Enterprise AI ROI: 95% of Pilots Show No Profit Yet

Update, Oct. 25: The gap this post describes is not closing. McKinsey’s 2026 State of AI survey (1,719 respondents across 97 countries, reported by TechTarget on Sept. 25, 2026) found 44% of organisations now scaling AI across the enterprise, up from 38% a year earlier, and 80% of employees saying AI improves their individual productivity — yet only 37% of leaders could attribute any positive EBIT effect to AI, exactly where that figure sat in 2025, and just 6% qualified as AI “high performers” (AI worth at least 5% of EBIT), also flat year on year. Adoption is sprinting; the bottom line is barely crawling. Read MIT’s 95% precisely: it is a measurement failure before it is an AI failure — pilots that never set a baseline or a cost per outcome cannot show a return even where one exists.

Somewhere in corporate America right now, a chief executive is staring at an AI invoice that runs to seven figures and a results slide that runs to adjectives. Faster. Smarter. Transformed. The words are doing a lot of lifting, because the numbers are not. This is the quiet scandal of the AI boom: not that the technology fails, but that companies keep buying it as though buying were the same thing as benefiting.

Ask any boardroom what AI has delivered this year and you will hear a confident story. Ask for the line on the profit-and-loss statement where that story lives, and the room goes noticeably quieter. That gap, between the purchase and the payoff, is the subject of this article. It has a name in the research, it has a size, and — this is the part that should worry investors — it has barely narrowed even as the models themselves have got dramatically better.

We checked each of the big survey claims behind this story against the original survey write-ups and the firms' own pages before writing it, and the figures that did not survive that check did not make it in. What remains is still damning enough. Enterprises are not being conned by bad technology. They are being undone by a much older corporate habit: mistaking a demo for a deployment, a licence count for a result, and enthusiasm for evidence.

The stakes are not academic. Gartner forecast in January 2026 that worldwide spending on AI would total $2.52 trillion in 2026, a 44% increase on the year before, with AI infrastructure alone adding $401 billion. That is not a technology budget any more. That is a bet the size of a national economy, placed largely on the assumption that returns will catch up with spending. So far, for most companies, they have not.

Photo: Server racks at the National Energy Research Scientific Computing Center (NERSC). File photo, not a picture of any company named in this article. Credit: Wikimedia Commons, CC0 public domain dedication.

The number that stopped boardrooms

Every debate about enterprise AI eventually collides with one statistic. In July 2025, MIT's Project NANDA published The GenAI Divide: State of AI in Business 2025, and the headline travelled faster than the report itself: 95% of generative AI pilots delivered no measurable impact on profit and loss. Only about 5% of integrated pilots were generating rapid revenue acceleration or significant value, against an estimated $30 billion to $40 billion already invested in enterprise generative AI.

It is worth pausing on what that number actually says, because the popular version of it is wrong in an instructive way. The report did not find that 95% of AI projects crashed, or that the models did not work. It found that 95% produced no measurable P&L impact. A pilot can run beautifully, delight its users, and still move no financial needle at all, because nobody wired it into the process that makes money, or because nobody measured the baseline before switching it on. That distinction matters. It is the difference between a technology failure and a management failure, and the evidence points overwhelmingly at the second.

The report's own methodology deserves the same honesty. Its research period ran from January to June 2025, and it rested on a review of more than 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organisations, and survey responses from 153 senior leaders collected at four industry conferences. Later press coverage quoted larger interview and survey counts, and readers should treat the report's own figures, stated here, as the authoritative ones. It was preliminary research, not a census of every company using AI, and its sample was drawn from initiatives the researchers could access or read about, not randomly selected. None of that makes the finding meaningless. It makes it a warning light, not a death certificate, and any executive quoting it as proof that all AI fails is misreading it just as badly as the executive ignoring it.

What gave the finding its force was corroboration. McKinsey's 2025 State of AI survey, covering 1,993 respondents, found that 88% of organisations used AI in at least one business function, yet only 39% reported any enterprise-level EBIT impact at all, and most of those put that impact below 5% of EBIT. McKinsey itself put it bluntly in a later write-up: despite widespread piloting and experimentation, just 39% of organisations were reporting EBIT impact at the enterprise level. Read that sentence twice. Nearly nine in ten companies use AI. Fewer than four in ten can point to it in their earnings. That is the ROI gap in one paragraph, and no amount of keynote enthusiasm closes it.

