The most consistent finding in this year's enterprise AI research is that spending and returns have come apart. A July 2026 study of 639 senior enterprise AI leaders found that the share of enterprises whose AI ROI does not outpace their spend held at 57 percent, unchanged from 2025, even as 93 percent reported improved production capability, up from 88 percent. More AI is reaching production, and for most companies the financial return still is not following it. At the same time, a small minority is pulling sharply ahead. PwC's 2026 AI Performance Study found that 74 percent of AI's economic value is captured by just 20 percent of organizations. The gap between those two groups is a measurement problem before it is a strategy problem.
Two things are true at once. Almost every company can now point to AI improving some task somewhere. And almost none of them can show, in a number a CFO would defend, that the improvement outweighs what it cost. The second failure is what keeps a company in the 80 percent.
Why does AI ROI fail to outpace spend for most companies?
Because the inputs are easy to grow and the return depends on a step most organizations cannot see. Provisioning is a purchase. Deploying a model into production is an engineering milestone. Both climb quickly and get counted accurately. The return depends on what the July 2026 research calls the last mile: the distance between a model running in production and the business users actually changing how they work because of it. That last stretch is where value is created, and it is the one part of the chain most companies do not instrument.
The spend keeps climbing regardless
The investment is not waiting for proof. Deloitte's 2026 State of AI in the Enterprise, drawing on 3,235 director- to C-suite-level leaders across 24 countries, found 84 percent of organizations increasing their AI investment year over year. The vendors are spending to close the gap for their customers: Microsoft launched a 2.5 billion dollar effort called the Frontier Company, embedding roughly 6,000 engineers directly inside enterprise clients to move stalled pilots into production, as reported by TechTimes in July 2026. When a hyperscaler assigns six thousand people to the last mile, that is a signal about where the value is stuck, and it is not stuck at the point of purchase.
What separates the 20 percent that capture AI's value?
Not budget, and not tool count. PwC's study, which scored 1,217 executives against an index of 60 management and investment practices, found the leaders differ in what they do with AI once it is in place. They are approximately 2.6 times more likely than peers to say AI improves their ability to reinvent their business model, roughly twice as likely to redesign workflows around AI rather than add tools to existing ones, and 2.8 times more likely to have increased the number of decisions made without human intervention, while going further on governance. As PwC's Global Chief AI Officer Joe Atkinson put it, "only a minority are converting that activity into measurable financial returns."
It is a measurement capability before it is a strategy
Read those leader behaviors as operations rather than as advice, and a single dependency runs through all of them. You cannot redesign a workflow around AI without knowing what the workflow actually does today, at the level of tasks and systems rather than the org chart. You cannot safely move decisions to automation without a baseline of how those decisions were made and how well. And you cannot govern what you cannot observe. Each leader move assumes the company already knows what its work looks like. That assumption is exactly what most organizations cannot satisfy.
#### Why the pilot succeeds and the rollout stalls
This is the mechanics of the last mile. A pilot is run under conditions the team controls and watches closely, so it produces a clean result. The rollout removes the watching. Once the tool is live across dozens of teams, nobody is measuring whether the work changed, so the organization loses the thread precisely where the money is. The pilot proved the tool can work. The rollout was supposed to prove it did work, and that second measurement was never built.
What laggards measure versus what the leaders measure
| Question | The laggard's metric | What it actually shows | What the leaders measure |
|---|---|---|---|
| How much AI do we have? | Dollars spent, tools deployed | Commitment, not return | Output change per team against a comparable non-adopting team |
| Is it in production? | Models live, pilots shipped | An engineering milestone | Whether behavior downstream of the model actually changed |
| Are people using it? | Seats and licenses active | Access, not habit | Frequency and depth of use on real work |
| Did it pay off? | Vendor usage dashboard | Activity inside one vendor's product | Cross-tool return, disclosed with a confidence score |
| Where do we invest next? | Aggregate ROI estimate | An average that hides both extremes | Team-level variance, which is where the next dollar goes |
The right-hand column has one property in common. Every item requires reading the operational systems where the work happens, not the spend report or the vendor console above them. That is the structural reason a procurement record and a vendor dashboard cannot answer the ROI question: one measures the input, and the other sees only its own product and has a commercial interest in the number it reports.
