Adoption happened fast. Most teams are running Copilot and other AI tools in environments they still can't fully see. Read more about that in CH 1Confidence ≠ control→.

The permissions sprawl that makes AI risky is the same sprawl that makes it expensive. And the governance work that contains the risk is the same work that makes the spend explainable. IT teams tend to experience these as separate conversations, one with security and one with finance. But the data suggests they're the same conversation with two audiences.

IT leaders are already finding that AI costs are taking up a meaningful share of the IT budget, and most teams can't yet trace them well enough to say whether they're working. That's not a Copilot problem. It's a visibility problem in a trench coat.

This chapter of The State of Microsoft 365 looks at what AI is actually costing, why those costs are hard to pin down, how governance work changes investment confidence, and where AI spend sits next to the other things quietly draining the budget, like storage growth, unused licenses, duplicate tenants nobody consolidated after the acquisition.

Let's dig into the insights we collected from 1,700+ IT pros and expert advice from Microsoft MVPs to explain why these costs are so hard to trace and what to do about it for a quick win, a medium lift, and a strategic investment.

What AI is costing

AI costs show up in licensing, in storage, in compute, and in the quiet expense of creating content and running tools nobody is tracking. Start with the size of it.

Asked what percentage of their total IT budget has been affected by AI infrastructure and licensing over the past 12 months, the answers cluster firmly in the middle of the range. 59% put it between 6% and 20%. Another 22% are already past a fifth of the annual IT budget. Only 4% report no measurable impact at all. For most teams this isn't a pilot-sized number. It's a line finance has noticed. AI governance survey

“

AI is currently coming out of the IT budget, and IT is still treated as a cost center rather than an innovation center. The business asks for AI. IT pays for it.

Drew Madelung Microsoft MVP

And the bill scales with how far into deployment you are.

What AI takes out of the IT budget, by how far Copilot has been deployed

Both thresholds shown together: the earlier table reported one and the prose the other. * base under 75 — read as directional.

Across all organizations, 22% say AI has taken more than a fifth of the budget. Among teams that have fully deployed Copilot, that rises to 32%. So if you are mid-deployment, your current costs are a preview rather than a total.

And if you have not deployed at all, your floor still is not zero. A fifth of that group reports AI taking more than a tenth of the budget — because readiness work, storage growth and other AI tools land on the same budget line whether or not Copilot has been switched on. Note the one place the table stops being tidy: at the higher threshold, teams with no Copilot deployment (12%) edge past teams that are partway through one (9%). The base is small and directional, but the direction is real. Not deploying is not the same as not paying.

None of which means the answer is to deploy less. It means being deliberate about who gets a seat.

AI licensing tops the cost list

The operations survey asked a different population about cost pressures across the whole Microsoft 365 estate. AI licensing did not replace the familiar costs of storage, duplicate tenants and on-prem infrastructure. It passed them.

Storage bloat is a structural problem teams have managed for years. It accumulates quietly, it's nobody's fault in particular, and it has had years to become the thing everyone complains about. Copilot licensing is a line item that has existed for a year or two at most organizations, and it's already top of the list. On a question that only let teams name two things, 41% spent one of their two on it, ahead of storage, ahead of the tenants left over from the last acquisition, ahead of every license nobody is using. IT operations survey

Five hundred employees is where AI licensing bites

Respondents could name up to two cost challenges, so this is 42% of a two-item budget across the whole sample.

Five hundred employees is where this stops being a line item. Below 250, 30% of teams name AI licensing as a top cost challenge. Through 499 it creeps to 35%. At 500 it jumps eight points, and then it just sits there — everything above 1,000 lands within half a point of 45%. Whatever happens at that size happens fast, and it does not get worse.

Third place on the list is a different kind of answer. A quarter of teams say a lack of visibility into what they are spending versus what they are actually using is one of their top challenges.

That isn't a cost. It's not knowing what the costs are.

