AI Project Failure Rates: Where the 95% Comes From

Few of the famous “X% of AI projects fail” numbers measured AI projects failing. RAND’s 80% is RAND repeating “some estimates” from a 2022 magazine profile of an AI software CEO. MIT’s 95% is stated at least four different ways in a report that never shows the arithmetic, and the one chart it prints points nearer 75%. The IDC playbook behind the 88% never prints that number. It shows 4 launches out of 33 tries. Gartner’s 30% and 40% began as predictions. The studies with control groups and usage logs tell a duller story: AI speeds people up on work it suits, slows them down on work it doesn’t, and paid licences can sit unopened on any given day.

You should know what I sell before you read a single number. I set up Claude for small teams, and my pitch is that plenty of companies pay for AI and still do the work by hand. A scary failure rate would make that pitch easier. So every number from someone else links to where it came from, I say which ones are measured and which are guesses, and there are no client results of mine on this page. The few numbers about my own company come from the private repo we run it on, so those you’ll have to take on my word. Many pages that quote these numbers end with a booking link. Mine does too. Weigh it accordingly.

What each number actually measured

I had Claude go to the document behind each figure, then had three separate Claude runs check the draft against those sources. They caught real mistakes in it, so don’t take the page’s word for anything: the links are there to click. The word “fail” covers four different things in these sources: projects someone abandoned, projects not in production yet, executives who didn’t get the value they hoped for, and, in a few cases, nothing anyone counted.

The measured ones are mostly executives answering surveys, published by companies that sell the cure. IBM sells AI consulting, Lenovo sells AI servers, Gartner sells research. That doesn’t make the numbers false. It does mean someone along the way was happy to see the problem look big.

Four numbers, followed home

RAND’s 80% is a footnote to a CEO profile

RAND’s 2024 report on why AI projects fail says, early in its summary, “By some estimates, more than 80 percent of AI projects fail.” The footnote is a Fortune article from July 2022 about the CEO of Aible, a company that sells AI software. The article says “a slew of recent surveys” put the failure rate between 83% and 92%, and doesn’t name one. Fortune offers them as proof that AI “doesn’t deliver the financial returns company officials expect,” which is a narrower claim than failing.

RAND’s own research was 65 interviews in 2023, mostly with data scientists and engineers, and it left out projects that just used a pretrained model. What it found is useful and gets far less airtime than the 80%: the most common root cause is business leaders misunderstanding how to set the project up, starting with what problem the AI is supposed to solve.

MIT’s 95% disagrees with its own chart

The MIT report came from Project NANDA, a group at the MIT Media Lab, in July 2025. It calls itself “Preliminary Findings.” It rests on 52 interviews, 153 survey responses gathered at four industry conferences, and a review of 300+ public initiatives. Its only listed reviewer is one of its own authors, and it calls its figures “directionally accurate based on individual interviews rather than official company reporting.”

The report states the number at least four ways: “95% of organizations are getting zero return,” “Just 5% of integrated AI pilots are extracting millions in value,” “only 5% of custom enterprise AI tools reach production,” and “The 95% failure rate for enterprise AI solutions.” So Fortune’s headline, “MIT report: 95% of generative AI pilots at companies are failing,” was no great distortion. It repeated what the report said.

The trouble is in the report. It never shows how it got to 95%, and the one funnel it prints gives a different answer. For task-specific tools, custom or bought, 60% of organizations investigated one, 20% piloted one and 5% successfully implemented one. If those bars are the same organizations, about one pilot in four succeeded and three in four fell short. That’s a reading of the chart, not something MIT states, and 80,000 Hours did the same sum first. MIT notes that sample sizes vary, and “successfully implemented” is a high bar: “marked and sustained” impact, more than just reaching production. Still, three in four is a long way from nineteen in twenty. Futuriom’s version: “It’s like saying ‘95% of drivers failed at driving a car,’ when 80% of those people don’t even have cars” (Futuriom).

The same report says generic chatbots “appear to show high pilot-to-implementation rates (~83%),” then adds that “this masks a deeper split in perceived value.” We couldn’t find a headline that led with that. Two days after its first story, Fortune wrote that a selloff “appeared to be sparked in part by” the report, with Nvidia down 3.5% and Palantir nearly 10%. By 80,000 Hours’ account the authors work on AI agent frameworks, and the report names NANDA among the systems that will fix the problem. The original PDF link now redirects to the group’s page.

The IDC playbook never prints 88%

Page 8 of Lenovo’s CIO Playbook 2025, researched by IDC and commissioned by Lenovo, has a small graphic: 33 AI proofs of concept, 4 production launches. It cites a different IDC survey from the rest of the playbook and gives no sample size. We couldn’t find an IDC document that says 88%. The number shows up in a CIO.com article that credits IDC, and it matches 4 of 33 turned inside out. From there it grows. One dev shop’s blog now calls it “IDC research across thousands of enterprise projects.” Lenovo’s 2026 release says 46% of proofs of concept reached production. That’s a different survey, so it isn’t a trend, and it gets a fraction of the attention.

VentureBeat’s 87% became a citation

In July 2019 VentureBeat ran a “Partner Content” write-up of a conference panel headlined “Why do 87% of data science projects never make it into production?” The number has no source. The one sentence behind it says “only 13% of data science projects, or just one out of every 10,” and 13% is not one in ten. A 2022 paper on arXiv then cites it flatly: “According to VentureBeat [1], 87% of data science projects never make it into production.”

How a guess becomes a fact

Most of these took the same road. A hedged sentence loses its hedge, then picks up a method it never had (“tracked 2,400 initiatives,” “thousands of enterprise projects”), then goes on a sales page. Researchers call the first part the Woozle effect: a claim earns trust by being repeated, and nobody goes back to check. Brandolini’s law says it takes far more work to knock down a bad number than to make one up, which is my guess at why the corrections seem to have traveled less far than the headlines.

