Hype Index: 'Over 40% of Agentic AI Projects Will Be Canceled by 2027'

TL;DR: Gartner's June 2025 prediction that over 40% of agentic AI projects will be canceled by end-2027 is the rare AI claim that survives an audit and might even be conservative: ordinary software projects were canceled at a 31% rate back in 1995, and every 2026 measurement of agent deployments points the same direction. The noise is in the retelling: of the top 20 results citing it, 8 distort it in the headline, and nobody quotes the more damning number from the same press release, that Gartner estimates only about 130 of the thousands of "agentic AI" vendors are real. We score it 35% noise / 65% signal, the most signal this series has ever awarded.

What is the claim?

Verbatim from Gartner's June 25, 2025 press release: "Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls." Gartner analyst Anushree Verma's diagnosis came attached: "Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied."

Fourteen months later this is the most-cited sentence in agent skepticism, deployed both as proof the category is doomed and dismissed as analyst theater. Time to rate it.

What is actually true?

More than Gartner's critics want, for three reasons. First, the base rate makes 40% unremarkable. The Standish Group's famous CHAOS report measured 31.1% of ordinary software projects being canceled before completion in 1995, in the pre-AI, pre-hype era of boring enterprise software. Predicting 40% cancellation for projects that are, by the analyst's own description, hype-driven proofs of concept is barely a prediction at all; it is regression to a well-documented mean plus a hype tax.

Second, every 2026 measurement rhymes with it. Forrester's June 2026 state-of-agentic assessment found "three-quarters of enterprise leaders tell us they're adopting agentic AI. Only a small minority have it running in meaningful production beyond 'agentish' chatbots." Teradata's survey has 78% of enterprises piloting agents and 14% scaling one. Gartner's own April 2026 survey found 20% of AI projects in infrastructure and operations failing outright. A pipeline shaped like that sheds 40% of itself without anyone noticing, and the frameworks underneath do not help: of 17 whose multi-agent docs I read, three cap the cost and none can undo an action already taken.

Third, the buried lede from the same press release is worse. Gartner estimated "only about 130 of the thousands of agentic AI vendors are real," coining "agent washing" for the rebranding of chatbots and RPA as agents. The cancellation prediction got the headlines; the vendor arithmetic, thousands claiming, ~130 real, is the stronger indictment and almost never travels with it. It is the supply-side of the demo-to-production gap: a market where most sellers are relabeled automation will, definitionally, generate canceled projects.

What is not true?

Two things, both about how the number is used rather than the number. It is a prediction, not a measurement, and it has no disclosed methodology; one 2026 critique searched the primary document for the survey or calculation behind the 40% and reports finding none. Fourteen months on, no agent-specific cancellation data has been published against it, by Gartner or anyone, so treating it as settled fact is exactly the zombie-stat lifecycle at an earlier stage.

And the retelling is already mutating it. My audit of the top 20 results citing the prediction: 18 of 20 get it right in body copy, an unusually good ratio for this genre, but 8 of 20 distort it at the headline, six swapping "canceled" for "fail" (a different claim; a canceled pilot is a decision, a failure is an outcome) and two rendering it in the present tense, as if the 40% were happening now. Two more misstate where it came from, one attributing it to a webinar poll of 3,412 attendees that measured investment levels, not cancellations. The claim is healthy; its citations are already growing the tumors.

What should you do instead?

Read it as a base rate, not a verdict. If you are running an agent project, 40% cancellation odds are the industry prior, and the escape moves are the documented ones: a defined owner, a scoped workflow with a human checkpoint, and a stop condition, which is precisely the configuration the surviving 14% in Teradata's data share. If you are evaluating vendors, use the other number: ask any "agentic AI" seller what distinguishes their product from a chatbot with an API key, because Gartner's arithmetic says the odds are literally thousands to 130 that the answer is marketing.

Verdict

A prediction with no published math, already being distorted by its citers, that is nonetheless directionally supported by every measurement taken since, anchored by a historical base rate that makes it look tame, and paired with a vendor-count stat that deserves the fame it never got.

Verdict: 35% noise / 65% signal. The first claim this series has rated where the responsible correction is: it is probably worse than Gartner said.