MIT's GenAI Divide Study: What the 95% Figure Actually Means

MIT NANDA's 2025 report, The GenAI Divide, found that 95% of the organisations it studied saw no measurable return from generative AI. The figure is weaker than its fame suggests: it rests on a small sample and a strict definition of success. Its explanation, though, holds up well. Most AI never learns the business it works in.

What did the MIT GenAI Divide study find?

The study found a sharp split between a small group getting real value from generative AI and everyone else. It reported that 95% of organisations saw no measurable return, while about 5% of integrated pilots produced large returns. The authors put the gap down to implementation, not to the quality of the AI models.

The report was published in July 2025 by Project NANDA, a research group at MIT. It spread quickly in August 2025, largely because of the 95% figure. Many headlines turned it into "95% of AI projects fail", which is not quite what it said.

The report's own framing was a divide. On one side sat organisations whose AI tools were used widely but changed little. On the other sat a minority who had built AI into specific workflows and saw measurable results. Most of the attention went to the size of the first group. The more useful material explains why the second group exists.

How was the study done?

The study combined three sources: a review of more than 300 public AI initiatives, structured interviews with representatives of 52 organisations, and a survey of 153 senior leaders. It is described as a preliminary study. Its headline finding on returns rests mainly on the interviews rather than on company financial reporting.

The definition of success matters as much as the sample. According to Marketing AI Institute's summary, the study counted a pilot as successful if it was deployed beyond the pilot phase with measurable KPIs, and if ROI impact was measured six months after the pilot.

That is a high bar. A tool that saves staff time every day, but was never formally deployed or measured, counts as no return under this definition. So does a promising pilot that is six weeks old. The 95% is therefore best read as "most organisations could not show a measured, deployed return within the study's window". That is still an important finding. It is just a different one from "95% of AI fails".

What are the criticisms of the MIT study?

Critics raised three main objections: the sample was small, the method behind the 95% figure was thinly explained, and the definition of success ignored real gains such as time saved or faster sales. The report itself acknowledges that its interview findings are directional rather than drawn from official company data.

The sharpest response came from Paul Roetzer, founder of the Marketing AI Institute. He said, "This is not a viable, statistically valid thing." He also argued that the study ignored efficiency gains, cost cuts, lead conversion and pipeline velocity. As he and others noted, the report describes its findings as "directionally accurate based on individual interviews rather than official company reporting".

Others have defended the direction while doubting the number. Their view, broadly, is that the report captured something real: returns have not shown up where leaders expected. Other surveys point the same way. According to McKinsey's 2025 State of AI survey, 88% of organisations use AI in at least one function, but only 39% report earnings impact at the enterprise level.

So the honest position is this. Do not quote the 95% as a precise failure rate. Treat it as one more sign, alongside better-sampled surveys, that adoption has run well ahead of measurable return.

What did the study find about shadow AI?

The study found that employees had adopted AI far faster than their employers. According to the report, only 40% of companies said they had bought an official AI subscription, yet workers at over 90% of companies used personal AI tools for work. Several commentators argued this was its most important finding.

This "shadow AI economy" reframes the 95%. In many businesses, the AI that delivers value is not the official project at all. Instead, it is staff using their own subscriptions to get through the week.

That has two consequences. The first is good news: people find AI useful enough to pay for it themselves. The second is a risk. Company data is going into accounts the business does not control, and whatever those assistants learn leaves when the person does. The guide to shadow AI covers the security side in detail.

The 90% figure is a share of companies, not of workers, so it means at least some staff at nine in ten companies. Microsoft and LinkedIn's 2024 Work Trend Index makes a similar point at worker level: 78% of AI users bring their own tools to work.

Why did most AI pilots fail, according to the authors?

The authors argued that the core barrier is learning, not infrastructure, regulation or talent. Most generative AI systems, they wrote, do not retain feedback, adapt to context or improve over time. People find them useful for quick tasks but do not trust them with work that depends on knowing the business.

This is the part of the report most worth reading. It describes a familiar experience. An assistant writes a fluent first draft, but it does not know the customer, the product's limits or how your company talks. So someone rewrites it. The next day, the assistant has learned nothing from the rewrite.

The report also found that external partnerships succeeded about twice as often as internal builds, and that only 5% of custom enterprise AI tools reached production. Both findings fit the same story. Getting AI to learn a business is hard, and the organisations that manage it tend to bring in people who have done it before.

What should a business take from the MIT report?

Take the explanation, not the headline. Returns come from AI that learns the business: workflows built around shared context, kept current and measured. Treat staff's personal AI use as both a signal of demand and a risk to manage. And measure any AI work against a baseline, so your business is never part of a statistic you cannot check.

Four practical lessons follow:

1. Give AI the business. Build the shared context, such as product facts, customer language and methods, that every assistant reads. That is what a company brain does.
2. Build the learning loop. When a person corrects AI output, the correction should improve the shared context, not vanish into one chat.
3. Bring shadow AI inside. Move staff onto business accounts and keep their best context where the company owns it.
4. Measure from the start. Set a baseline on a business number before AI touches a workflow.

If you want an outside view on where AI would pay in your business, Skein runs a fixed-scope AI strategy diagnostic. It suits a growing B2B business already paying for AI that cannot point to a return. A business with a clear plan and a working pilot may only need help measuring it.

Frequently asked questions

What is the MIT 95% AI study?

It is MIT NANDA's July 2025 report, The GenAI Divide: State of AI in Business 2025. It found that 95% of organisations studied saw no measurable return from generative AI. The study is preliminary, drawing on 52 organisation interviews, a survey of 153 leaders and a review of over 300 public initiatives.

Is the MIT 95% figure accurate?

Treat it as directional, not precise. The sample was small and success was defined strictly, as deployment beyond pilot with measured returns six months later. Critics, including Paul Roetzer of the Marketing AI Institute, questioned its validity. Better-sampled surveys, such as McKinsey's, still show a large gap between AI adoption and earnings impact.

Why do 95% of AI pilots fail?

According to the report's authors, the main reason is a learning gap rather than weak models. Most generative AI systems do not retain feedback, adapt to context or improve over time. Without the business's own knowledge, they produce generic work people will not trust with important tasks.

What is the shadow AI economy?

It is the report's term for employees using personal AI tools for work outside official company channels. The report found only 40% of companies had bought an official AI subscription, while workers at over 90% of companies used personal AI tools. It shows strong demand, and a data risk.

What does the MIT study mean for small businesses?

The lesson applies at any size. Returns come from AI built into specific, measured workflows with the business's own context, not from licences alone. Smaller businesses can act faster: pick one weekly job, give the AI the facts it needs, measure the result, and bring staff onto business accounts.

Related reading