The MIT report behind the 95% statistic

What the report actually says

Start with the sentence itself, because almost nobody who cites it has read past it. MIT Project NANDA's report, The GenAI Divide: State of AI in Business 2025, states:

"Despite $30 to 40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return." (Punctuation normalised, wording unchanged: the original prints the dollar range with a dash.)

Read that sentence again and notice what it is not saying. It does not say 95% of pilots fail. It says 95% of organisations are getting zero return, against an outlay MIT puts at $30 to $40 billion. Zero return and pilot failure are different things, and the rest of this page is mostly about that gap.

The method, and the sample behind it

MIT describes its own method plainly:

"This report is based on a multi-method research design that includes a systematic review of over 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected across four major industry conferences."

That is three different instruments stitched into one report: a desk review of 300+ public initiatives, interviews with 52 organisations, and a survey of 153 senior leaders reached at 4 industry conferences. Worth holding onto that last detail. People who attend AI conferences are not a random sample of business leaders, and MIT does not claim they are. Treat the survey slice as opinion from an AI-forward crowd, not a population estimate.

The learning gap, in MIT's own words

MIT's stated cause, and it is a claim rather than a separately measured figure:

"The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time."

"Learning gap" is MIT's own term, not a journalist's shorthand: it is the title of the report's own section, "Why Pilots Stall: The Learning Gap Behind the Divide". Worth naming it for what it is even so. It is MIT's explanation for the 95%, offered by the same team that measured the 95%. That does not make it wrong. It is not an independently audited root cause either.

The denominator trap

Here is the comparison that actually travels: general-purpose tools succeed roughly 83% of the time, custom tools succeed 5% of the time. It sounds like one measurement with two results. It is two measurements with two different denominators, placed next to each other as though they answered the same question.

MIT's own words, on general-purpose tools against custom or vendor-sold ones:

"Tools like ChatGPT and Copilot are widely adopted. Over 80 percent of organizations have explored or piloted them, and nearly 40 percent report deployment. But these tools primarily enhance individual productivity, not P&L performance. Meanwhile, enterprise-grade systems, custom or vendor-sold, are being quietly rejected. Sixty percent of organizations evaluated such tools, but only 20 percent reached pilot stage and just 5 percent reached production."

And, on the general-purpose figure specifically:

"Generic LLM chatbots appear to show high pilot-to-implementation rates (~83%). However, this masks a deeper split in perceived value."

Line the denominators up. The 83% is a pilot-to-implementation rate: of the organisations that piloted a general-purpose tool, most went on to some form of implementation. The 5% is not that. It is the share of all organisations that evaluated a custom or vendor-sold enterprise tool and reached production, out of the full funnel MIT describes: 60% evaluated one, 20% reached pilot stage, and 5% reached production.

Run the same division MIT ran for the general-purpose figure, pilot stage to production rather than evaluation to production, and the gap narrows: 5 of the 20 organisations that reached a pilot went on to production, which is 25%, not 5%. Set against a pilot-to- implementation rate near 80% for general-purpose tools, that is still a real gap. It is nothing like sixteen to one, which is what the uncorrected comparison implies.

One more honest note. MIT's own bucket is "custom or vendor-sold", not "custom" on its own, so this figure does not by itself separate a bespoke build from a purchased enterprise tool. That question, build against buy, gets its own page: build vs buy AI.

What "95%" is a percentage of

Worth stating in one place, because it is the fact most likely to get flattened in the retelling. The 95% is organisations, not projects. It is zero return, not failure to ship. An organisation can deploy a working generative AI tool, keep it running, and still land inside the 95% if that tool never moves the profit and loss statement. MIT's own sentence draws the line precisely: organisations getting zero return, full stop.

The provenance problem

One more thing before you cite this anywhere. This is universally called "the MIT report". We could not find it hosted on any MIT domain, or any Project NANDA domain, and read the copy that exists from a third-party mirror. It has not been peer reviewed. None of that makes the underlying interviews, or the review of over 300 public initiatives, worthless. But a report that does not host its own PDF on its own website is not behaving like the institution whose name is doing most of the work when it gets quoted. Treat every figure in it, including the ones on this page, as contested until MIT publishes it directly.

Reconciling MIT with RAND

MIT's 95% and RAND's more than 80% are both real, and they are not the same measurement. RAND counts projects that fail: by some estimates, more than 80% of AI projects fail, about twice the rate of non-AI IT projects. MIT counts organisations getting zero return from generative AI specifically. A business can run several AI projects, watch most of them fail by RAND's count, and still sit inside or outside MIT's 95% depending only on whether generative AI moved its numbers at all. For the full table showing every widely quoted failure and adoption figure next to what it actually measures, see the AI project failure rate, reconciled. We are not restating that table here.

Questions, answered from the sources

Up: why AI projects fail. Also on this thread: RAND's failure report, read the same way, and build vs buy AI, on what this funnel does and does not settle.