Skip to content
AI Board

Blog

Why Most AI Projects Fail: Every Real Number, With Sources (2026)

Gartner, MIT and S&P Global on AI project failure: what each number actually measures, with sources, and the finding everyone misses about who wins.

Your board signed off on an AI programme this year. Maybe you sponsored it.

Now read what the analysts your board actually trusts are publishing about programmes exactly like it.

Gartner expects over 40% of agentic projects to be canceled. MIT measured almost none of the pilots moving the P&L. S&P Global watched abandonment more than double in a single year. These are not doom-bloggers. They are the same firms whose slides sit in your board pack.

Here is every real number, what each one actually measured, and the links to check it yourself. Then the part almost nobody quotes: the finding that decides which side of the gap you land on.

The numbers, honestly

Most of what circulates on LinkedIn is a mangled version of the stats below. "95% of AI fails" is not what MIT said. "40% of AI is dead" is not what Gartner said. The real claims are narrower, better sourced, and somehow more alarming for it.

Source (date)What it actually measuredThe number
Gartner, Jun 2025Agentic-AI projects it predicts will be canceled by end of 2027Over 40%
Gartner, Jul 2024, updated Apr 2026GenAI projects abandoned after proof of concept≥30% predicted → ≥50% actual
MIT NANDA, Aug 2025Enterprise GenAI pilots with no measurable P&L impact~95%
S&P Global, 2025Companies that abandoned most of their AI initiatives42% (up from 17% in 2024)
RAND, 2024AI projects that fail, vs comparable non-AI ITOver 80% (about 2×)

Two of those numbers are not snapshots. They are trend lines, and both point the same way. Read this one as a chart:

What worsened, year over yearThenNowTrend
Companies abandoning most AI initiatives (S&P Global)17% (2024)42% (2025)
GenAI projects abandoned after proof of concept (Gartner: forecast → measured)≥30% (2024)≥50% (2025)

Gartner predicted in 2024 that at least 30% of generative-AI projects would be abandoned after proof of concept. In April 2026 it reported the real figure came in at least 50%. The prediction did not just come true, it came true worse. Over the same window, S&P Global watched the share of companies abandoning most of their AI initiatives climb from 17% to 42%. The direction is not improving. It is accelerating.

Is the 95% number real? Here's the pushback.

The MIT "95%" is the most-quoted and most-abused figure of the lot, so treat it honestly. It comes from MIT's NANDA initiative, "The GenAI Divide," built on 150 leader interviews, a 350-person survey and 300 public deployments. It measures pilots with no measurable P&L impact, not "95% of AI projects fail."

And it is contested. Critics at the Marketing AI Institute and VentureBeat note the sample is small and the definition of "failure" is loose; the same underlying report, read differently, describes a booming shadow-AI economy rather than a graveyard. We link the pushback on purpose. A number you cannot defend under scrutiny is not an asset.

The honest read: whether the true figure is 95%, 80%, or "merely" half, every credible source agrees the failure rate for corporate AI programmes is high, rising, and worse than ordinary IT.

Why they fail

The sources agree on the reasons, and not one of them is "the AI wasn't good enough."

Gartner attributes the abandonments to escalating costs, unclear business value, and inadequate risk controls. RAND, whose report is the single most-cited answer to "why do AI projects fail", points to teams misunderstanding the problem, poor data quality, tech-first thinking, and missing infrastructure. More than 80% of AI projects fail, RAND found, roughly twice the rate of comparable non-AI IT projects.

MIT adds a build-versus-buy finding that stings: buying from specialized vendors succeeds around 67% of the time, while internal builds succeed about a third as often. The instinct to "build our own" is, statistically, the instinct to lose.

And a chunk of the market was never real to begin with. Of the thousands of vendors claiming "agentic AI," Gartner estimates only about 130 are the genuine article. The rest is "agent washing." Some projects fail because they were sold a category that did not exist.

Notice the pattern. Cost, data, governance, scope, vendor theatre: organizational failures, every one. Not a single reason on the list is "the technology cannot do this."

The finding everyone skips

Here is the sentence that gets cut from every summary of the MIT report.

