Unconfirmed: TechCrunch wrote on 14 November that S&P Global Market Intelligence says about 42% of AI initiatives are ultimately abandoned by their corporate parents1, a figure Automatica has not confirmed.
The sentence is the whole of the sourcing. TechCrunch lists insufficient funding, technical challenges, competition, difficulty scaling and weak user demand as reasons. It does not name the S&P report, link to it, or describe the method, sample or period. In the material Automatica reviewed, neither S&P nor TechCrunch has said more about the number.
The scope matters. TechCrunch introduces the figure as covering AI bets that "never even survive inside the companies that build them". That separates it from the startups the article opens with. Character.ai filed for bankruptcy in early October, and Ollie said it would shut on 6 November. Ollie's chief executive, Bill Lennon, cited being unable to stand out, failing to grow and not raising enough capital. Both are independent companies, so neither illustrates the figure or its denominator.
The definition of "abandoned" is the open question. The list around the statistic covers products that "shut down, pivoted, or significantly missed expectations". That is TechCrunch's own scope and says nothing about how S&P drew its line.
The corporate cases in the list show how blurry the line is. TechCrunch says OpenAI discontinued the stand-alone ChatGPT Atlas browser on 9 August and absorbed its most useful features into ChatGPT. The product ended but the capability continued. Whether S&P would count that as abandonment is unknown.
The time frame is also missing. "Ultimately" implies an end-state share, perhaps over a project's whole life, but no window is given. The figure therefore cannot be compared with an annual failure rate. It also does not say whether the base is pilots, proofs of concept or shipped products.
What would confirm it: S&P publishing the underlying report, or TechCrunch pointing to it, with five details. Those are the definition of "abandoned", the unit counted (projects, pilots or products), the sample size, the company types and regions covered, and the period measured.
What would undercut it: a report that gives a different number, or a narrower population than TechCrunch's wording implies. A statement from S&P that its data was mischaracterised would do the same. A TechCrunch correction or clarification would also settle it.
Automatica has no earlier reporting on this figure. TechCrunch is the only source in hand, and a second-hand citation is not corroboration. Until the S&P document surfaces, the 42% is TechCrunch's attribution, not a measured rate.