Businesses poured billions into AI, and 95% of it came back with nothing
Businesses have spent tens of billions of dollars putting generative AI to work, and the return so far has been close to nothing. The largest study of the question, from the NANDA initiative at MIT, found that around 95% of enterprise AI pilots produced no measurable effect on profit, while a small handful pulled clearly ahead. That gap has almost nothing to do with the models themselves and almost everything to do with how the tools get chosen, wired in and used, which is the part most owners never get told. This brief gathers the verified numbers on where the money went, how staff actually reach for AI, and what the winners quietly did differently, all traced back to the primary research so you can check each figure yourself.
Where the AI spend actually went
In the biggest study yet, 95% of enterprise AI pilots returned nothing measurable on the bottom line.
The honest starting point is the money. Researchers at the NANDA initiative at MIT examined how enterprise generative AI spending actually performed and published the result in 2025 as The GenAI Divide. Across the organisations they studied, somewhere around 95% of the pilots delivered no measurable impact on profit, even though enterprises had put an estimated 30 to 40 billion dollars into the effort.
Only about 5% of the projects reached the kind of rapid, visible return that everyone had been promised. The rest did not fail loudly so much as stall quietly, in the awkward space between a working demo and a tool the business genuinely runs on.
That is the number that should stop an owner in their tracks, because it says the common outcome of an AI project is not a modest win or a modest loss, it is a spend that simply goes nowhere while the invoice still arrives.
What happened to the AI pilots
MIT NANDA, 2025The verified numbers
Sourced- 95% no measurable return. That was the outcome across the enterprise generative AI pilots the NANDA initiative at MIT studied in 2025.
- 30 to 40 billion dollars invested. The estimated enterprise spend behind those stalled pilots (MIT NANDA).
- Only 5% pulled ahead. A small share reached the rapid revenue acceleration the rest never saw (MIT NANDA).
The barrier is method, not the model
The tools themselves work well enough, and what goes missing is the integration, the process and the learning around them.
What makes the finding useful rather than merely alarming is the reason behind it. The MIT researchers were direct that the problem is not infrastructure, not regulation, and not a shortage of clever models, but learning, because most generative AI systems do not remember feedback, adapt to a particular business, or improve with use unless someone deliberately makes them do so. A tool that is bought and then left on its own stays a novelty.
The same report found that more than half of generative AI budgets had gone into sales and marketing, while the clearest returns were sitting further back in the business, in the ordinary operational work that quietly eats hours. It also found that buying capability from specialised vendors and shaping it to fit succeeded far more often than building it in house, by a margin of roughly three to one, which is a polite way of saying that most internal AI projects are quietly reinventing wheels that already exist.
Budget versus return
MIT NANDA, 2025Why the spend does not land
Sourced- It is a learning problem. Most systems do not retain feedback or adapt to the business unless someone builds that in (MIT NANDA).
- The money chased the wrong work. Over half of budgets went to sales and marketing while the returns hid in back office operations (MIT NANDA).
- Internal builds struggled. Buying and adapting proven capability succeeded around three times as often as building from scratch (MIT NANDA).
Leaders and staff are not looking at the same tool
Enthusiasm without direction is where most of the wasted effort lives.
Underneath the spending sits a quieter problem, which is that the people buying AI and the people using it rarely agree on what it is for. Microsoft's 2025 Work Trend Index, drawn from 31,000 workers across 31 countries, found that 67% of leaders felt familiar with AI agents while only 40% of employees said the same.
Staff are reaching for these tools anyway, mostly because they answer at any hour and faster than a busy colleague would, but they are doing it in whatever ad hoc way suits the moment, without a shared idea of which jobs AI should own and which it should stay well away from. That distance between eager adoption and any real direction stays invisible on a balance sheet, right up until the output starts causing problems of its own.
The alignment gap
Sourced- 67% of leaders, 40% of staff. The gap in who feels familiar with AI agents (Microsoft, 2025).
- Adoption without a plan. Staff reach for whatever tool is at hand, with no agreed view of what AI should own.
- Direction is the fix. Review how your team already uses AI.
