On 1 September 2026 Gartner published the results of a survey run from January to April among 1,303 leaders at organisations with annual revenue of at least $50 million. The headline of the press release keeps the failure: only 22% of organisations have successfully scaled AI across multiple business units, or adopted an AI-first approach. The rest of the text describes anything but a retreat: 85% of functional leaders plan to increase spending in 2026, after dedicating an average of 12% of their functional budgets to AI in 2025.
The number that will travel is elsewhere. Gartner singles out a group of high performers, and that group reported positive returns in 81% of its AI initiatives. The number is striking, it is accurate, and it rests on a construction worth knowing before you quote it.
What the survey is worth
Start with the measurement itself. This is a self-reported survey of functional leaders at sizeable organisations. The press release gives neither the breakdown by sector nor the exact wording of the questions. And Gartner sells advisory on managing the value of AI: a study concluding that those who measure do better serves its business. That does not make the figures wrong. It is a reason to separate what the survey establishes from what it recommends.
The 81% is partly in the definition
Here is how Gartner defines its high performers: companies that constantly track the ROI of their AI initiatives, treat AI as a portfolio of value, and regularly assess project performance and reallocate or discontinue underperforming initiatives.
Reread that last clause. A group defined by the fact that it stops its disappointing projects will, among the projects it keeps, show a high share that pay off. Failures leave the denominator by construction. That is survivorship bias, written into the definition of the group itself.
The press release does not say whether discontinued projects are counted. If they are not, the 81% measures the quality of the pruning at least as much as the quality of the projects.
That is no reason to dismiss the number. It is a reason to make it say what it actually says. It does not show that measuring makes projects succeed. It suggests that measuring lets you know which ones to stop, which is a more modest claim, and probably a more useful one.
A comparison that is not one
The press release sets the two groups side by side, and that is where a quick read goes wrong. High performers see positive returns on 81% of their initiatives; low performers did not know the rate of return for 29% of theirs.
These are not two values of the same measure. The first is a share of successes, the second a share of unknowns. The release does not give the low performers' rate of positive returns, so the gap a reader thinks they see between 81 and 29 does not exist in the published data. What the comparison really shows is that the line between the two groups runs through knowing, not through winning.
The simplest number is the most worrying
One figure in the release needs no caveat at all: roughly 11% of organisations are entirely unaware of what their function spent on AI in 2025. Not the return: the spend.
And the whole population is speeding up, since 85% of leaders plan to raise their budgets. Tina Nunno, a Distinguished Vice President at Gartner, puts it plainly: "This lack of financial visibility heightens risk as spending accelerates". You cannot reallocate or stop what you cannot see. The discipline Gartner describes in its top performers therefore starts well before any ROI calculation, with an elementary accounting question: how much, and for what.
We covered the other side of the same problem in companies cutting AI projects without knowing why. Gartner's data confirms that diagnosis from the other end: stopping is not the problem, since top performers stop too. They simply know what they are stopping.
Code generation, the only popular use case that pays
The second half of the release compares, for IT, the most pursued use cases with those most often delivering positive returns. The two lists barely resemble each other.
The most pursued are cybersecurity threat detection and response (54%), IT service desk automation (54%) and automated code generation and refactoring (44%). Those most often delivering positive returns are intelligent IT asset and cost optimisation (40%), synthetic data generation (28%) and, again, automated code generation and refactoring (23%).
Code generation is the only use case on both lists. The other two favourites do not appear among the ones that pay. That is the point Gartner makes: the most widespread use cases are rarely the most profitable, and the value tends to come from less common ones, chosen because they fit a specific need.
For anyone building software with agents, being on both lists does not mean it works everywhere. It means it can work, and that only some of those who try report a positive return. The two percentages answer different questions, and I am not dividing one by the other. But the gap raises a practical question: what separates those who get a return from the rest?
What we do with it
Gartner's answer comes down to three habits: track the return, manage a portfolio, stop what underperforms. For code generation, the first one means attaching the spend to something you can judge. A monthly amount per account lets you stop nothing, because you do not know what to stop. That was already the lesson of Uber burning its annual AI budget in four months. A cost attached to a feature and its version, on the other hand, can be judged.
That is exactly what Survol records, since we steer products built by agents: the cost of each session is attached to the feature and the version it serves. But the question arises with or without us, and the 11% who do not know their own spend show that it is not yet being asked everywhere.
The takeaway
The release will be quoted for its 81%. An honest reading takes something else from it: the difference between organisations that succeed with AI and the others does not show first in their results, it shows in what they know about their results. That is less striking than a success rate. It is also the only part of the number you can copy.
Sources
- Gartner, "Gartner Survey Finds Only 22% of Organizations Have Successfully Scaled AI Across Multiple Business Units", press release, Stamford, 1 September 2026. Survey run January to April 2026 among 1,303 respondents from organisations with annual revenue of at least $50 million in fiscal year 2025. Every figure in this article comes from it. Gartner sells advisory on managing the value of AI.
