Why AI projects fail (and how to make sure yours doesn't)

"95% of AI projects fail." You've seen it on LinkedIn, in board packs and probably in a vendor's opening slide. It comes from an MIT report published in July 2025, and it says something a bit different.

The real research on why AI projects fail is more useful than the headline, and a lot less mysterious.

The short answer: AI projects mostly fail for organisational reasons, not technical ones. RAND's interviews with 65 AI practitioners found the most common cause is leaders misunderstanding or miscommunicating the problem AI should solve. Poor data comes second. After that: chasing technology over real problems, missing infrastructure to deploy anything, and asking AI to do things it can't. Generative AI adds one more: tools that don't fit the workflow.

Key points:

• The famous failure rates measure different things, and most come with caveats people drop.
• The biggest single cause is a fuzzy problem with no owner and no measurable outcome.
• Pilots die in the gap between a demo and a real workflow.
• Almost every cause is visible before you spend money, if you look.

How many AI projects actually fail?

Depends who you ask and what they mean by "fail". Here's what the primary sources say, in their own terms.

The report does not establish that 95% of all AI projects fail. MIT's figure is about organisations seeing measurable P&L return from generative AI, judged mostly through interviews. RAND's 80% is borrowed. The best measured number here is S&P Global's: a big and growing minority of companies are binning most of what they start.

The exact percentage matters less than the pattern. The studies describe different populations and outcomes. Several point to workflow, data and ownership problems; model limitations can also matter.

Why do AI projects fail?

RAND's 2024 report, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, is the most cited serious study on this. The team interviewed 65 data scientists and engineers, 50 from industry and 15 from academia. It was written with the US Department of Defense in mind, but the findings provide useful questions for other sectors, rather than a measured failure rate for them.

They found five root causes:

1. The wrong problem. Stakeholders misunderstand or miscommunicate what problem AI should solve. Models get optimised for the wrong metric, or don't fit the workflow.
2. Not enough of the right data to train an effective model.
3. Technology first. More focus on "the latest and greatest technology than on solving real problems for its intended users."
4. No infrastructure to manage data and deploy finished models.
5. Problems too hard for AI. In RAND's words, "AI is not a magic wand."

The first is the big one. 84% of RAND's industry interviewees cited leadership-driven causes as the primary reason projects fail. Usually that's because the person who understands the business problem and the person building the solution aren't talking properly.

Gartner's data backs up the second cause. In February 2025 it reported that 63% of organisations either don't have, or aren't sure they have, the right data management practices for AI. It predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data. If you're not sure where you stand, start with an AI readiness checklist.

Gartner's 2024 prediction about generative AI blames "poor data quality, inadequate risk controls, escalating costs or unclear business value." Unclear value is the one I'd underline. If you can't say how you'll measure AI ROI before you start, you won't be able to prove it afterwards, and projects that can't prove value are the first to be cut.

Why do so many AI pilots never reach production?

This is where MIT's report earns its attention. Its sharpest finding isn't the 95%. It's the funnel. For enterprise-grade generative AI tools, 60% of organisations evaluated them, 20% reached a pilot, and just 5% reached production.

MIT's explanation: most fail because of "brittle workflows, lack of contextual learning, and misalignment with day-to-day operations." The researchers call the core barrier a learning gap. Most generative AI systems "do not retain feedback, adapt to context, or improve over time." People happily use ChatGPT for a first draft, then won't trust an internal tool that forgets everything they told it last week.

Two more MIT findings are worth knowing if you're buying:

• Bought beat built. In MIT's sample, partnerships with external vendors offering learning-capable, customised tools reached deployment about 67% of the time, compared with about 33% for internal builds. The authors flag that these are self-reported and may not account for all confounding factors. More in build vs buy AI.
• Money goes to the wrong place. Roughly half of generative AI budgets went to sales and marketing, though back-office automation often paid better.

S&P Global saw something related from a different angle. Companies with above-average project failure rates were more likely to cite staff resistance, customer resistance and reputational worries. A pilot that works technically can still die because nobody changed how the work gets done, or nobody trained the people expected to use it. Which is why AI training for employees belongs in the project plan, not after it.

Are AI agent projects failing too?

Gartner thinks many will. Its June 2025 prediction is that over 40% of agentic AI projects will be cancelled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls." Same three culprits as before, in new clothes.

Gartner also flagged "agent washing": vendors rebranding chatbots, assistants and robotic process automation as agents. It estimates only about 130 of the thousands of agentic AI vendors are real. Analyst Anushree Verma put it bluntly: most agentic propositions "lack significant value or return on investment" because current models can't yet reliably achieve complex goals on their own.