What the surveys actually measure

Part of the confusion is that the surveys are not all measuring the same thing, and vendors quote whichever number flatters their pitch. MIT measured measurable P&L impact from generative AI pilots. McKinsey measured whether respondents attributed any enterprise-wide EBIT impact to AI. Boston Consulting Group measured something else again: how many companies generate substantial value from AI at scale. In BCG's 2025 research, only 5% of companies qualified as future-built, generating substantial value, while 35% were scaling with some returns and 60% reported minimal revenue and cost gains. BCG's own July 2026 write-up repeated the point in starker form: in its AI Radar 2026 survey, only 5% of companies reported generating value at scale.

These are different yardsticks, and conflating them produces nonsense. But they point in the same direction, which is why they are worth taking together rather than picking apart. Adoption is near-universal. Enterprise-level financial impact is a minority sport. And the minority is small by every measure anyone has tried: 5% here, 6% there, 39% on the loosest definition McKinsey offered, with most of that 39% sitting below 5% of EBIT. When three research houses using three different methods keep landing in the same neighbourhood, the neighbourhood is real.

There is also a second measurement trap, quieter but just as expensive. Companies count what is easy to count: licences bought, employees trained, prompts typed, dashboards opened. None of those is a return. A chatbot that 10,000 employees open every morning and that changes no customer outcome, no cost line, and no revenue line has delivered engagement, not value. The firms that do capture value tend to be almost boring about this distinction. They pick one process, one metric, one owner, and they refuse to declare victory until the metric moves. Everyone else declares victory at the demo, and then wonders why finance will not renew the budget.

This is also why the bubble debate around AI keeps talking past itself. Ask whether AI is overhyped and you get two honest answers at once: the technology is genuinely transformative, and the spending is genuinely ahead of the returns. Our own tulip-monkey test for AI hype made the same distinction, separating the question of whether AI works from the question of whether current prices and promises assume it works everywhere, immediately, for everyone. The ROI gap is the evidence for the second half of that sentence.

How to measure enterprise AI ROI (so the CFO believes it)

Start with a baseline before you build: what does the task cost today, in labour, tokens, review time and rework? Then price every outcome the same way — cost per ticket closed, per document processed, per defect caught — and track the defect and rework rate alongside it. Hours saved count as savings only when they convert into revenue, a cancelled hire or lower contractor spend; a dashboard that says “40% faster” without a money line is not a return, it is a vibe. The pattern is visible even at the firms selling the dream: Accenture sold clients on AI — then put its own staff on a $29 allowance. And the spending behind the gap keeps climbing: Big Tech will burn $1 trillion on AI. Even the banks are pulling back where the return is thinnest: Wall Street is deleting AI agents from your bank account.

The learning gap: tools that cannot learn the job

MIT's researchers gave the central problem a name that has stuck: the learning gap. Most generative AI systems, they argued, do not retain feedback, adapt to context, or improve over time. A general-purpose chatbot is superb at a first draft and hopeless at a hundredth client file, because it forgets the ninety-nine that came before. As the report put it, the tool's very limitations reveal the core issue: it forgets context, does not learn, and cannot evolve. Inside an enterprise, where value lives in exactly that accumulated context — which client prefers what, which supplier always delivers late, which invoice format always breaks — a tool that starts fresh every session is a brilliant intern who never gets promoted.

Aditya Challapally, the lead author of MIT's NANDA project, explained the pattern to Fortune in August 2025, and his explanation is worth quoting because it inverts the usual excuse. Executives tend to blame regulation or model quality. Challapally pointed somewhere else entirely. He said of the successful minority: "They pick one pain point, execute well, and partner smartly with companies who use their tools." Startups led by 19- or 20-year-olds, he noted, had seen revenues jump from zero to $20 million in a year, precisely because they chose one problem and stayed with it. The large enterprise, by contrast, launches forty pilots, staffs none of them properly, and then commissions a report on why AI does not work.

There is a human detail in that story that deserves more attention than it gets. The successful teams are not buying better models. They are doing the unglamorous work of fitting the model to the workflow: cleaning one dataset instead of announcing an enterprise data strategy, fixing one broken handoff instead of redesigning the whole company on a slide, giving one team ownership of one number. BCG calls this the 10-20-70 principle in its wider work: roughly 10% of AI success depends on algorithms, 20% on technology and data, and 70% on people, processes, and culture. You can argue about the exact split. You cannot argue with the direction. The technology is the smallest part of the problem, and it receives the largest share of the budget and the attention.

The shadow AI economy makes the point from the other side. MIT found that only about 40% of companies had purchased an official large language model subscription, yet workers at more than 90% of companies reported regularly using personal AI tools for work. Employees, in other words, crossed the divide on their own, with consumer tools, while their employers' official programmes stalled in committee. That should embarrass every CIO who has presented a slide about AI governance. The workforce did not wait for permission. It found the tools useful for individual work and useless, in their official form, for enterprise work, and it behaved accordingly. The gap is not between believers and sceptics. It is between individual productivity, which AI delivers easily, and enterprise value, which it delivers only when somebody redesigns the work.