The AI cost visibility problem

Nobody we surveyed is confused about whether AI is costing them. Even the teams that have not deployed Copilot have AI taking a chunk of the IT budget. The problem is that most teams cannot connect the cost to specific outcomes.

Where the concern shows up in Copilot specifically

We asked IT leaders what worries them most about Copilot. Data quality and access management came back effectively tied at the top — 463 respondents against 461, out of 943.

Realizing the ROI of Copilot is a top concern for 34% of IT leaders, a clear six points above cost control. Teams are more worried about proving Copilot is worth it than about what it costs. A price is a number you can look up. A return is something you have to demonstrate, and demonstrating it means tracing spend to outcomes across an environment most teams can't fully see.

The two governance concerns at the top of this list are the subject of Chapter 1, and they're worth seeing again here CH 1Confidence ≠ control→. Because the thing teams say is hardest about proving AI's value isn't the pricing model. It's visibility, and visibility is a governance capability.

Cost control falling while AI licensing tops the Microsoft 365 cost list isn't a contradiction. One question asks what is expensive across Microsoft 365; the other asks what worries you about Copilot. A line item can be the biggest on your bill without being the thing that keeps you up.

Read the five slopes together and the shape is the point. Four fall. Only lack of internal expertise on AI governance rises, the one worry a year of Copilot deployment was supposed to fix.

Governing AI agents has no 2025 line above it because it wasn't on last year's list at all. In its first year of being asked, one in five named it.

What makes AI ROI hard to measure

Ask what makes AI return on investment hard to measure and the list splits in two. Cost visibility, governance complexity and unreliable data lead. Then a real gap, and below it the answers you would expect to top a question like this: evaluation windows that are too short, goals that were never clear, benefits too diffuse to attribute.

Why AI ROI is hard to prove

We asked What are the biggest challenges in measuring AI ROI today? Select all that apply.

Cost visibility and governance complexity top the list.

That ordering is a finding in itself. The classic reasons ROI is hard to prove—fuzzy objectives, not enough time—are the ones teams named least. What they named instead were three versions of the same problem: they can't see clearly enough to do the arithmetic.

The overlap makes it sharper.

Almost nobody blames the goals alone

88% named at least one of the three clarity problems — cost visibility, governance complexity, unreliable data. A third named nothing else, three times the number who blamed goals or timelines on their own.

88% of teams named at least one of the three clarity problems. A third named nothing else at all—no complaint about goals, no complaint about timelines—which is three times the number who blamed goals and timelines on their own. Now set that against how the same population describes its own AI strategy. AI governance survey

Clear on the goal. Blind on the cost.

Three in four say their AI business goals are well or extremely well defined. Of those same teams, more than half name cost visibility as one of the biggest things blocking AI ROI.

That's the sharpest juxtaposition in this chapter, and it's worth stating plainly. Three quarters of organizations know what they want AI to achieve. More than half of those same teams can't see what it costs. The bottleneck isn't strategic clarity. It's visibility.

Permissions, content sprawl, lifecycle, ownership—the reason the numbers won't reconcile is that the environment underneath them was never mapped. That's the uncomfortable part of this chapter. The cost question can't be answered without going back to the basics.

The teams watching most closely report the biggest visibility problem

We expected teams doing more monitoring to find cost visibility easier. They don't. The rate is the same across every group that monitors at all.

Naming cost visibility a top ROI blocker, by monitoring posture

The two smallest groups are directional given their base sizes, but the direction is the opposite of the intuitive one. Teams with no monitoring at all are the least likely to say cost visibility is a problem.

This is the third time this pattern has appeared in the data. Chapter 1 found that concern about AI reaching unreviewed content rose with monitoring maturity, and that teams which had done cleanup reported more incidents because they could see them CH 1Confidence ≠ control→. Here the same shape shows up in cost. The likeliest explanation isn't that these teams' costs are clearer. It's that nobody has looked. Reporting a problem requires noticing it first, so if anything, the 51% headline understates how many organizations have a cost visibility problem, because the teams least equipped to detect one are also the least likely to report one.