I build AI agents in my own lab, Ascendancy IQ, where they do real work under human direction with approval gates and controls. My view: agents are genuinely useful for well-bounded, repeatable work with a human checking the important decisions. They are not yet a replacement for judgement. If you're shopping, AI agents for small business covers what's realistic.

Do small businesses fail at AI for the same reasons?

Mostly, yes, with a twist. Nearly all the research above is about large organisations. MIT defines enterprises as firms with over $100 million in revenue, and those are the ones it found leading on pilots but lagging on getting them into production.

Small businesses rarely fail through over-engineered infrastructure. They fail in simpler ways: a tool bought on a demo and never wired into anything, a project that depends on one enthusiastic person who then leaves, or nobody owning the result. The data problem is the same, just smaller. A spreadsheet held together by one person's memory is not AI-ready data.

A smaller firm may be able to change a bounded workflow quickly, provided the owner has the time and authority. That's exactly what McKinsey's 2026 State of AI survey says separates the high performers: they "fundamentally redesign workflows" around AI rather than inserting it into existing ones.

What I've seen

I've spent more than 20 years turning technology into working businesses. The uncomfortable thing about AI project failure is how ordinary it is. Most of it is old-fashioned project failure with a new label.

The common mistake is starting with the tool. "We should do something with AI" is not a problem statement. "We lose a third of inbound enquiries because nobody answers after 5pm" is.

Nobody owns the outcome. IT owns the tool, operations owns the process, and the result belongs to nobody. RAND recommends committing a team to a problem for at least a year. Most AI projects I hear about don't get a quarter.

The pilot proves the wrong thing. It shows the AI can do the task in a demo, not that it can do it inside your real workflow, with your real data, at your real volume. A demo is a promise. Production is the proof.

Running costs arrive late. Usage bills, oversight time and maintenance show up after launch, and that's when "escalating costs" turns into a cancellation.

How to make sure your AI project doesn't fail

1. Write the problem in one sentence, in business terms, with a number in it. If you can't, stop here.
2. Name one owner for the outcome, not the technology, with authority to change the process.
3. Check the data before you buy anything. Where it lives, how clean it is, who can give access.
4. Set a baseline and kill criteria before launch: what success looks like by when, and what result makes you stop.
5. Pilot inside the real workflow, with real users and real data, scoped so it can go to production if it works. My guide to running an AI proof of concept covers how.
6. Budget for running it: usage, oversight, maintenance and training, for at least two years.
7. Put basic governance in place early. A light AI governance framework answers the risk questions before they cancel the project.
8. Run a pre-mortem. Imagine it's a year from now and the project failed. Write down why. Then fix those things first.

Frequently asked questions

What is the AI project failure rate? There's no single reliable figure. S&P Global found 42% of companies abandoned most of their AI initiatives before production in its 2025 survey, up from 17% a year earlier. Higher figures, such as RAND's "more than 80 percent", are estimates cited from other sources.

Is it true that 95% of AI projects fail? Not quite. MIT NANDA's 2025 report said 95% of organisations were getting zero return from generative AI, based mainly on interviews and self-reported outcomes. The authors themselves describe the figures as directional.

What is the most common reason AI projects fail? According to RAND's interviews with experienced practitioners, it's misunderstanding or miscommunicating the problem AI is meant to solve. Poor or insufficient data is the second most common cause.

Why do generative AI pilots fail to scale? MIT found most fail because of brittle workflows, tools that don't learn from feedback, and poor fit with daily operations. Gartner adds poor data quality, weak risk controls, rising costs and unclear business value.

Do AI projects fail more often than other IT projects? RAND cites estimates that AI projects fail at about twice the rate of IT projects that don't involve AI. That comparison comes from secondary sources, so treat it as indicative rather than measured.

Sources

• The GenAI Divide: State of AI in Business 2025, Challapally, Pease, Raskar and Chari, MIT NANDA, July 2025 (PDF)
• The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, Ryseff, De Bruhl and Newberry, RAND, 2024
• Generative AI shows rapid growth but yields mixed results, S&P Global Market Intelligence (451 Research), 2025
• Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025, Gartner, July 2024
• Lack of AI-Ready Data Puts AI Projects at Risk, Gartner, February 2025
• Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner, June 2025
• The state of AI in 2026: On the road to ROI, McKinsey, August 2026

If you'd rather find the reasons your AI project might fail before you've paid for them, let's talk.