Where the money actually goes

Follow the money and the ROI gap stops looking mysterious. MIT found that more than 50% of generative AI budgets were absorbed by sales and marketing tools, while the biggest realised returns sat in back-office automation: eliminating business process outsourcing, cutting external agency costs, and streamlining operations. Companies are spending where the demos are prettiest and the internal politics are easiest, not where the returns are. A marketing chatbot photographs well in a board presentation. An invoice-exception bot does not, and it quietly pays for itself.

Then there is the cost side that nobody puts in the business case. A pilot that costs almost nothing for twenty users can cost a fortune at twenty thousand, once inference charges, retrieval infrastructure, data storage, monitoring, compliance reviews, and retraining are all counted. S&P Global Market Intelligence found, in a survey of 1,006 IT and line-of-business professionals in North America and Europe, that 42% of companies abandoned the majority of their AI initiatives in 2025, up from 17% in 2024, with the average organisation scrapping 46% of its projects between proof of concept and broad adoption. The top obstacles were not model intelligence. They were data privacy, security risk, and cost, each cited by roughly 38% of respondents for the first two. Gartner, meanwhile, reported that at least 50% of generative AI projects were abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, and unclear business value. Pilots do not die because the model is stupid. They die because nobody priced the whole system, and nobody cleared the security review until after the build.

The financing layer adds a further twist that investors are only now waking up to. In October 2026, S&P Global cautioned that guarantees and financing structures tied to AI infrastructure may be creating larger liabilities than public disclosures indicate, urging boards to stress-test cash flow projections and re-examine return assumptions. Our own reporting on the debt machine behind AI's infrastructure boom traced the same pattern: the build-out is increasingly funded with structures that look tidy on a balance sheet and less tidy under stress. When the returns on that infrastructure depend on enterprise customers who have not yet found their own returns, the circle is uncomfortably tight. Somebody, somewhere, has to be the first to actually make the money.

None of this means the spending is irrational. Infrastructure, once built, tends to find uses its builders did not imagine, and the history of general-purpose technologies is full of overbuilding followed by a long, profitable harvest. The dot-com crash did not kill the internet; it killed the companies that assumed the harvest would arrive on schedule. Our look at what the Nasdaq's 78% crash after 2000 teaches about AI made exactly that point. The question for an enterprise buyer is not whether AI will matter in ten years. It is whether this year's invoice will be justified by this year's return, and for most buyers, the honest answer is still no.

What the 5% that profit from AI do differently

So who is in the 5%? The research is unusually consistent on this, and the answer is almost disappointingly old-fashioned. The winners redesign workflows instead of decorating them. TechTarget's September 2026 analysis of the latest McKinsey data found that nearly three-quarters of AI high performers fundamentally redesign their business workflows around AI, compared with only about one-quarter of other respondents. BCG's leaders pursue a multi-year ambition concentrated where AI can create the greatest business value, rather than scattering pilots across dozens of departments in the hope that one of them lands. McKinsey's leaders set goals beyond efficiency, pursue growth and innovation too, and redesign workflows nearly three times as often as their peers, with senior leaders visibly owning the programme.

The payoff for that discipline is real, and it is large. BCG found that its top 5% of AI leaders delivered three times greater cost reduction, 1.6 times higher EBIT margins, and 2.7 times the return on invested capital than their peers. In its 2026 Applied AI Index, the future-built group delivered 2.3 times the three-year shareholder return of laggards, 2.4 times the revenue growth, and 2.8 times the growth in EBITDA. These are not marginal gains. They are the kind of gaps that decide which companies lead a sector in five years and which ones get acquired. The tragedy of the ROI gap is not that AI cannot deliver. It is that the delivery mechanism is known, documented, and widely ignored.

There is a counter-voice worth hearing here, because the hype is not entirely hollow. Matt Garman, the chief executive of Amazon Web Services, was asked at the HumanX conference in San Francisco in April 2026 whether AI was overhyped, and his answer was disarming. "People ask me that all the time. I think it's like one of the funnier questions I get," he said, before asking the audience to raise their hands if they were seeing positive ROI from AI or expected to within six months. He estimated about 70% raised their hands, and said the answer he usually gets is closer to 90%, adding that Amazon's software developers are roughly 4.5 times more efficient with AI. "That's not an efficiency that people just give up," he said. "And I think that there's more to come." Garman also conceded the other half of the story without being pressed: "There will be billion-dollar valued AI startups that don't make it. I'm 100 percent sure that's true."

