Spending on AI is at record highs, and for most companies, the results still don't show up in the numbers that matter. Here's where the money is actually going, and where the next dollar should go instead.
Record AI spending and disappointing adoption aren’t a contradiction, they’re the same story told from two different ledgers. The money is mostly buying access: seats, infrastructure, model capacity. Adoption needs workflow redesign and reinforcement, which is a much smaller and far less funded line item. This piece breaks down where the budget actually goes, what the companies seeing real value do instead, and a short test for where your own next AI dollar belongs.
A finance director at a mid-sized logistics company is looking at the renewal notice for the company’s enterprise AI license. Attached to it is the usage dashboard her IT team pulled the week before, and it isn’t close: fewer than half the seats logged even one session in the past month. She signs the renewal anyway. The licenses are bundled into a three-year enterprise agreement, and unwinding it now would cost more than just letting the unused seats sit there. The AI line on next quarter’s budget goes up again, and so does the running joke in her team’s group chat about the tool nobody actually opens.
That’s not really a story about one company being bad at procurement. It’s closer to the default. The gap between what companies are spending on AI and what they’re actually getting back has quietly become the real headline, and it’s a strange one, because the money is not the missing ingredient. There has never been more of it.
Worldwide spending on AI is forecast to hit $2.59 trillion in 2026, up 47% from the year before, according to Gartner.[1] That’s one of the fastest runs of technology spending growth on record, from a research firm that tracks this for a living.
At almost exactly the same time, McKinsey’s newest global AI survey put the share of organizations attributing any measurable bottom-line impact to AI at just 37%, a number that hasn’t moved since the year before despite all that additional spending.[2]
Worldwide AI spending growth vs. the share of organizations reporting a measurable bottom-line impact from AI, both for 2026. Sources: Gartner [1], McKinsey [2].
Put those two numbers side by side and you get the actual shape of the problem. The investment curve is accelerating. The results curve is flat. Gartner’s own read on why is worth sitting with: its analysts describe most organizations as still favoring tactical AI projects with incremental efficiency gains, rather than anything that would count as real disruption, even as the dollars pour in. Spending fast and changing slowly can both be true of the same company at the same time, and right now, for most companies, they are.
The most detailed look at where that gap shows up in practice comes from MIT’s Project NANDA, which spent the first half of 2025 reviewing 300 disclosed enterprise GenAI deployments, interviewing representatives from 52 organizations, and collecting survey responses from 153 senior leaders. Its headline finding: despite $30 to $40 billion in enterprise GenAI investment, 95% of the organizations it studied were getting zero measurable return.[3]
That’s not a “give it another quarter” gap. That’s most of an entire category of spending producing nothing anyone can point to on a P&L. The real answer to “why isn’t AI working for us” almost never starts with the tool.
Access to a tool is not the same as a workflow that has changed.
None of this means the spending is wasted in some absolute sense, or that AI doesn’t work. It means the money is buying something real: capacity, seats, model access. What it isn’t reliably buying is the thing companies actually want, which is different work happening because of it. That distinction is what the rest of this piece is about. If you want the deeper mechanics of why a specific pilot stalls once it’s already technically live, that ground is covered separately in why your AI pilots never scale. This piece stays one level up: where the money is going, and what closes the gap between spending it and actually using it.
Ask most leadership teams where their AI budget is going and you’ll get an answer built around seats and speed: more licenses, more compute, a faster rollout. It’s a natural thing to reach for, because it’s the part of the budget conversation that’s easiest to measure and easiest to approve. Nobody has ever been challenged in a budget review for buying too much capacity.
The trouble is that capacity and adoption are not the same purchase, and the current spending mix makes that obvious once you break it apart. Of Gartner’s $2.59 trillion 2026 forecast, more than half, about $1.43 trillion, sits in AI infrastructure alone: the servers, cloud capacity and network fabric that make the models runnable at scale.[1] AI software, the layer that actually shows up inside someone’s daily work, accounts for a much smaller slice: around $453 billion.
Gartner’s worldwide AI spending forecast for 2026, by market segment. Software is the layer employees actually touch day to day, and it’s the smaller of the two. Source: Gartner [1].
None of that infrastructure spending is wrong to make. Model capacity genuinely has to exist before anyone can use it for anything. But it’s worth being honest about what that dollar buys on its own: more infrastructure changes what’s technically possible. It doesn’t, by itself, change what a claims processor does differently on a Tuesday, or whether a regional manager’s weekly summary gets built any differently than it did last year.
Inside most companies, the AI line of the budget tends to break into roughly the same four buckets, whether or not anyone has ever written them down that way:
That last bucket is the one that actually produces the four-stage chain companies say they want. Training builds the underlying skill. Workflow design gives that skill a specific, recurring task to live inside. Reinforcement, the repetition, review and manager follow-up, turns the new habit into something that survives past the second week. Adoption is simply what you see once the first three have actually held. Right now the money is landing almost entirely on the first stage, occasionally on the second, and rarely on the third.
