What the 95% AI failure statistic actually says

What does the report actually say?

The exact sentence is that "95% of organizations are getting zero return." Organizations, not pilots, not projects, not investments. The report is The GenAI Divide: State of AI in Business 2025, published in July 2025 by four authors at MIT NANDA, and it describes itself on its own second page as "Preliminary Findings."

Its method: a systematic review of more than 300 publicly disclosed AI initiatives, structured interviews with representatives of 52 organizations, and survey responses from 153 senior leaders collected across four industry conferences, over a research period of January to June 2025. It carries a disclaimer stating that the views are the authors' own and do not reflect the positions of any affiliated employers.

None of that makes it bad research. It makes it a well-documented industry survey rather than the law of physics it now gets quoted as.

Where does the 5% come from?

From one line of a chart that has two lines on it, and almost everybody has only ever seen the steep one. The report tracks general-purpose tools like ChatGPT and Copilot separately from embedded, task-specific GenAI built into a particular workflow. They behave nothing alike.

General-purpose tools: 80% of organizations investigated them, 60% piloted, 40% successfully implemented. Task-specific tools: 50% investigated, 20% piloted, 5% reached production. The famous 5% is the end of the second line only.

So the honest one-sentence version is that custom enterprise AI tools reach production about 5% of the time, while the general-purpose tools your staff are already using deployed successfully at four organizations in ten.

So what share of pilots actually failed?

About three quarters, for custom tools. If 20% of organizations piloted one and 5% got it into production, then roughly a quarter of the businesses that actually ran a pilot cleared the bar. That is a sobering number and a genuinely useful one. It is not 95%.

The 95% figure counts every organization in the sample, including the ones that never piloted anything. An organization that looked at a custom tool, decided against it and carried on is counted in the 95% getting zero return, which is accurate as arithmetic and misleading as a headline about pilots.

For general-purpose tools the same arithmetic runs the other way: 60% piloted, 40% implemented, so the majority of those pilots succeeded.

What counted as success?

A demanding bar, declared openly: deployment beyond the pilot phase with measurable KPIs, and ROI measured six months after the pilot. Anyone quoting the failure rate should quote the bar alongside it, because the bar is doing a lot of the work.

The report is candid about what that costs it. It notes that a "six-month observation period may be insufficient to fully assess successful deployment for complex enterprise systems, potentially understating success rates for longer-term implementations."

That is the authors telling you their own number is probably pessimistic. It appears in none of the coverage I have read.

How solid is the underlying data?

The report answers this itself, in a line directly under the chart everyone screenshots: "These figures are directionally accurate based on individual interviews rather than official company reporting. Sample sizes vary by category, and success definitions may differ across organizations."

It also names its own selection bias, observing that organizations willing to discuss AI implementation problems may differ systematically from those that declined. Fifty-two interviews is a small base for a number now cited as though it were a census.

This is what good preliminary research looks like. The failure is not in the report. It is in a chain of retellings that dropped every qualifier the authors put there deliberately.

Why did the wrong version travel so far?

Because it is rounder, it confirms something plenty of people already half-believed, and checking it means finding a PDF that most articles quoting it do not link to. A number with those three properties will outrun its own corrections more or less forever.

Compare the two versions. The one that spread is "95% of AI pilots fail." The one in the report is that 95% of organizations are getting zero return, where return means measurable KPIs six months after a pilot, in a sample the authors describe as directionally accurate and based on 52 interviews. Nobody forwards the second one.

It is worth noticing what that does to your own position. The same figure gets quoted at you by someone selling AI, as evidence you need help doing it properly, and by someone skeptical of it, as evidence the whole thing is hype. Both are quoting a sentence that is not in the report. The useful question in either conversation is which of the two lines on that chart they are talking about.

What did the 5% do differently?

They bought or built for one specific process rather than for a capability. The report's finding is that the divide is not explained by model quality or regulation but by approach, and that what separates the two sides is whether the system learns: whether it retains feedback, adapts to context, and improves in the workflow it was put into.

The organizations on the working side of it demanded process-specific customization, drove adoption from the front line rather than by central mandate, and held vendors to business metrics rather than usage metrics. The report describes them as behaving like outsourcing clients rather than software subscribers.

Read against a business running on paper, that is a familiar shape. The thing that works is narrow, aimed at a process somebody can name, and measured against how that process ran before.

Does buying beat building?

The report found that external partnerships reached deployment about 67% of the time against about 33% for internally built tools, roughly a two-to-one difference. And then it says, in its own words, that this "may reflect organizational capabilities rather than implementation approach alone," and that "the correlation between external partnerships and success does not necessarily prove causation."

I want to be plain about something here. That 67% is the single most flattering number in this report for a consultancy like mine, and it is the one I am least willing to quote as proof. A business that chooses an outside partner may simply be better at procurement, clearer about what it wants, or more willing to spend on the parts that are not software. Those same qualities would make it more likely to succeed on its own.

The authors labelled that limitation themselves. Anybody putting the 67% in front of you without it, including anybody selling you an engagement, is quoting the half of the sentence that suits them.

What should you actually take from it?

Three things survive every caveat above, and all three are useful if you run an operation rather than a technology company.

• ·Generic tools spread easily and mostly improve individuals. Widely piloted, widely deployed, and the report is explicit that they enhance individual productivity rather than P&L performance. Useful, but do not expect them to show up in your numbers.
• ·Custom tools fail on workflow fit, not on intelligence. The stated reasons are brittle workflows, no contextual learning, and misalignment with daily operations. Every one of those is a description of the process, not of the model.
• ·Most of the zero-return finding is unmeasured rather than absent. Success required measurable KPIs six months on. Plenty of organizations cannot produce that for processes they have run for twenty years, let alone for a pilot.

What does that mean for your own pilot?

Take the baseline first, because you cannot join the 5% by definition if nobody wrote down what the process cost before you changed it. Hours, error rates, cycle times, for the specific process you are about to touch. It is a week of dull work and it is the difference between a result and an anecdote.

Then keep the scope to one named process. The report's own split says the general-purpose tools deploy easily and move nothing on the P&L, while the things that move the P&L are narrow and fitted. That is an argument for starting smaller than feels ambitious.

And measure at the end against the number you wrote at the start. Not against the impression in the room.

Common questions

Sources

• MIT NANDA, The GenAI Divide: State of AI in Business 2025. Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari, July 2025. Research period January to June 2025. Method: systematic review of 300+ publicly disclosed AI initiatives, structured interviews with 52 organizations, and surveys of 153 senior leaders across four industry conferences. Labelled preliminary findings, not peer reviewed, and carrying a disclaimer that the views are the authors' own. All figures here are quoted from the report, including its stated limitations.