The same study also found that employees at over 90% of companies already use personal AI tools for work, while only about 40% of those companies hold an official AI subscription. The programmes failed. The people did not.

Sit with that. The same organizations posting no measurable P&L impact are full of individuals quietly getting faster, sharper, and more capable on their own. The failure is institutional. The adoption is personal.

Which reframes the whole conversation. The gap that is opening is not company versus company. It is executive versus executive: the leader who works AI-natively against the one who sponsored a programme and waited for a rollout.

If reading this triggered a flicker of "am I already behind", you are in the majority. Korn Ferry found 71% of US CEOs experience imposter syndrome. The feeling is nearly universal at your level. The response does not have to be.

Because the stakes are lopsided. Your AI project failing is survivable; boards forgive a scrapped pilot; per S&P Global they do it 42% of the time. Being the last AI-native executive at the table is not survivable. That is the fear worth having, and it is the one with a clean, small exit.

What to do instead of a programme

If institutional AI is a coin-flip at best, stop starting with the institution.

Start at the top, on the one machine you fully control: your own laptop. Point executive-grade AI at your own company's data. Ask it the questions you would have waited two weeks and three meetings for. Days, not an IT programme. No vendor-selection committee, no six-month rollout, no line item your board can cancel, which, per Gartner, it has a better-than-40% chance of doing.

You are not launching another institutional programme, the kind with the failure rates above. You are doing personally what the data already shows individuals doing: getting AI-native on your own terms.

AI Board is your personal AI assistant that makes you an AI-native executive: an AI CEO, CFO and CTO on your own laptop, grounded in your company's data and getting sharper as your data grows. Private by design, fast, and ahead of the executives who wait.

Be honest about the limit, though. This is not org-wide transformation, and it does not pretend to be. Getting your whole company AI-native still takes the whole company: the change management, the data plumbing, the slow institutional work those RAND and MIT figures describe. What starts on your laptop is one thing only: you, the individual executive, refusing to be the bottleneck while the institution catches up. If you want the fuller picture of that role, this is what an AI-native executive is, and what an AI board member does once your data sits behind it.

That personal adoption is not hypothetical. It is already happening under you. Your team's shadow AI economy is exactly this, without permission. It is happening beside you, since the same squeeze is closing on managers. The only open question is whether it is happening at the head of the table.

FAQ

What percentage of AI projects fail?

It depends on what you measure, and the honest answer is a range. RAND found more than 80% of AI projects fail, about twice the rate of comparable non-AI IT. MIT's 2025 NANDA study found roughly 95% of enterprise generative-AI pilots delivered no measurable P&L impact. S&P Global found 42% of companies abandoned most of their AI initiatives in 2025, up from 17% a year earlier. Different lenses, one direction: high and rising.

Why do most AI projects fail?

Not because the technology is weak. Gartner cites escalating costs, unclear business value, and inadequate risk controls. RAND points to misunderstanding the problem, poor data quality, tech-first thinking, and infrastructure gaps. MIT found internal builds succeed roughly a third as often as buying from specialized vendors. The failures are organizational, not technical.

Is the MIT 95% statistic real?

It is real, from MIT's NANDA "GenAI Divide" report, but it is narrower and more contested than the headline suggests. It measured generative-AI pilots showing no measurable P&L impact, not "95% of AI projects." Critics note the small sample and a loose definition of failure. Quote it precisely, or not at all.

Should we still start an AI project?

Yes, but be clear-eyed about which kind. Large institutional programmes carry the failure rates above. An individual executive working AI-natively on their own data is a different, smaller, faster bet that does not depend on an org-wide rollout succeeding. Start where the odds are in your favor, at your level, and let the institution catch up.


Every number above is public, sourced, and checkable, because the executives who act on the real figures will be reading them long before the ones waiting for the post-mortem. See what an AI-native executive does with their own company's data →, and meet your AI CEO.

Become AI-native before your competition does

Ride the AI wave instead of swimming behind it. Join the waitlist and we'll email you when your seat is ready.

Runs on your own laptop. Your data never leaves it.