42% of companies are giving up on their AI projects. Here’s the pattern.

Somewhere between the kickoff meeting and today, a lot of companies quietly gave up. Not on AI in general — on the specific projects they funded, staffed, and announced. According to S&P Global Market Intelligence's 451 Research group, in a survey of roughly 1,000 respondents, the share of companies abandoning the majority of their AI initiatives jumped from 17% to 42% in just one year. That's not a handful of stragglers walking away. That's a plurality of the market abandoning most of what it started.

It's worth being precise about what that number does and doesn't say, because it gets misquoted constantly. It is not "42% of AI projects fail." The real finding is sharper and, honestly, more alarming: a fast-growing share of companies are abandoning the majority of their initiatives — whole portfolios, not stray experiments.

We've written before about a related but distinct failure mode. In Why 95% of AI pilots never reach the P&L, MIT NANDA's research showed most GenAI pilots quietly produce no financial impact and fade out. This is the harsher sequel. These aren't pilots fading into irrelevance — they're projects getting cancelled outright, after money was already spent.

The scrapping happens before production, not after

The same 451 Research survey explains a lot of the 42%: the average organization scrapped 46% of its AI proof-of-concepts before they ever reached production. Not after a disappointing quarter in the field — before the thing was ever put in front of a real workflow. Roughly half the work an organization does on AI dies in the lab.

That's the setup for abandonment at scale. A company runs a dozen proof-of-concepts with no consistent way to judge them. Half die on review before anyone measures anything. The survivors rarely have real measurement built in either. Eventually leadership looks at the graveyard, counts how much of the year's AI spend produced nothing provable, and decides to cut the majority — not because AI failed, but because nobody could show it worked.

The pattern, once you've seen it a few times

Three things line up in almost every abandonment we'd expect to see behind these numbers:

• No baseline. Nobody measured the process before AI touched it, so there's no honest before-and-after to defend the project when budgets tighten.
• No owner. The project belongs to whoever was enthusiastic in the kickoff meeting, not to the workflow itself — so when that person moves on, nobody is left to defend it.
• Pilot sprawl. Companies run five or six proof-of-concepts in parallel instead of one at a time. No single pilot gets the scrutiny, or the resources, to actually prove itself before the next budget cycle arrives.

None of these are AI problems. They're ordinary project-management failures that AI happens to expose at scale, because AI initiatives multiply faster than most teams are set up to govern.

Wide adoption, thin proof

McKinsey's "The State of AI in 2025" adds the number that makes the whole picture click into place. 88% of organizations report regular AI use in at least one business function — adoption is nearly universal. But only 39% attribute any EBIT impact to AI, and most of those put that impact below 5%. Widespread use, thin proof. That gap is exactly the soil abandonment grows in: when usage is high but nobody can point to the line on the P&L it moved, the project is one budget review away from getting cut, regardless of whether it was quietly working.

More proof, not more caution

The fix isn't to run fewer AI experiments out of fear of joining the 42%. It's to run fewer at once, and make each one earn its keep. Pick one workflow. Measure it honestly before you touch it. Give the pilot an owner who will still have the job in six months. Let it prove itself against that baseline instead of against a vibe. It's a smaller, slower-looking start than most companies attempt — which is exactly why the companies who do it don't end up in the abandonment column a year later.

That's the shape of how we work: an audit that tells you which workflow is actually worth piloting, and which ones aren't, followed by a single measured pilot that either proves itself against a real baseline or gets killed on purpose — deliberately, not by attrition.

---

Source: S&P Global Market Intelligence / 451 Research, "Voice of the Enterprise: AI and Machine Learning, Use Cases 2025"; McKinsey, "The State of AI in 2025" (November 2025). Ready to run one pilot instead of six? Start with the AI Opportunity Audit or go straight to a Pilot Sprint.