How do you measure enterprise AI ROI instead of counting spend?
You measure the work, not the wallet. Treating AI impact as a first-class, measurable signal read from the systems where work actually happens is the design difference between the two groups. That is the AI Impact signal family, and it sits inside a Human Capital Operating System, the intelligence layer above the existing stack. It does not replace the HRIS, the analytics suite, or the productivity tools. It reads across all of them, which is the one vantage point none of them holds alone, and the one the last mile requires. For the step-by-step version of this, we set it out in how to measure AI ROI across your workforce.
The discipline that makes the number defensible
The method matters more than the dashboard, and the reason most companies land in the 80 percent is that their method was never built to survive finance. We take that argument apart separately in the AI ROI measurement problem. Levos compares an adopting team against a comparable non-adopting team, matched on tenure, role mix, work type and tool stack, and attaches a confidence score to every result. That is controlled cohort analysis with confidence scoring. It is observational, and it is disclosed as observational rather than presented as full attribution, which is documented openly in the Levos measurement methodology. A CFO does not need a certainty the data cannot support. A CFO needs a number with its limits stated, which is a different and more useful thing than a vendor's confidence.
#### The privacy line, stated plainly
Behavioral measurement earns its access by being disciplined about it. Individual-level data flows only to a person's direct manager. Aggregated team views require five or more people. Executive-level exceptions are escalation-based rather than standing. A measurement system that quietly becomes a surveillance system forfeits the trust it depends on, and the trust deficit is already one of the most cited reasons AI programs stall.
What to do with this
If your AI reporting leads with dollars committed and tools deployed, you are measuring the input and calling it the return. That is how a company stays in the 80 percent while spending like the 20 percent. The move is not a bigger budget. It is a behavioral baseline established before the next tool lands, so the question of what changed has an answer that does not depend on a pilot you can no longer reproduce or a survey of how people feel.
Replace the spend report with an observation of the work, and the board conversation changes from how much AI we bought to what it earned, on which teams, doing which work, and where the next dollar should go. That is the conversation the leaders are already having. It is also the same idea we develop in depth in what a Human Capital Operating System is.
Frequently asked questions
What is the AI ROI gap? The widening gap between the roughly 20 percent of companies capturing most of AI's financial value and the majority that are not. PwC found 20 percent of organizations capture 74 percent of AI's economic value. The divide tracks with the ability to measure what AI changes in the work, not with spend or tool count.
Why does AI ROI fail to outpace spend for most companies? Because spend and production capability grow easily while the return depends on a last mile that most companies cannot see. A July 2026 study found 57 percent of enterprises still do not see ROI outpace spend, unchanged since 2025, even as 93 percent report improved production.
What do the companies capturing AI value do differently? Per PwC, they are 2.6 times more likely to use AI to reinvent the business model, about twice as likely to redesign workflows rather than add tools, and 2.8 times more likely to automate decisions, while investing more in governance. Each move depends on knowing what the work looks like today.
How do you measure enterprise AI ROI properly? Establish a behavioral baseline, then compare adopting teams against comparable non-adopting teams matched on tenure, role mix, work type and tool stack, and attach a confidence score. Report it as observational, not as full attribution.
Is AI spend a measure of AI ROI? No. Deloitte found 84 percent of organizations increasing AI investment while only 20 percent say AI drives revenue growth today. Spend is an input. Only a change in the work is a return.
Measure what AI earns, not what it costs
Levos is opening early access to a small cohort of mid-market leaders who want AI return measured from the work their teams already produce, in a number the CFO can defend.
Design partner cohort today: 150 to 500 employees. Expanding to 500 to 2,000 in the second half of 2026.
