Enterprise AI spending is huge and the return is not.

Across more than 65 studies published in 2025 and 2026, the failure numbers look contradictory only because each one counts something different. Reconcile them and one pattern holds.

AI works in the demo and stalls at the last mile, the point where a person has to operate it inside a real business. That last mile is a human problem, not a model problem, and the scarce input is the people who can build and run AI.

In short

The claim that most enterprise AI fails is directionally true and routinely exaggerated by stacking numbers that count different things.

• MIT's State of AI in Business 2025 found 95 percent of organizations see no return, while
• IBM's study of 2,000 CEOs found only 25 percent of initiatives hit their expected ROI and
• McKinsey's 2025 State of AI found 39 percent report any EBIT impact. Three numbers, three denominators.

The people who run AI say the barrier is themselves, not the technology. In a 2026 benchmark of Fortune 1000 data and AI leaders, 93 percent named culture and people as the main barrier to being AI-driven, not technology. Deloitte calls insufficient worker skills the biggest barrier to integrating AI, and ManpowerGroup reports AI development and AI literacy are now the two hardest skills to hire on Earth.

The productivity that was supposed to justify the spend is partly an illusion. A controlled trial found experienced developers were 19 percent slower with AI while believing they were 20 percent faster, and low-quality AI output now costs a 10,000-person company about 9 million dollars a year to clean up. The people who could close that gap are scarce and getting scarcer, with early-career employment in the most AI-exposed jobs down 16 percent.

--------------------------------------------------------

--------------------------------------------------------

How this was built

More than 65 studies from 2025 and 2026 were collected, with primary research preferred over aggregators. Every figure was traced to its original publisher and checked against a live source. Where numbers appeared to conflict, they were reconciled by construct instead of averaged, because most apparent contradictions come from studies counting different things on different populations. Vendor-funded studies are flagged wherever their incentive points at the result. No proprietary or client data was used anywhere in this analysis. The reconciliations below are the part no stat roundup does.

Finding one. The failure numbers, reconciled

The widely repeated line that most enterprise AI fails is true in spirit and misleading in the arithmetic. The four figures people stack count four different things on four different groups.

‍

Abandonment is not the same as no return. A forecast is not a measured rate. A 2024 estimate is not a 2026 finding. Stacked together they manufacture a scarier headline than any single study supports. Read separately, the most disinterested estimates land at 25 percent of initiatives hitting expected ROI and 39 percent showing any EBIT impact. The optimistic outlier is worth naming for balance, since Google Cloud reported 74 percent achieved ROI in year one, and Google Cloud sells the AI agents in question.

Adoption is measured just as loosely. Ramp's payment data put business AI adoption at 50 percent, the US Census Bureau put firm-level use near 18 to 20 percent, and a Federal Reserve note reconciled the two by showing 18 percent of firms have adopted AI while 78 percent of the labor force works at a firm that has. Big firms are in, most small firms are not, and the headline you cite depends on whether you count firms, spend or employment.

Finding two. The barrier is the human layer, not the model

Ask the people running enterprise AI what breaks it and they point at themselves. In the Fortune 1000 data-leadership benchmark, 93 percent named culture and people the main barrier, not technology. The largest surveys agree.

‍

One major study disagrees, and the disagreement is worth sitting with instead of hiding. MIT's Project NANDA states plainly that the core barrier is learning, not infrastructure, regulation or talent. Reconciled against the rest, that is a smaller gap than it looks. Learning is a human-layer function too. Whether an organization calls its constraint skills, training, learning or hiring, the thing in short supply is human capability and the ability to grow it, not compute.

The governance half of the human layer is thinner still. ServiceNow found 59 percent of enterprises now use agentic AI while only 26 percent have the governance to control it. Deploying faster than you can govern is a people-and-process failure waiting to happen, not a model limitation.

Finding three. The productivity illusion and the verification tax

Adoption is real. The productivity behind it is partly an illusion, and what survives carries a hidden tax. This is the mechanism behind the stall, and it is the part almost nobody measures.

Start with the cleanest evidence. A randomized controlled trial by METR put experienced developers on real tasks and found they were 19 percent slower with AI tools, while the same developers believed AI had made them 20 percent faster. Felt speed and measured speed moved in opposite directions. Google's DORA program found the same shape at scale, with 90 percent of developers using AI, 30 percent reporting little or no trust in its output, and AI showing a negative relationship with software delivery stability.

Then the tax. Low-quality AI output that looks finished but is not, now labeled workslop, reached 40 percent of desk workers in a month and costs roughly 9 million dollars a year in cleanup for a 10,000-person company, per research from BetterUp Labs and Stanford. The verification burden lands hardest on the best people, since Upwork found 88 percent of the highest AI-productivity workers are experiencing burnout. Speed that has to be re-checked is not speed, and a tool that burns out your strongest operators is not free.

This is why adoption does not convert to return. The gain is smaller than it feels, the output needs a human to verify it, and verifying it well takes exactly the scarce skilled people the rest of this analysis is about.