I had Claude read sixteen consultancy and vendor pages that use these numbers. Most open with the statistic, put you in the failing group and offer a call. Of the fourteen whose links it could inspect, three linked a primary report. It cuts the other way too. An AI vendor and an AI training company have both argued against the 95%, and they have something to sell as well.

Mark Twain popularized “lies, damned lies, and statistics” in 1907 and credited it to Disraeli. The phrase isn’t in any of Disraeli’s works, and the earliest known uses turn up years after he died. Even the line about bad statistics has a bad source.

What the careful studies found

The studies built to measure results, with control groups, usage logs or government records, agree with each other a good deal more than the headlines do.

AI helps on work it fits. Customer support agents with an AI assistant resolved 15% more issues per hour, and less-experienced agents about 30% more (Brynjolfsson, Li and Raymond). Professionals doing writing tasks in a randomized experiment finished faster and wrote better.

It hurts on work it doesn’t fit. In an experiment with 758 BCG consultants, those using GPT-4 were 19 percentage points less likely to get the right answer on a task built to fall outside what the model does well (Dell’Acqua et al.). In early 2025, 16 experienced open-source developers took 19% longer with AI on their own code, and afterward believed it had saved them about 20% (METR). METR’s 2026 follow-up leans toward a speedup now, and METR itself calls that data “an unreliable signal.” The part I’d keep in my pocket is the gap between how it felt and what the clock said, since most “time saved” figures are people’s own estimates.

It helps some owners and hurts others. A randomized trial gave 640 Kenyan entrepreneurs an AI mentor for about five months. On average, no detectable change. High performers did about 15% better and low performers about 8% worse, and the authors trace the gap to which advice people chose to act on (Otis et al.).

Outside the lab, the effect is small so far. Denmark matched AI use for 25,000 workers in 11 highly exposed occupations to official records of their pay and hours and found “precise null effects on earnings and recorded hours” (Humlum and Vestergaard). The UK Department for Business and Trade gave out 1,000 Copilot licences, mostly to volunteers, and watched the logs. About 30% of holders used it on a given working day. Drafting got faster, scheduling got slower, and the evaluation “did not find evidence that time savings have led to improved productivity” (DBT). In Australia’s Treasury trial, 6% of participants expected Copilot to help with little or none of their workload. Afterward, 59% said that’s what it did, among 100 people surveyed both times (Treasury).

Gallup’s May 2026 reading puts this in proportion: 15% of US employees use AI daily, 52% use it at least a few times a year, and 47% say their organization has integrated it at all (Gallup). That’s all workers, not people with a seat, so it can’t tell you how many licences sit idle. As far as we can find, no vendor publishes that.

The problem I believe is real: nobody turns it off

None of this makes AI adoption a success story. In a Gartner survey of 2,002 CIOs and technology executives, released in October 2026, 13% of CIOs reported significant or transformational value from AI so far. The same survey holds the number I find most useful. 31% said they had guidelines for deciding when a pilot shouldn’t go to production, and 17% had guidelines for abandoning low-value AI that’s already live (Gartner). These are big-company CIOs saying whether a practice exists, so treat it as a sign and no more.

Deployments do get trimmed. In KPMG’s 2026 survey of 2,145 senior leaders, about half had delayed or scaled back AI agents because costs began to outrun value, and another 22% had questioned the decision and changed nothing (KPMG). Subscriptions are getting harder to leave. In Ramp’s card data, barely half of companies that subscribed to an AI product in 2022 were still subscribed at year end, and by the end of 2024 more than 80% were (Ramp). Ramp counts payments, not use, so some of that is people keeping tools they like.

What happens after a tool goes live is my own reading, and we can’t find anyone who has measured it. A tool with no exit rule renews by default, because a line item stays put until somebody moves to cut it. People have studied the habit since the 1970s, and in IT projects since the 1990s, under the name escalation of commitment: whoever backed a project keeps funding it after it stops working. I think that’s what paying for AI and still doing the work looks like from the inside, and it describes the risk better than any failure rate.

We’re no better. AI made it nearly free to add tooling to our company brain, so we added it fast, and between March and May we deleted 52 commands, skills and agents. In August we built three feedback loops so the system could learn from its own work. Two months later the calibration store had zero rows and the usage log had zero lines.

Our fix is a rule in that repo: a new build doesn’t get a full yes until it has kill criteria, a date and a test for deleting it. The ledger has two rows. One verdict was don’t build. The other says to delete the job if, by December 5, fewer than three flags a month clear within a week. Neither has an outcome yet, so I can’t tell you the rule works. Ask me in December.

How to check the next AI statistic you see

1. Find the original sentence. The report, not the article about it. If it takes more than a few minutes to find, you’ve learned something.
2. Find the denominator. 95% of what? Organizations, pilots, custom tools and “initiatives” give you four different numbers.
3. Check the tense and the method. “Will be abandoned by 2025” is a prediction. “At least 50% were” is a claim. Ask how each one was counted.
4. Ask who paid for it, and what they sell. That goes for whoever showed you the number, me included.

I’ll update this page as sources change. If you find a mistake or a better source, tell me. I’ll fix it and credit you.

If the numbers look like your team

A setup won’t help a team that has no work suited to Claude. And if your people already use their AI tools every day, you probably don’t need me. If you’re paying for seats that mostly sit idle, the ten-question team check shows where you stand without talking to anyone. Whoever sets it up, write down the date and the test for turning it off before you start. Setup starts at $3,997. For what counting your own AI output honestly looks like, here’s how ours made 1,042 tasks in four months, and why most of them were noise.