What teams are doing about cost today

Knowing the bill is hard to read hasn't stopped teams from trying. We asked how IT pros have approached Microsoft 365 costs—not just AI costs—over the past twelve months.

Half of organizations are actively working the problem. But a third are watching the number rise without acting on it, and one in nine has seen costs increase and done nothing yet. The gap between monitoring and optimizing is the same gap that runs through the governance chapter between periodic review and continuous control CH 03 Governance. Watching a number isn't the same as having a lever attached to it. IT operations survey

Governance is the key to AI ROI

So far: AI is taking a real share of the budget, it tops the list of cost challenges in Microsoft 365, and proving ROI is difficult. So we asked how confident organizations are that a measurable return is coming in the next eighteen months, and whether AI-related governance work changes that confidence.

Who is most confident in AI ROI

86% expect measurable ROI from Microsoft 365 AI initiatives within 18 months. Only 39% are very confident of it, and 13% are uncertain or not confident. Goals are clear and expectation is high. Conviction is what's thin, and what stands between expectation and conviction is, by the sample's own account, the ability to see the costs and manage the governance complexity.

Which raises the question worth chasing through the rest of this chapter. Who are the 39%, and what makes them different? The first answer runs against the intuitive assumption that bigger bills make people nervous.

The bigger the AI bill, the surer they are of the return

From 21% at the bottom of the range to 77% at the top, in a clean line. The teams under the most cost pressure are the most convinced it's worth it.

Confidence rises with spend, in a clean line, from 21% at the bottom to 77% at the top. The two extreme bands have small bases and should be read as directional, but the gradient across the well-populated middle is unambiguous. There are at least two readings and the data doesn't separate them. Organizations spending more may be seeing more, having pushed past pilot scale into workloads where value is measurable. Or conviction may drive the spending rather than follow it. Most likely both.

None of this means cost anxiety is misplaced. In the operations study AI licensing is the most-cited cost challenge in Microsoft 365, and cost control is a top Copilot concern for more than a quarter of teams. The pressure is real. What this gradient adds is that the teams under the most pressure are the most convinced it's worthwhile.

How governance work shapes AI investment confidence

Asked to what extent AI-related governance activities—permission audits, cleanup, lifecycle management—affect their confidence to invest further in AI, 78% said those activities have a significant or moderate impact. 32% said significant. Fewer than one in twenty say they make no difference. As beliefs go, that's close to consensus. AI governance survey

Governance isn't a tax on the AI budget. It's a precondition of the AI investment case. That reframes the internal argument. The usual framing puts governance and AI in competition for the same money, with governance cast as the responsible-but-boring option. The data says the opposite: the organizations that can prove they have some control over their tenants are the ones that can justify the next batch of AI investment, and the ones that can't are the ones whose AI programs stall at the business case.

Beliefs are cheap, though, and a survey is a comfortable place to hold one. The more useful question is whether the teams who actually did the work feel differently from the teams who didn't.

Three different measures of "did you do the work". Same slope every time.

The first two panels report the share very confident of a measurable AI return within 18 months. The third is a different study measuring a different outcome—the share running proactive governance monitoring—which is exactly why it's worth reading alongside. Two independent studies, two different framings, one finding: the organizations that can see and control their environment are the organizations that can see and control what it costs.

Confidence drops as you move down each ladder, but the steps aren't even. On maturity, the sharpest fall comes between governance that's consistently enforced and governance that exists as process without being operationalized, a 14-point drop, against barely a point between the two rungs below it. On cleanup the effect is stronger still: teams that cleaned up estate-wide before scaling AI are more than twice as likely to be very confident of a return as those that cleaned up only in places, and more than five times as likely as those that only have a plan.

Before anyone builds a business case on that gradient, be clear about what it proves. These are correlations. The data can't tell us whether cleanup produces returns, whether the organizations disciplined enough to run an estate-wide cleanup are also the ones disciplined enough to measure and realize returns, or some mixture. What it does establish is that governance maturity and confidence in AI value move together, consistently, across two independent measures. Nothing here supports treating them as competing priorities.

