuring return to initiative-level ROI that funding is reallocated on. Where this report describes a company as “weak” on a dimension it means the bottom two rungs; “strong” means the top two.

These are self-ratings on descriptive ladders, not audited capability assessments. The top two rungs on each are the most likely to be over-claimed, since they describe outcomes rather than mechanisms. Finding 12 quantifies how far that goes on outcome measurement specifically: half of the companies who say they can only see a total still claim board-defensible ROI or better.

How the numbers were tested

Comparisons between groups are two-proportion z-tests. Differences are described as real only where p < 0.05; where a gap fails that bar it is either reported as not significant or left out. No finding rests on a group smaller than 30 respondents. No cut below 30 respondents is reported. At n=252 the margin of error on a whole-sample proportion is approximately ±6.2 percentage points at 95% confidence; subgroup margins are wider.

What this survey can and cannot show

Every variable here is self-reported by one person per organization in a single sitting. That is standard for executive research and it carries a known limitation: a respondent who rates their capability highly may also rate their year favorably, so some part of the gap between high-capability and low-capability groups reflects a consistent point of view rather than a difference in what happened. This is common-method bias, and survey data alone cannot rule it out. Where this report compares a self-rated capability against a self-reported outcome, read the result as two things traveling together.

One piece of evidence cuts the other way. If the capability ladders were simply self-flattery, self-ratings would rise cleanly with position on them. They do not. Finding 12 shows the group doing allocation manually rating itself below two groups doing less.

How to read the source tags

Each finding closes with the survey questions it draws on. A multiplication sign indicates a cross-tabulation, so Q19×Q21 means allocation maturity crossed against outcome measurement, and the figures after n= give the size of each group being compared, in the order they appear in the text.