A Gartner survey released in September 2026, covering 1,303 functional leaders at organizations with annual revenue of at least $50 million, produced a number that cuts against the dominant AI narrative: only 22% of organizations have successfully scaled AI across multiple business units or adopted a truly AI-first approach. The other 78% remain stuck at the pilot stage — proving value inside one team, unable to repeat it at scale, even as 85% of leaders plan to increase AI spending this year.
What separates scalers from pilots isn’t tool volume
The most telling number in the survey isn’t the 22% — it’s what separates that group from everyone else. High-performing organizations reported a positive return on 81% of their AI initiatives. Low performers, by contrast, simply don’t know the return on 29% of the projects already running. It’s not that those projects failed — it’s that no one measured whether they worked. The gap between the two groups isn’t technical, it’s management discipline: companies that scale treat every AI initiative as a portfolio position, not as a one-off bet that keeps running on budgetary inertia once approved.
Investment portfolios have a discipline AI projects rarely get
A portfolio manager never approves a position without three things: an expected-return thesis, an objective exit criterion, and a recurring review that stacks that position against every other one, competing for the same capital. Most AI projects inside a company have none of the three. Someone approves the pilot, the pilot rolls into next quarter’s budget out of habit, and the question “is this still worth what it costs?” never makes it onto the agenda — because no ritual forces that question to surface. The project doesn’t die because it failed; it simply never gets compared to anything.
Four practices of companies that treat AI as a portfolio
None of these require new tooling — they require a management ritual most companies still don’t have:
1. A return metric defined before the project starts, not after. If the metric only gets picked when someone asks “how’s it going,” it tends to be whichever one makes the project look good — not the one that actually measures what matters to the business.
2. A recurring review with every initiative side by side. An AI project evaluated in isolation, out of context, almost always looks reasonable. Placed next to six others competing for the same budget and the same leadership attention, it becomes clear which ones actually deserve to continue.
3. A stop criterion written before the project starts. Deciding what justifies canceling a project after it has already consumed months of investment is deciding under bias — no one wants to admit the bet didn’t work. The criterion needs to exist beforehand, while it’s still a cold decision.
4. Active reallocation of whatever gets freed up. Stopping a project without redirecting the freed budget and team toward the initiative that’s actually working wastes half the value of the decision. Portfolio discipline exists to concentrate resources on what generates return, not just to cut what doesn’t.
This decision doesn’t belong to the data team
In the view of João Paulo Batistella, an innovation executive with a background at Ericsson, Telefônica/Vivo, and five years as CEO of EISA, deciding how AI enters the business model is a CEO-and-board matter, not an isolated technical implementation — when AI is treated only as an IT tool, the company risks optimizing a process that maybe shouldn’t keep existing in that form at all. The same logic applies to portfolio discipline: comparing initiatives, deciding what stops and what gets more resources, is a capital-allocation decision, not a technical-architecture one. A data team can instrument the metric; only whoever owns the business outcome can decide, with real authority, what leaves the portfolio.
Follow Eleva Tecnologia for more analysis on technology applied to business: follow @ElevaTechno on X, @elevatechnologies on Instagram or Eleva Technologies on LinkedIn, or learn about the group at elevatec.net/about.