There is a reason only 51% of teams managed an org-wide cleanup. It's hard to do, especially with scripting or native tooling alone.

The geography of AI spend and return

The budget picture isn't evenly distributed across countries, and it doesn't map neatly onto confidence.

AI budget share and ROI conviction, by country

The shaded corridor is one standard deviation either side of the fitted trend. Four countries sit inside it — their conviction is about what their spend predicts. Germany and Netherlands sit below it, and United States above.

Spend and conviction broadly travel together, which is the corridor. What the corridor is for is the three countries that leave it. The US isn't simply high on both. Its conviction runs further ahead than even its spend predicts. Germany goes the other way, and hardest: the second-highest AI budget share in the sample paired with one of the lowest confidence levels, 55% against 26%. The Netherlands sits in the same quadrant on a smaller base.

High investment, low certainty of return. The rest of the German data explains it.

Germany against the sample average

One bar above the line, four below. Spending more and having built less.

German organizations sit above the sample average on AI spend and below it on every governance measure we asked about. Their confidence lands where those governance numbers would predict, not where their spending would. That's the argument of this whole chapter in miniature: what a team has built shapes its conviction more reliably than what it has spent.

Is this just national modesty? Some of it, probably. German respondents rate themselves lower on every confidence scale in the study, which is exactly what a conservative answering style looks like. Here is why it doesn't explain the gap. Three of the five measures above aren't opinions. Budget share, completed cleanup and monitoring approach are things a team either did or didn't do. Germany is spending more and has done less. No amount of modesty moves those.

Germany is one profile among seven, and the smaller samples don't resolve as cleanly. But the finding underneath holds: conviction follows the work. What the rest of this chapter adds is that the same work reaches costs that have nothing to do with AI at all.

Cost in the broader estate

When we asked what is expensive about Microsoft 365, AI licensing came back on top. The rest of that list is the estate itself: storage that keeps growing, licenses nobody uses, tenants left over from acquisitions, on-premises infrastructure still running alongside the cloud. None of it's new. All of it was expensive before Copilot arrived.

There is a practical difference between AI costs and the rest of this list, and it is worth being precise about. You can reduce your AI bill by buying fewer seats. But that means less AI, which is rarely what anyone wants six months into a rollout. Reducing storage, reclaiming unused licenses, consolidating duplicate tenants: those usually do not cost you capability. They cost you time and effort. That is a different kind of decision, and historically it is the one teams chose not to prioritize.

Storage growth is a content sprawl problem CH 4The estate→. Duplicate environments from M&A are a migration problem CH 5Migration→. Unused licenses are a lifecycle problem CH 3Governance→. They arrive as invoices, but they originate as governance debt — and each gets more expensive the longer it waits, because storage grows, renewals recur, and duplicate tenants accumulate their own sprawl.

There is an honest complication on that chart too. 16% of teams name the cost of third-party tools needed to fill gaps in native capability as one of their biggest cost challenges. Buying something to do this work is itself a cost, and any argument for it has to clear that bar rather than pretend it isn't there. What the calculation usually leaves out is the other side of the ledger.

88% say cost pressures are a moderate or major driver of their migration and consolidation plans, with 36% calling it a major driver. Only one team in fifty says it plays no part. But look at where the weight sits: half call it one consideration among several, and only about a third say it is a primary reason they are moving. What matters here is the direction of causation. Teams are not migrating and then discovering savings. They are migrating because the cost of not doing so became visible.

Only one cost problem actually changes the migration decision

Share calling cost a major driver of migration or consolidation.

Among teams carrying duplicate environments or tenants from M&A activity, the share calling cost a major driver jumps to 44%, well clear of everyone else. No other cost problem on the list moves the number more than three points. That makes sense once you look at what a duplicate tenant actually is. Storage growth and unused licenses are costs you can attack without changing anything structural. Two tenants running in parallel can't be optimized away. The only way to stop paying for the second one is to consolidate.