Finding four. The talent market, reconciled and priced

If the constraint is people, the market for those people tells the story in prices. Demand is climbing, supply is thin, the premium is real but smaller than the headlines, and the entry-level pipeline is contracting.

Demand. LinkedIn ranks AI engineer the number one fastest-growing job in the US. Forward-deployed engineer postings rose between 729 and 800 percent year over year. And the demand has left tech, since AI-skill postings grew 144 percent against 7 percent for all postings, with employment placement agencies leading every industry at 69 percent. The staffing industry itself is now one of the fastest-growing buyers of AI skills.

Price, reconciled. The AI wage premium is real and smaller than the loudest number suggests. Job-ad studies put it high, at 56 percent from PwC and 28 percent from Lightcast. Real payroll data tells a cooler story, with Ravio measuring a 12 percent premium at the individual level, and Payscale finding 55 percent of employers pay no AI premium at all. The honest read is that the premium exists, it is largest for scarce senior builders, and most of the market has not repriced yet. At the top, senior forward-deployed engineers still clear 560,000 to 785,000 dollars in total compensation at frontier labs.

Supply. Readiness sits low across every study once the wording is normalized. Kyndryl found 23 percent of workforces fully ready, down six points year over year, and 52 percent of leaders say AI skills are harder to find.

Pipeline, reconciled. Whether AI is cutting entry-level jobs is the sharpest fight in the data, and the honest answer is that it is bending the pipeline, not breaking the market. Stanford measured a 16 percent relative employment decline for workers aged 22 to 25 in the most AI-exposed jobs, and a Harvard study of 65 million workers found firm-level causal evidence that GenAI cuts junior hiring about 9 percent and is seniority-biased. Yet the Yale Budget Lab finds no economy-wide disruption yet, Forrester forecasts that over half of AI-attributed layoffs will be quietly reversed, and Gusto shows small businesses still hiring near a million new grads in 2026. AI is rearranging demand toward people who can build and run it, not deleting the need to hire.

What it means

Put the four findings together and the AI ROI problem stops looking like a technology problem. The money is spent, the tools work in the demo, and the return stalls at the exact point where a person has to operate the thing inside a real business. That point is a hiring and enablement problem.

MSH CEO Oz Rashid describes the pattern from the field. Someone shows a cool use case, then never uses it that way again, and that lack of consistency is the real gap for a lot of companies. The reconciled evidence says the same thing at scale. The scarce input in enterprise AI is the human who can build, deploy, govern and run it, and the discipline to make one win repeatable. Buying more model does not close that gap. Growing and hiring the people does. For any leader looking at a flat AI return, the more useful question is not which model to buy, it is who owns this and whether they can do it. Closing that last mile between AI ambition and AI capability is the work MSH does, placing the AI talent and helping build the operations around it.

Frequently asked questions

What is the AI execution gap? It is the distance between what organizations spend on AI and the return they can measure, traced to the last mile where a person has to operate AI inside a real business. Across 65-plus 2025 and 2026 studies, the gap tracks to skills, hiring and readiness far more than to the technology itself.

Why do the AI failure statistics conflict? Because they count different things. The 95 percent measures no P&L impact, the 42 percent measures abandonment, the 40 percent is a forecast of cancellations and the 80 percent is an older estimate. Stacking them is the most common mistake in AI coverage. Read on their own terms, the disinterested numbers put expected ROI at 25 percent and any EBIT impact at 39 percent.

Is the 95 percent AI failure rate real? It is a real finding from one MIT study measuring organizations with no measured return, and it is not the same as 95 percent of AI projects collapsing. The majority stall short of measured ROI rather than fail outright, and the cause the data points to is the human execution layer, not the model.

Is the AI wage premium real? Real and smaller than the headlines. Job-ad studies show 28 to 56 percent, real payroll data shows about 12 percent at the individual level, and 55 percent of employers pay no AI premium at all. It is largest for scarce senior builders and most of the market has not repriced yet.

Is the AI talent shortage real or hype? Real, and visible in prices and in readiness. AI development and AI literacy are now the hardest skills to fill worldwide, workforce readiness sits below half in every survey once normalized, and the executives running AI say people, not technology, are the main barrier.

Does AI reduce the need to hire? It shifts the need instead of removing it. Entry-level openings in exposed roles are contracting, while demand for people who can build, run and verify AI is climbing. The net effect in the data is a harder, more expensive hire, not an easier one.

How was this analysis built? By collecting more than 65 primary and near-primary studies from 2025 and 2026, tracing every figure to its original publisher, checking each against a live source and reconciling conflicting numbers by construct instead of averaging them. No proprietary or client data was used.

Explore the roles behind the gap

• AI and machine learning engineer staffing
• Forward-deployed engineer recruitment
• Chief AI Officer recruitment
• AI Center of Excellence consulting
• AI enablement guide

Ready to close your own execution gap?

The studies agree on the shape of the problem even where they argue about the numbers. AI spend turns into AI return at the last mile, the moment the right people own it. Book a consultation with MSH to put the AI builders, operators and leaders in place who move your initiatives from pilot to production.

Book a Consultation