When AI output makes more work
41% of workers have been handed AI output that looked finished but only created more work.
The clearest sign that ungoverned AI use costs money rather than saving it now has a name. Researchers at Stanford's Social Media Lab and BetterUp Labs coined the term workslop in a Harvard Business Review piece in September 2025, describing it as AI generated content that looks like solid work but carries none of the substance needed to move a task forward.
In their survey of more than a thousand full time employees, 41% said they had received exactly that, and each instance cost the person on the receiving end close to two hours of effort to untangle and redo. The damage did not stop at the lost time either, because more than half of the people handed workslop thought less of the colleague who sent it, and a good share said they would rather not work with them again.
When a business hands staff powerful tools without teaching them how to use them well, this is the bill that quietly arrives, paid in rework and eroded trust rather than in a line item anyone thought to budget for.
The cost of workslop
Stanford and BetterUp, 2025What ungoverned use costs
Sourced- 41% received workslop. AI output that looked finished but lacked the substance to move the task along (Stanford and BetterUp).
- Close to two hours each. The rework every piece of workslop pushed onto the person who received it (Stanford and BetterUp).
- Trust took the hit too. Over half thought less of the sender, and many did not want to work with them again (Stanford and BetterUp).
They ran AI as a process, not a purchase
The winners closed the gap with ordinary discipline, not a cleverer model.
None of this means the technology is the problem, and the small group of businesses that did get a return is the proof. What set them apart was ordinary discipline rather than anything exotic. They picked a real problem worth solving instead of chasing whatever demo looked impressive, they wired the tool into the systems and the data their people already worked in, and they trained staff on the specific workflows they would use rather than handing out a login and hoping for the best.
They also tended to buy proven capability and shape it to fit, instead of quietly rebuilding it from scratch, and they aimed it at the unglamorous operational work where the hours genuinely hide. This is precisely the gap VibeZero exists to close, which is why our work starts with an honest read of where AI genuinely fits, then moves through a prioritised plan, the automation and build work that puts it into your systems, and the team training that decides whether any of it actually sticks.
How to land in the 5%
Action- Find where AI genuinely fits. AI readiness audit
- Build the plan around real problems. AI consulting
- Wire it into the daily work. Automation
- Train the team who will use it. AI training
- Own it past launch. Fractional Chief AI Officer
Frequently asked questions
The largest study of the question, the NANDA initiative's 2025 report from MIT, found that around 95% of enterprise generative AI pilots produced no measurable effect on profit, and it put the cause down to how the tools are used rather than to the technology, because most systems are bought and then left alone instead of being wired into real workflows, fed the business's own data, and improved over time. The projects that succeed treat AI as a process to run and a skill to build rather than a product to purchase.
Mostly in an unmanaged and ad hoc way, reaching for whatever tool is at hand because it is available around the clock and answers faster than a colleague would, which Microsoft's 2025 Work Trend Index captured alongside a wide gap in understanding, with 67% of leaders feeling familiar with AI agents against only 40% of employees. Without a shared view of which jobs AI should own, that enthusiasm tends to produce inconsistent and often low value output rather than the steady gains people expected.
Workslop is a term coined by researchers at Stanford's Social Media Lab and BetterUp Labs in a September 2025 Harvard Business Review article for AI generated content that looks like finished work but lacks the substance to actually move a task along, so the effort simply shifts onto whoever receives it. Their survey found that 41% of workers had been handed it, at a cost of close to two hours of rework each time, along with a real hit to how much colleagues trusted the sender.
The businesses that see a return start by choosing a genuine problem worth solving, then wire the tool into the systems and data their people already use, train staff on the specific tasks rather than handing out a login, and keep the work pointed at the operational areas where the hours quietly accumulate. VibeZero runs an AI readiness audit to find where AI genuinely fits, builds the plan and the integrations, and trains your team so the tools are actually used, which is the difference between joining the 5% that gains and the 95% that does not.
Every number, cited
Related Field Note
Read nextMost AI spend returns nothing, yours does not have to
honest answers, no pitch deck, no commitment.