A lot of what gets labelled “AI strategy” inside a company is really this same spending pattern with a slide deck attached: a budget increase, a rollout timeline, and very little that touches how the actual work happens. That specific pattern is covered in why most AI strategies are just for show, and the short version applies directly here: a strategy document is not a workflow redesign, and a bigger budget line doesn’t tell you which one you actually have.
Take a regional sales manager who’s started using AI to draft his weekly pipeline summary. He’s genuinely faster at it: the write-up that used to eat most of a Friday afternoon now takes him about twenty minutes. On paper, that’s exactly the kind of individual win AI is supposed to produce, and for him personally, it’s real.
But the review meeting his summary feeds into hasn’t changed. The same two people still re-check every number by hand before it goes to the VP, because nobody redesigned that review step to account for how the summary actually gets built now. His saved time evaporates two steps downstream, and the team’s forecast accuracy looks exactly like it did a year ago.
That gap between one person’s tool use and the organization’s actual results is close to the whole story of this stretch of AI adoption. It’s also exactly what Future Factors’ Tool x Workflows x Behavior framework describes at the individual level: having access to a capable tool doesn’t make someone an AI-powered professional on its own. It takes a workflow that genuinely uses the output, and a habit of checking and reinforcing it, before that access turns into anything durable. That idea gets a proper explainer in what makes an AI-powered professional different, so it won’t get re-run here.
McKinsey’s own numbers make the same point at scale. Eight in ten respondents to its 2026 survey say AI has genuinely improved their own productivity.[2] Those individual gains still haven’t turned into broad financial impact for the organizations employing them, which is exactly the disconnect between personal use and enterprise-level results this section is describing.
A tool that saves someone time upstream is not the same as a workflow that produces a better result downstream.
Where AI is genuinely doing real work inside a process like this, it helps to be explicit about which part is actually being delegated, rather than letting “AI helped with this” quietly become “AI is responsible for this.”
| What AI does | What the manager still owns | How it gets checked |
|---|---|---|
| Drafts the first-pass weekly pipeline summary from raw notes and CRM data | Deciding whether the summary’s framing still matches what actually happened in each deal | Reviewed against the CRM record before it’s shared, not after |
| Formats the numbers into the layout the VP expects to see | Confirming the review step itself, not just the drafting step, reflects how the summary is now built | The review meeting is the check; it has to be redesigned, not just shortened |
A worked example built around the pipeline-summary scenario above. The split changes by task; what doesn’t change is that something stays the manager’s to own.
The unresolved question underneath all of this is who actually owns checking the output before it moves downstream, and what happens when nobody does. That’s a governance question as much as a workflow one, and it’s covered directly in the AI governance gap. The short version that matters here: “no one owns checking it” is one of the most common reasons a real time-saving on one person’s desk never becomes an organizational result.
It’s tempting to assume the small number of companies actually seeing a return on AI just have a better tool, a bigger budget, or a head start. McKinsey’s survey lets you check that assumption directly, because it tracks a specific group it calls “AI high performers”: organizations that attribute at least 5% of their EBIT to AI and describe the impact as significant. That group has held steady at about 6% of respondents for two years running.[2]
What separates that 6% has less to do with the tool than most leaders assume. High performers are far more likely to have redesigned the actual workflow AI sits inside, rather than adding AI on top of the workflow that was already there. Nearly three-quarters of them report fundamentally redesigning a workflow because of their AI use, compared with about a quarter of everyone else. They’re also about twice as likely to say senior leadership visibly backs the initiative, and twice as likely to have a defined process for measuring what AI actually changes, rather than deciding after the fact whether it “seems to be working.” They’re 3.3 times more likely to be planning to fundamentally transform their business with AI, rather than just running it alongside the business as-is.
| Practice | AI high performers | Everyone else |
|---|---|---|
| Fundamentally redesigned a workflow because of AI use | ~74% | ~25% |
| Senior leadership visibly committed to the initiative | About 2x as likely | Baseline |
| Has a defined process for measuring AI’s impact | About 2x as likely | Baseline |
| Plans to fundamentally transform the business with AI | 3.3x as likely | Baseline |
“AI high performers” are the roughly 6% of organizations attributing at least 5% of EBIT to AI and describing the impact as significant. Source: McKinsey, The State of AI in 2026 [2].
None of that requires a different product. It requires treating the rollout as a change to a specific process, with an owner, a defined redesign, and a way of checking whether it held, rather than as a licensing decision that IT executes and everyone else absorbs.
