TL;DR
AI ROI is the measurable financial return you get for what you spend on AI, and right now most enterprises are not getting one. MIT found that despite $30 to $40 billion in enterprise spending, about 95% of generative AI pilots produced no measurable bottom-line impact. The gap is not the technology; it is how companies choose, integrate, and measure their AI, and the 5% who win do all three deliberately.
In July 2025, MIT's Project NANDA published a report that landed like a cold shower on enterprise boardrooms. After reviewing more than 300 AI initiatives, its researchers found that despite $30 to $40 billion in enterprise investment into generative AI, roughly 95% of organizations were "getting zero return," while just 5% of integrated pilots were "extracting millions in value." One executive quoted in the study put it bluntly: "The hype on LinkedIn says everything has changed. Nothing fundamental has shifted." If you are struggling to show AI ROI, you are not failing at AI. You are simply living the median result.
What is AI ROI, and why is it so hard to prove?
AI ROI is the net financial gain from an AI investment divided by its total cost, expressed as a percentage. The formula is easy. The proving is not. Costs are scattered across API bills, cloud compute, licenses, integration, and the human time spent reviewing outputs, while benefits like "faster" or "better" resist a dollar figure. That is why so many programs never officially fail: the companies with no measurable return often do not measure rigorously, so they simply spend indefinitely on projects whose value they cannot demonstrate.
The $40 billion question: why 95% get nothing back
The MIT number is not an outlier. S&P Global found 42% of companies abandoned most of their AI projects in 2025, more than double the year before. Morgan Stanley reported that only 21% of S&P 500 companies could cite a measurable AI benefit at all. McKinsey found that while about 88% of organizations use AI somewhere, only around 6% are high performers capturing significant value. Independent market data from Menlo Ventures pegs 2025 enterprise generative AI spend at roughly $37 billion, so the "$40 billion spent" in the headline is real money. An IBM survey echoed it: only about 25% of AI initiatives delivered the expected return, and 56% of CEOs admitted zero significant financial benefit. Investors have noticed too, with Citi flagging a credit-spread penalty for companies labeled AI "adopters" rather than "enablers."
This is the AI revenue leak: not one dramatic failure, but billions dribbling out through pilots that never connect to the bottom line. Weak AI ROI is rarely a model problem. RAND estimates over 80% of AI projects fail, usually because of data and integration, not the algorithm.
This is one of five enterprise AI blindspots we're tackling at the CloudNuro AI Summit on August 13. 100+ enterprise leaders. Industry leaders from Google, Yahoo!, Jio, and Pavestone VC. View the agenda and reserve your seat.
Why AI pilots fail
Dig into the failures and the same five causes surface.
- Teams start with a tool instead of a business problem.
- Data is messy, so outputs are unreliable.
- Budgets go to the most visible use cases rather than the ones that pay off.
- No executive truly owns the outcome.
- And no one set a baseline, so success is unmeasurable even when it happens.
Budget misallocation deserves special blame. MIT found that more than half of generative AI budgets go to sales and marketing, while the highest and most measurable returns sit in unglamorous back-office automation. Companies keep funding the demo that impresses the board and starving the workflow that would actually improve AI ROI.
Buy vs. build: the 67% advantage
One decision predicts success better than almost any other: whether you buy or build. MIT's research found that purchased or partnered AI tools succeed about 67% of the time, roughly double the one-third success rate of internal builds. The reason is not that in-house teams are worse. It is that building a reliable AI product means owning the data pipelines, evaluation, guardrails, and ongoing maintenance, and most teams underestimate all four.
This does not mean never build. It means being honest about where you have a genuine edge. Buy the commodity capabilities a vendor has already hardened, and reserve internal builds for the narrow, proprietary use cases that are core to your business. Choosing "build" for something you could have bought is one of the quiet ways AI ROI leaks away, in engineer-months that never ship.
What the 5% do differently
The winners are not buying better models. They run AI like a portfolio, with discipline.
- They define value capture before approval, deciding up front whether a saved hour becomes redeployed capacity or a cost that actually leaves the budget.
- They baseline everything, measuring the process for weeks before deployment so the "after" is provable.
- And they see their full AI spend, because you cannot improve the AI ROI on money you cannot even find.
Do this and the payoff is large: firms that get it right report roughly $3.70 back for every dollar, and back-office automation programs commonly land measurable returns in the millions per year.
We're closing the CloudNuro AI Summit (August 13, virtual) with a panel on exactly this: which AI investments should scale, which should stop, and what the 5% did differently. 100+ enterprise leaders are joining. Reserve your seat.
A CFO framework for measuring AI ROI
If you cannot measure it, you cannot defend it in a budget review. A workable AI ROI framework has four parts.
- Start with a baseline: quantify the current cost of the process in labor hours, error rates, and dollars, captured for weeks before anything ships.
- Then track net cost impact, which is the cost before minus the cost after, minus the fully loaded AI cost of tooling, integration, maintenance, and human oversight.
- Next, estimate revenue attribution honestly, counting only incremental revenue you can actually trace to the AI, and saying so when you cannot.
- Finally, watch time to value; U.S. finance use cases average around eight months from first dollar to first measurable impact.
Run those four numbers per use case and the portfolio decision makes itself: scale what clears the bar, and stop what does not. That discipline, not a better model, is what separates the 5% from everyone else.
The bottom line
The 95% headline is not a verdict on artificial intelligence. It is a verdict on how most companies buy, deploy, and measure it. The technology works; the value simply leaks out between the pilot and the P&L. Closing that leak does not require a bigger budget. It requires seeing every dollar of AI spend, tying each investment to a defined outcome, and having the discipline to scale what works and stop what does not. Better AI ROI starts with a question every CFO should be able to answer today: for every dollar we put into AI, how many are we getting back?
Read next: The AI Unit Economics Blindspot. The ROI gap starts with the cost gap. https://www.cloudnuro.ai/ai-summit
.webp)
.svg){kind=small}