Why Do AI Projects Fail? What the Ones That Work Do Differently

Most AI projects fail for business reasons, not technical ones. The problem was never defined, the AI had none of the business's context, nobody owned the result, or nobody trusted the output. The projects that work start from one weekly job, give the AI the business to work from, and keep a person approving what goes out.

How often do AI projects fail?

Often, though the exact rate depends on what counts as failure. RAND reported in 2024 that by some estimates more than 80% of AI projects fail, about twice the rate of IT projects without AI. Gartner and MIT NANDA have published separate figures for generative AI that point the same way.

The headline numbers come from different studies with different definitions, so they are not interchangeable. RAND's figure covers AI and machine learning projects broadly. According to Gartner, at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, because of poor data, weak risk controls, rising costs or unclear value. Gartner has since predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, for much the same reasons.

MIT NANDA's 2025 study, The GenAI Divide, went further. It found that 95% of the organisations it studied saw no measurable return from generative AI. It is a preliminary study, built on more than 300 public initiatives, 52 interviews and a survey of 153 leaders. However, its explanation is the useful part, and we come back to it below.

Meanwhile, adoption keeps rising. According to McKinsey's 2025 State of AI survey, 88% of organisations now use AI in at least one function, but only 39% report any impact on earnings at the enterprise level. Use is common. Return is not.

What are the root causes of AI project failure?

RAND interviewed 65 experienced data scientists and engineers and found five root causes. The most common is that the problem is misunderstood or miscommunicated. The others are missing data, chasing technology over the real problem, inadequate infrastructure, and applying AI to problems too hard for it to solve.

Here are RAND's five, in plain terms:

1. The wrong problem. Leaders and technical teams do not share a clear picture of what the project is for, so the team optimises the wrong thing.
2. Not enough of the right data. The model cannot learn what the organisation never recorded.
3. Technology first. The project starts from the newest tool rather than from a real problem worth solving.
4. Weak infrastructure. There is no reliable way to manage the data or put the model into everyday use.
5. Problems too hard for AI. Some tasks are beyond what the technology can do reliably today.

Notice that only the fifth cause is about the AI itself. The other four are about how the business defines, feeds and runs the work. That pattern holds even more strongly once you move from custom models to the generative AI tools most businesses actually use.

Why does generative AI fail differently?

Generative AI fails differently because most businesses never build a model at all. They buy subscriptions. As a result, the failure is rarely a project that crashes; it is a tool that produces confident, generic work nobody quite trusts, used differently by every person, with no way to measure what it returns.

MIT NANDA's authors did not blame model quality or regulation. Instead, they pointed to systems that do not retain feedback, adapt to context or improve over time. In other words, the AI never learns the business.

Consider what that looks like day to day. According to Microsoft and LinkedIn's 2024 Work Trend Index, 78% of people who use AI at work bring their own tools. Each of those assistants knows only what one person has told it. So when a salesperson asks for a follow-up email, the AI writes it from the internet's idea of a sales email, not from your pricing, your customers or the objection that came up on the call.

The output is plausible, which is the trap. Bad output gets caught. Plausible output gets sent. Over time, the team learns that AI work needs checking line by line, and the time saved quietly disappears.

The four failures most growing businesses actually have

Most AI failures in a business of 10 to 200 people look like one of these:

• Subscriptions without context. Everyone has a seat; nobody's AI knows the business.
• Pilots nobody owns. An enthusiast builds something useful, then moves on, and it stops. The guide to AI pilot purgatory covers why this happens.
• Output nobody trusts. The work is fast but needs so much checking that it saves little.
• Spend nobody can measure. Leaders cannot say what the AI returns, so they cannot decide what to do next.

What do the AI projects that work have in common?

The projects that work start from a specific, frequent job rather than a technology. They give the AI the business's own facts and methods, they have a named owner, and they keep a person approving anything published, sent or spent. They also measure the result in business terms, not hours saved.

Four habits show up again and again.

They start small and frequent. One workflow the team runs every week is a better first project than a broad transformation. It produces a result quickly, and the result is easy to judge.

They give the AI the business. The work reads from a shared company brain: what is true about the product, who the customer is, how the brand sounds, which claims are approved. Every person's AI then starts from the same place.

They have an owner. Someone is accountable for whether the workflow keeps working, which is the difference between a pilot and a process.

They keep the gate. A named person approves anything that goes out under the company's name. That is not a brake on adoption. It is the reason people are willing to use the output at all.

MIT NANDA also found that external partnerships succeeded about twice as often as internal builds. That is not an argument for outsourcing everything. It does suggest that bringing in people who have made the mistakes before shortens the path.

How do you stop your AI project failing?

Pick one weekly job, write down how your best person does it, and give the AI your business's facts before you give it the task. Name an owner, keep a person approving the output, and measure the result against something the business already tracks, such as pipeline or time to first draft.

A practical sequence looks like this:

1. Define the job and the measure. "Draft the weekly client report" and "time from request to approved draft" beat "use AI in marketing".
2. Write down the method. Capture how your best person does the job, in their words.
3. Build the context first. Load the facts the job needs before the AI writes a word. A Company OS is the general form of this step.
4. Keep the human gate. Decide who approves the output and what they check.
5. Measure for a month, then decide. Expand what worked, stop what did not, and record why.

If you want help choosing where to start, Skein runs a fixed-scope AI strategy diagnostic that ranks the opportunities in a business by return and effort. It suits growing B2B businesses already paying for AI that cannot say what it returns. A business that has a clear plan and only needs one tool configured may not need a diagnostic at all.

Frequently asked questions

What percentage of AI projects fail?

Estimates vary with the definition. RAND reported in 2024 that by some estimates more than 80% of AI projects fail. Gartner predicted at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. MIT NANDA's preliminary 2025 study found 95% of organisations saw no measurable return.

What is the main reason AI projects fail?

According to RAND's interviews with 65 data scientists and engineers, the leading cause is misunderstanding or miscommunicating the problem the AI should solve. For generative AI in particular, MIT NANDA pointed to systems that lack the business's context and do not learn from feedback, rather than to weak models.

Why do AI pilots never reach production?

Pilots stall when nobody owns them after the first enthusiasm, when they were built around a tool rather than a frequent business job, or when the output cannot be trusted without heavy checking. A pilot that starts from a weekly workflow, with a named owner and a clear measure, has a far better chance of becoming a process.

Is the AI model the problem when a project fails?

Rarely. Only one of RAND's five root causes concerns the limits of the technology itself. The others are about the problem definition, the data, the infrastructure and a focus on tools over needs. MIT NANDA likewise attributed most generative AI failure to implementation and missing context rather than model quality.

How do you measure whether an AI project worked?

Measure it against something the business already tracks, such as pipeline, cost per lead, time to first draft or hours of rework, rather than hours saved in isolation. Set the measure before the project starts, run it for a fixed period, and compare it with the same work done without AI.

Related reading