Both halves can be true at once, and that is the point most commentary misses. Developer productivity is real. Individual productivity is real. Eighty percent of employees in McKinsey's latest survey said AI improves their individual productivity. What is not yet real, for most companies, is enterprise-level profit. Garman is selling the infrastructure on which that profit might one day be built, and his customers' raised hands are a measure of belief as much as of audited returns. The Register's reporter at the same event noted, with admirable dryness, that it would be remarkable if so large a share of the audience actually knew their company's AI costs and returns. Belief is not a line item. Sooner or later, finance asks for one.

Here is my judgment, stated plainly because this subject suffers from too much hedging. The AI ROI gap is not a verdict on AI. It is a verdict on how large companies buy technology: enthusiastically, decoratively, and without an owner. The firms that will win the next five years are not the ones with the cleverest models or the largest AI budgets. They are the ones that treat AI as an operating-model change, fund one painful workflow redesign at a time, and measure the result in money rather than applause. Everyone else is renting a revolution by the seat licence, and the renewal date is coming. Whether enterprise AI can close that gap will decide, more than any model release, whether this boom ends in a harvest or a hangover.

The 2026 update: the gap is narrowing, slowly

It would be dishonest to end the story in 2025, because the 2026 data shows movement, and movement in the right direction. McKinsey's latest global survey, of 1,719 respondents across 97 countries and reported in September 2026, found that 44% of organisations reported scaling AI across the enterprise, up from 38% a year earlier. Yet the financial ceiling barely moved: only 37% of leaders could attribute any positive EBIT effect to AI, a statistic unchanged from the previous year, and just 6% qualified as true AI high performers, meaning AI accounted for at least 5% of their EBIT, flat on 2025. Adoption sprinted. Returns crawled. The pattern held.

BCG's 2026 Applied AI Index, published in October 2026 and covering 1,330 senior executives, tells a brighter version of the same story. The future-built group grew from 5% to 7.5%, and a further 41% are now scaling AI and outperforming, meaning nearly half of companies are generating some value from AI. Nicolas De Bellefonds, BCG's global AI leader and a co-author of the report, put it well: "The perception is that AI isn't delivering. Our data says the picture is more nuanced: nearly half of companies are already creating real value from it, while many others are still struggling to translate investment into impact." That nuance matters. The gap is no longer "almost nobody benefits." It is "almost half benefit somewhat, a small elite benefits hugely, and the rest are still paying for the demo."

The market, meanwhile, has voted with its wallet in a way that should make enterprise buyers nervous. A J.P. Morgan analysis published on 8 October 2026 found that 42 AI-linked stocks drove most of the S&P 500's price gains since January 2024, contributing roughly 60% to 80% across price return, earnings growth, and growth in capital spending and research and development. The same analysis found sharp earnings growth among companies supplying AI infrastructure, while businesses adopting the technology had yet to show a comparable earnings advantage. In other words, the people selling the picks and shovels are getting rich. The people digging are still waiting for the gold. That is a normal phase of a technology cycle. It is also a phase that ends, one way or another, and the ending depends on whether the diggers start finding anything.

How to tell hype from return: a buyer's checklist

For any executive, investor, or manager reading this and wondering whether their own AI programme is in the 5% or the 95%, there is a short, brutal test. First, name the single metric the programme will move, and show the baseline measured before the programme started. If there is no baseline, there is no ROI claim, only a story. Second, name the executive owner who sits outside the IT team and whose bonus depends on that metric. If the owner is the vendor, or nobody, the programme is decoration. Third, count the full cost: licences, inference, integration, data cleanup, training, governance, and the human hours spent fixing AI output. If the business case only counts the licence, it is fiction. Fourth, ask whether the workflow was redesigned or merely decorated. If employees still copy, reconcile, and re-enter the same information by hand around the AI, the AI is a passenger, not a driver. Fifth, set a kill date. A pilot that cannot show a measured result by an agreed date should be stopped, and the budget moved to the one that can.

That checklist sounds harsh, and it is meant to. The alternative is what most companies are doing now: renewing on hope, expanding on anecdote, and discovering at the next budget review that the CFO has finally learned to ask the question that matters. The firms in the 5% are not smarter about AI. They are stricter about evidence. They buy less, finish more, and measure everything. In a market where everyone is buying, that discipline is the actual competitive advantage, and it costs nothing except the courage to say no to a good demo.

