Why most AI projects fail — and the shape of the ones that don’t

A vendor writing about failure statistics has an obvious incentive to scare you and then sell the cure, so let us be precise about what the numbers do and do not say — and honest that they apply to us too. A failed build hurts the builder as much as the buyer; this checklist is the one we run on our own projects.

What the numbers actually say

• MIT’s widely reported 2025 study (Project NANDA) found 95% of corporate GenAI pilots produced no measurable P&L impact; about 5% captured value at scale.
• Industry analyses put overall AI project failure at 80%+ — roughly twice the rate of comparable non-AI IT projects.
• S&P Global-reported survey data showed 42% of companies abandoned most of their AI initiatives in 2025, up from 17% a year earlier.
• Organisations scrapped close to half of AI proofs-of-concept before they reached production.
• One finding worth sitting with: in MIT’s data, buying or partnering succeeded around twice as often as building purely in-house.

Read carefully, these are not statements about AI capability. They are statements about how organisations run AI projects. That is genuinely good news, because organisational failure has a checklist.

The five causes, in the order they kill projects

1. No definition of success

"We want to use AI for customer service" is not a goal; it is a mood. Projects that survive start from a number: reduce first-response time from four hours to ten minutes, cut invoice-processing cost per document by half. If success has no number, failure has no moment — the project just quietly stops being mentioned.

2. Data that was messier than anyone admitted

Every failed-project postmortem contains the sentence "the data was worse than we thought." The records were in seventeen spreadsheet formats, the CRM had six years of duplicates, the documents were scans. None of this is fatal — but discovering it in week six instead of week zero is. Look honestly at your own data before anyone quotes you anything; it is the core of our AI readiness checklist.

3. Output nobody's workflow actually consumes

The most common shape of the 95%: a pilot that works in a demo tab while the team keeps doing the job in the old tools, because the AI lives outside the system where the work happens. If using the AI requires changing apps, copying results, or trusting output nobody verified, the workflow wins and the pilot loses. Integration is not the last mile of an AI project — it is most of the project.

4. Technology-chasing

Projects picked because the board asked "what is our AI story?" fail at a different rate than projects picked because a specific process was expensive, slow, or error-prone. The test is embarrassingly simple: if the same proposal without the word "AI" would not get funded, it should not get funded with it. We keep a public list of the overhyped use cases we decline.

5. Sponsorship that fades after the demo

AI projects need decisions mid-flight: what counts as correct, which edge cases matter, when to ship. When the sponsoring executive stops attending after the kickoff, those decisions queue up, and a queued decision is a stalled project. The fix is structural, not motivational — short projects with weekly demos give sponsorship less time and less reason to fade.

The checklist that avoids them

• One workflow, not a platform. Scope the first project to a single expensive process. Platforms are what you build after two wins.
• A number, agreed in writing. The definition of done, with the metric it moves and who measures it.
• Data audit before quote. Anyone pricing your build without looking at your data is pricing someone else’s.
• A working slice by week two. Real pipeline, real data, one narrow case. This single demand filters most of the failure modes above — it is the spine of how a build should run week by week.
• Evals from week three. Quality gets a number and a regression test, or "almost done" means nothing.
• Ship into the workflow, not beside it. The AI appears where the work already happens — the inbox, the CRM, the WhatsApp line — with a human approving anywhere mistakes cost money.
• Six-to-eight-week increments. Long enough to ship something real, short enough that sponsorship and budgets survive it.

Nothing on that list is clever. That is the point — the projects that fail are usually missing three or four of these basics, not some advanced technique. And if a project does fail, run the honest autopsy: our AI failure postmortem template covers that.

What this means for you

• Treat the 95% number as a warning about process, not a verdict on AI.
• Before starting anything: name the workflow, the number, and the person who owns it.
• Demand the working slice and the evals from whoever builds it — internal team or vendor. The questions to ask a vendor are in our 10-question guide.
• If a pilot has run for a quarter with no measurable movement, kill it publicly and take the lesson. Zombie pilots cost more than failed ones.

Have a workflow in mind and want the failure modes checked against it before you spend anything? Book a 30-minute call. We will tell you honestly which side of the statistics your project is likely to land on — and it is fine if the answer is "do not build this."