McKinsey's own survey: a third of companies are quietly building software instead of buying it -- while enterprise AI's bottom-line payoff hasn't moved in a year

McKinsey's State of AI 2026 survey -- 1,719 respondents across 97 countries, fielded May 4 through June 8 and published August 25 -- landed on a genuinely new number: 32% of organizations now report they've decided against buying at least one software product or feature because agentic coding tools let them build it in-house instead. Coverage of the finding mostly stopped there, with headlines like "the build-vs-buy shift." The same survey, read past that one number, tells a less triumphant story: the share of companies reporting any real bottom-line payoff from AI hasn't moved since last year.

The build-vs-buy number isn't evenly spread. Among the survey's self-identified AI "high performers" -- organizations attributing 5% or more of EBIT to AI and reporting significant value from it, just 6% of all respondents -- 50% say they've skipped a software purchase to build instead, against 31% of everyone else. Technology and healthcare organizations report the pattern most often. That's the group actually capturing value from AI choosing to build rather than buy at nearly double the rate of the pack -- a real signal, not noise, even inside a survey with plenty of noise elsewhere in it.

That top-line 32% average hides a sharp split by how much value an organization is actually getting from AI already -- which is where the more interesting number sits.

Zoom out from the high performers and enterprise scaling is up in the aggregate too: 44% of organizations now report scaling AI agents across at least one business function, versus 38% in 2025. Among large enterprises specifically -- those with more than $1 billion in revenue -- that figure reaches 54%, against roughly a third for smaller organizations. Scaling of coding agents specifically sits lower: about 20% of organizations overall, rising to 31% at large enterprises. Technology (41%), healthcare payers and providers (39%), and professional services and energy (both 38%) lead the scaling numbers by industry.

That gap -- adoption and scaling both genuinely rising, financial payoff flat -- is the actual story here, and it's the one the "build-vs-buy shift" headline leaves out. Eighty percent of respondents say AI has improved their own individual productivity, and half say it's improved their decision-making. Almost none of that is showing up as a bottom-line number companies are willing to report: just 6% describe AI's EBIT impact as significant, identical to 2025 -- and McKinsey finds high performers report AI cost constraints on coding-agent use roughly 3 times more often than everyone else, even as they build more.

Adoption and scaling are both genuinely up. The number that was supposed to follow them -- the bottom-line payoff -- is exactly where it was a year ago.

None of this is free money for the organizations doing the building, either. Roughly one in five report that AI operating costs -- including raw token spend -- now constrain how much they actually use the tools they've adopted, and McKinsey finds high performers hit that constraint roughly three times more often than the rest of the survey. That's the group with the most usage running hardest into the cost ceiling, which is exactly what you'd expect if the constraint is real rather than an early-adopter artifact. (This is the same token-pricing math behind Anthropic's own recent cache-pricing cut on Claude Fable 5.1 -- providers are visibly aware that agentic, high-volume usage is where the bills actually land.) A company weighing whether to build its next internal tool instead of buying one should price that ongoing run cost against the vendor's subscription fee, not just compare the one-time build effort to a purchase order.

The build-vs-buy trend is real, concentrated among the companies already getting value from AI, and worth watching as a leading indicator of where enterprise software spend goes next -- Salesforce, among others, has already restructured its own AI go-to-market around the assumption that customers want agents woven into products they already own rather than new software bought outright. But the number that would actually confirm AI is paying for itself at scale -- the EBIT-impact figure -- says the same thing it said last year. Both facts are true at once, and a survey that only reports the first one is telling half the story.