One final note for readers weighing who actually pays when an AI bubble pops: the answer, on the evidence of the ROI gap, is not mysterious. It will be the companies that bought the story without the metric, the investors who priced the harvest before the planting, and the employees whose workflows were decorated rather than redesigned. The technology will survive them all. It usually does. The question is whether your company will be among the survivors, or among the case studies.

Enterprise AI ROI FAQ: the questions boards actually ask

Why do 95% of pilots show no profit?

Mostly because they were never set up to show one. MIT’s finding is about pilots in its dataset that reported no measurable profit — and a large share of those pilots had no baseline, no cost per outcome and no owner for the return. Without those, even a tool that works cannot prove it.

Is 95% the same as “AI fails”?

No. Read precisely, the number describes a measurement failure first: pilots that never set a baseline or a unit cost cannot show a return even where one exists. That is why the 5% that do profit look boringly disciplined — baselines, unit costs, and a named owner for the saving.

Frequently asked questions

What percentage of companies see ROI from AI in 2026?

It depends on the yardstick, and that is the honest answer. MIT's 2025 research found only about 5% of generative AI pilots delivering measurable P&L impact. McKinsey's latest survey found 37% of leaders attributing any positive EBIT effect to AI, but only 6% qualifying as high performers with AI accounting for at least 5% of EBIT. BCG's 2026 Applied AI Index found 7.5% of companies future-built and a further 41% scaling with some value. The loose definitions give you roughly a third to a half. The strict definitions give you roughly one in twenty. Anyone quoting a single number without saying which definition they mean is selling something.

Why do most enterprise AI pilots fail to deliver ROI?

Because they are built to impress rather than to operate. The recurring causes across MIT, McKinsey, BCG, S&P Global, and Gartner are the same: no baseline metric defined before launch, no executive owner outside IT, data that is duplicated or poorly permissioned, security and privacy reviews that arrive after the build, and costs that multiply at scale. MIT calls the core problem the learning gap: tools that do not retain feedback or adapt to workflows. In plain language, companies buy a brilliant answer machine and then act surprised when it cannot do the job.

How much are companies spending on AI in 2026 without seeing returns?

Gartner forecast worldwide AI spending of $2.52 trillion in 2026, up 44% on the year before, with AI infrastructure alone adding $401 billion. Most of that spending is by technology providers building foundations, not by enterprises buying outcomes. The enterprise share is harder to isolate, but the return side is clear enough: with only 37% of leaders reporting any positive EBIT effect, a very large share of enterprise AI spending is currently unproven against financial returns. Spending is not the same as waste, infrastructure may pay off over years, but no honest accounting can call the current position proven.

Which companies are actually getting positive AI ROI in 2026?

The winners share a profile, not a sector. They redesign workflows around AI rather than layering tools onto old processes, they concentrate investment in a few core workflows, they set hard P&L targets with senior owners, and they partner rather than building everything alone. BCG's future-built companies show 2.3 times the three-year shareholder return of laggards and 2.4 times the revenue growth. On the infrastructure side, the suppliers of AI hardware and cloud capacity are reporting the clearest earnings gains, while most adopters are still waiting for theirs. If your vendor is making more money from AI than you are, you are the customer, not the winner, and that is worth noticing.

Why do most enterprise AI investments fail to show measurable ROI?

Often they do not fail to create value at all. They fail to prove it, which in a budget review amounts to the same thing. No baseline, no control group, no cost-per-outcome tracking, and success metrics like licences or prompts that no CFO accepts as money. Add to that the real failures: fragmented pilots with no path to scale, AI grafted onto processes it should have replaced, and targets set on productivity percentages rather than profit. Fix the measurement first, and some pilots turn out to have been working all along. Fix the workflow second, and the rest have a chance.

Sources: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025), via Fortune coverage and the Financial Express report of August 2025; McKinsey State of AI surveys 2025 and 2026, via McKinsey's own customer-care analysis and TechTarget's September 2026 analysis; Boston Consulting Group AI Radar 2026 and Applied AI Index 2026, via BCG's own July 2026 write-up and TechCentral's October 2026 report; Gartner press release, 15 January 2026, on worldwide AI spending; S&P Global Market Intelligence Voice of the Enterprise survey, via CloudServus, October 2026; S&P Global AI financing warning, October 2026; J.P. Morgan Asset and Wealth Management analysis, 8 October 2026, via EconoTimes; AWS chief executive Matt Garman at HumanX, April 2026, via The Register.

By Daniel Rao

Daniel Rao is a tech journalist covering AI and enterprise technology. He writes about how companies actually use new technology, not just how they announce it, and he would rather read a profit-and-loss statement than a press release.