Why AI projects fail: what the studies actually measured

Every AI failure statistic, traced to its source

The numbers in circulation come from four different kinds of evidence: forecasts, surveys of executives, interviews, and one estimate nobody can trace. They measure different things, and most of the confusion comes from treating them as the same thing.

Every row was checked against the publisher's own report, release or page. Full references are at the bottom of this page.

The 95% figure from MIT: what it measured

The claim that 95% of AI pilots fail comes from The GenAI Divide, a July 2025 report from MIT's NANDA project. What the report says is that 95% of organizations were getting zero return on their generative AI spending, measured as profit and loss impact six months after a pilot.

The evidence behind it is a review of more than 300 public AI initiatives, 52 interviews, and 153 survey responses collected at four industry conferences between January and June 2025. The report calls its own figures directionally accurate rather than official, notes the risk of self-selection, and labels itself preliminary findings. It also promotes protocols the same project builds.

The report's more useful findings get less attention. Only 5% of custom, task-specific AI tools made it from evaluation into production, while generic chatbots converted at a far higher rate. Tools bought from outside partners reached deployment about twice as often as tools built in house. And workers at more than 90% of companies were already using personal AI tools, while only 40% of companies had bought an official subscription. The gap the report found is between what people are doing with AI and what their organizations have managed to deploy.

Where the 80% comes from

Many pages go further and describe RAND's study as an analysis of 65 initiatives, or of more than 2,400. RAND interviewed 65 people: 50 experienced industry practitioners and 15 academics. The value of the report is in what those people said about causes, and that part is well worth reading.

Leadership-driven failure came first by a wide margin, named as the primary cause by 84% of the industry practitioners interviewed: leaders optimizing for the wrong business problem, setting unrealistic expectations, and asking for models that do not fit how the work is actually done. Missing or inadequate data came second.

The 70% figure is older than AI

"70% of change initiatives fail" predates generative AI by three decades. It began as Hammer and Champy's own unscientific estimate in a 1993 book about reengineering, and a 2011 review in the Journal of Change Management found no valid empirical evidence for it. It is now routinely applied to AI programs. We cover it in full on our page of questions about AI and jobs.

What actually makes AI projects stall

Set the failure rates aside and the studies with real methods agree on causes to a striking degree. None of the leading causes is the model.

1. The wrong problem, set from the top

   RAND's interviewees put leadership first: unclear problem definition, unrealistic expectations, and projects chosen for the technology rather than the business need.

2. Value never defined

   Gartner's survey found the top obstacle, named by 49%, was estimating and demonstrating the value of AI projects. A pilot with no agreed measure of success cannot win the decision at the end of it.

3. Data that is not there

   Second in RAND's interviews, and one of the reasons Gartner gave for expecting projects to be abandoned after proof of concept.

4. AI bolted onto the old workflow

   McKinsey's 2026 survey found the companies getting value redesign workflows rather than inserting AI into existing ones. MIT's report pointed at the same thing from the other side: brittle workflows and tools that never learned the context of the work.

5. Managers who stay on the sidelines

   Employees whose manager actively supports their team's AI use are 8.7 times as likely to say AI has transformed how work gets done, and fewer than one in three say their manager does (Gallup, State of the Global Workplace 2026). It is the largest single variable measured, and the least managed.

BCG reaches the same conclusion from a survey of 1,000 executives: about 70% of AI implementation challenges come from people and process, 20% from technology, and 10% from the algorithms themselves.

What the companies getting value do differently

The pattern in the evidence is unglamorous. The companies getting a return choose a business problem first, define what success looks like before they build, redesign the work around the tool, and put a manager in front of every team whose job is changing.

Technologists often assume that technical challenges are the bottlenecks in AI adoption. In reality, leadership is the limiting factor.

Judgment Day, chapter 1

In practice we run this as a 90-day readiness sprint: audit what is actually in use and what people believe about their own jobs, arm the managers before anyone else, run small experiments on a fixed clock against metrics that matter, and debrief. Every experiment ends in a decision to scale it, change it or stop it. A pilot with no date for that decision is how projects drift into the statistics above. The whole method is on one page, and the people side of it is on our page about AI change management.