In S&P Global Market Intelligence's 2025 survey of 1,006 IT and business professionals across North America and Europe, the share of companies abandoning most of their AI initiatives rose from 17 percent to 42 percent in one year, and the average organization scrapped 46 percent of its AI proofs of concept before production. The pattern behind those numbers is consistent: initiatives with no owner and no measured business number lose budget reviews by default. The fix starts with a one page inventory, not another purchase.
Before you read on. The Revenue Signal goes out every Thursday to executives who have to defend AI spending to a board: one decision, one verified source, and one move you can finish before the next issue. The Executive AI ROI Scorecard comes with the sign up. Get both here.
Somewhere in your company right now there is an AI subscription renewing automatically that nobody has opened in a month. There is probably also a pilot from last year that everyone remembers as promising and nobody can describe in numbers. Neither of these feels like a problem on any given Tuesday. In aggregate, across the market, they became the most expensive quiet failure of the decade so far. This article covers what the abandonment data actually says, why the projects die, what Klarna's public reversal teaches, and the one page document that separates disciplined portfolios from expensive ones.
S&P Global Market Intelligence runs an annual enterprise survey called Voice of the Enterprise: AI and Machine Learning. The 2025 edition, based on 1,006 midlevel and senior IT and line of business professionals across North America and Europe, produced three findings that belong in any board conversation about AI budgets.
First, the share of companies abandoning the majority of their AI initiatives before production rose from 17 percent to 42 percent year over year. The abandonment rate more than doubled while adoption itself kept accelerating. Second, the average organization scrapped 46 percent of its proofs of concept before they reached production, so nearly half of what got started never reached the people it was built for. Third, and least quoted, organizations reported using more measurement criteria than the year before, 7.1 of the 20 the survey tracks, up from 6.3, while rating fewer of those criteria as very successful, with cost and risk scores worsening.
A scope note, because this data has limits: it is a self reported survey of professionals describing their own organizations, in North America and Europe only. The percentages describe what respondents report, not audited outcomes. Read them as direction rather than decimal, and the direction is unambiguous.
The popular answer is that the technology disappointed. The data points somewhere less comfortable: the projects die at budget reviews, not in test environments.
Here is the mechanism, and it will be familiar to anyone who has sat through a renewal discussion. An AI initiative comes up for continued funding. If it has an owner and a measured number, the room can argue about evidence. If it has neither, the conversation defaults to impressions, and impressions lose to this year's cost pressure. The initiative was never disproven. It was never proven, and unproven loses by default once the invoice is large enough.
The measurement finding in the S&P data supports this reading. Companies are applying more metrics and finding less satisfaction in them, which is what measurement looks like when it arrives after the commitment instead of before it. A metric bolted onto a running project to justify it at review time is an autopsy. A number agreed before the first dollar is a steering wheel. Most of the 46 percent that died between proof of concept and production had the first kind, or none.
Data condition sits underneath all of it. Informatica's CDO Insights 2025 survey of 600 chief data and analytics leaders found data tied with technology as the top obstacle stopping AI from reaching production, each named by 43 percent of respondents. Informatica sells data tools, so treat the exact figure as directional, but the obstacle matches what shows up in practice: initiatives grounded in scattered, undocumented data produce generic output at full price, and generic output is the first thing cut.
In May 2025, Klarna CEO Sebastian Siemiatkowski told Bloomberg the company was hiring human customer service agents again, after a year of being the most publicized AI deployment in business. Klarna's assistant had been handling, by the company's own account, the equivalent work of roughly 700 full time agents. His summary of what went wrong was four words long: we went too far. The company moved to a hybrid model, AI on routine volume, people on everything needing judgment.
Two honest caveats keep this case useful. Klarna did not abandon AI, and its early performance numbers were company reported but never disproven. What the case demonstrates is narrower and more valuable: the original decision optimized for cost without pricing quality, and the correction, recruiting and training people to rebuild capacity that had been let go, was a real expense that no AI replacement business case had modeled. The lesson for a B2B executive is not that AI in service fails. It is that a decision made without a complete number, one that includes the cost of being wrong, is not a measured decision even when it comes wrapped in a dashboard.
The strongest objection to the alarm reading of the 42 percent goes like this: killing weak pilots is exactly what good portfolio management does, so a rising abandonment rate might be health, not failure.
Partly right. A company that never stops anything is not running a portfolio. The distinction that decides the question is what triggered the stop. Stops driven by a measured number coming in low are discipline. Stops driven by cost pressure after the money is spent, on projects nobody ever measured, are waste using the vocabulary of discipline. The survey data cannot separate the two kinds inside your company, and neither can this article. One document can: an inventory stating, for each initiative, what it was supposed to return and what it returned. Companies that can produce that document may treat their abandonments as hygiene. Companies that cannot should assume they are inside the 42 percent, funding the next casualty now.
The inventory is one page, and the first draft takes under an hour. For every AI initiative, tool, and subscription in the company, record four things:
Then mark every row M or U. M means measured: the business number exists and someone checked it in the last 90 days. U means unmeasured. The rules that follow are mechanical. Any U row facing a renewal inside 90 days goes on a decision list, and the default for a U row is stop it or start measuring it, never renew it as is. Rows you cannot fill in at all are not gaps in the exercise. They are the finding.
You can do this on paper, or use the version we built for it. The AI Initiative Inventory takes the four columns above, marks each row measured or unmeasured, builds the decision list of unmeasured initiatives facing a renewal, and gives you the file to download. It runs in your browser, nothing you enter is stored or sent anywhere, and it is free to use.
Run it once and the inventory usually says the same thing the market data says: the company needs less AI than it has, connected better than it is, with each surviving initiative carrying an owner and a number. That conclusion costs a page of writing. The alternative conclusion, reached the other way, costs whatever your share of the 46 percent turns out to be.
What percentage of AI projects fail to reach production?
In S&P Global Market Intelligence's 2025 survey, the average organization scrapped 46 percent of AI proofs of concept before production, and 42 percent of companies abandoned the majority of their AI initiatives, up from 17 percent a year earlier.
Does a high abandonment rate mean AI does not work?
No. It means most initiatives are started without an owner and a measurable business number, so they cannot win budget reviews. The same survey shows adoption rising alongside abandonment, which is consistent with an execution and measurement problem rather than a technology one.
Should we pause AI investment until the market matures?
Waiting has an unmeasured cost of its own, and the companies that keep initiatives alive are the ones that measure them. The practical middle is to inventory what you already fund, stop or fix the unmeasured initiatives, and attach a number to anything new before the first dollar.
How is this different from what a consultant's strategy deck delivers?
The inventory is yours to build and free to run. Where outside help earns its fee is in measuring what internal teams cannot see from inside: real spend against real return, actual presence in AI generated answers, and the true condition of the data layer.
Find out where your company actually stands
The Executive AI ROI Scorecard is a free 20 statement self assessment across four areas: AI spend and return, AI visibility, data condition, and workflow change. It takes about five minutes and turns the inventory above into a calibrated score. Get the Executive AI ROI Scorecard.
Next Thursday. The Revenue Signal takes apart one decision like this every week, with one verified source and one move you can act on before the next issue. It is free, and the Executive AI ROI Scorecard comes with the sign up. Subscribe and get the scorecard.
Figures above are the sources' own reported findings, not Revenue Experts client results. Survey data is self reported unless stated otherwise.