It’s also worth saying plainly that none of this looks identical depending on who’s actually driving it. A CFO deciding where next year’s AI budget goes is answering a different question than a frontline manager deciding whether to change how her team hands off a task, which is different again from what an HR lead needs to reinforce once a rollout goes live. The next section breaks out exactly how that changes by role, because “companies should redesign their workflows” is true and almost useless without saying who does what.
The question most leadership teams ask at renewal time is which tool to add next. That’s the wrong question if access was never actually the bottleneck, and for most companies, based on everything above, it wasn’t.
The better question is where the next dollar buys redesign and reinforcement instead of another seat, and the answer depends heavily on who’s asking it.
| Role | What the next AI dollar should fund | What they own after it’s spent |
|---|---|---|
| CFO | Reallocating part of the license renewal budget toward a named redesign project, not another vendor line | Defining what “impact” means before the project starts, in terms the board will actually accept |
| Frontline manager | Time to map the specific handoffs in her team’s workflow that AI now touches, and to change the ones still built around a human doing the whole task | Deciding which parts of the new workflow get checked by hand, and how often |
| HR lead | Structured training and reinforcement design for the specific workflows being changed, not another generic tool-onboarding session | Tracking whether the new habit is still happening a month later, not just who logged in during week one |
A worked example, not a universal script. Change the specifics to your own org chart; keep the shape, one named owner per row.
That last row deserves its own emphasis, because it’s the piece most rollouts skip. The easiest thing to measure after a rollout is usage: how many people logged in, how many prompts got sent. That’s also the least useful measure of whether adoption actually happened.
A more honest lens moves through three stages instead of one. Use: are people actually doing the new thing? Persistence: are they still doing it a month later, without being reminded? Impact: is the work itself measurably better, faster or different because of it? Usage numbers can be the easiest first signal to check, but a rollout that stops at “logins were up” has measured access, not adoption.
Before the next line item gets approved, it’s worth running the specific workflow the money would touch through three plain questions, out loud, in the room where the budget actually gets decided:
Run the specific workflow the next dollar would touch through these three questions before approving the line item.
A short list of what tends to go wrong is usually faster to recognize than a long explanation of why. Watch for these signs that a workflow was only touched by a tool, not actually redesigned:
The next AI budget dollar should buy redesign and reinforcement, not another seat.
Pick one workflow in your own organization that AI already touches. Run it through the three questions above, honestly, out loud. If the answer to any of them is no, that’s where the next budget conversation should start, not on which vendor gets added next.
Because most of that spending buys access, seats, model capacity, infrastructure, rather than the workflow redesign and reinforcement that turn access into changed work. Gartner puts 2026 worldwide AI spending at $2.59 trillion, up 47% year over year, while McKinsey’s own 2026 survey found only 37% of organizations attribute any measurable bottom-line impact to AI, a share unchanged from the year before. The two numbers moving in opposite directions is the actual story: money and results are not the same curve.
There’s no single agreed number, because different studies measure different things, but the pattern holds across them. McKinsey’s 2026 survey found 63% of organizations report no measurable EBIT impact from their AI use, and separately, MIT’s Project NANDA found 95% of the organizations it studied in 2025 were getting zero measurable return on their GenAI investment specifically. Both point the same direction: spending has scaled much faster than results.
Almost always something else. The technology is capable enough for most of what companies are trying to do with it; what’s usually missing is a workflow that’s actually been redesigned around it, a named owner who checks the output, and a way of reinforcing the new habit past the first few weeks. Companies that jump straight from buying a tool to expecting different results are skipping the two steps that actually produce adoption.
McKinsey’s data on its “AI high performers,” about 6% of surveyed organizations, shows they’re nearly three times as likely to have fundamentally redesigned a workflow around AI (roughly 74% have, versus about 25% of everyone else), and about twice as likely to have visible leadership backing and a defined way of measuring impact before the rollout starts. None of that requires a different tool. It requires treating the rollout as a change to a specific process rather than a licensing decision.
Toward three things specifically: redesigning the actual workflow steps AI touches, naming an owner responsible for checking and reinforcing the new way of working after go-live, and measuring adoption through Use, Persistence and Impact rather than login counts alone. Who owns which piece differs by role: a CFO typically owns the budget reallocation and the definition of impact, a frontline manager owns the actual workflow redesign, and an HR lead owns the reinforcement and training design that makes it stick.
All three statistics in this piece were checked directly against their original source on 4 September 2026: Gartner’s own press release for the AI spending forecast, McKinsey’s own August 2026 survey report page for the EBIT-impact and high-performer figures, and the executive summary of MIT Project NANDA’s own PDF report for the GenAI return figure, rather than any secondary write-up of it. A widely repeated figure on enterprise AI project abandonment, attributed to S&P Global Market Intelligence, appears across multiple technology-news outlets but could not be confirmed against S&P Global’s own underlying report, so it was left out rather than used at second hand.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.