McKinsey published its annual “State of AI” survey on August 25, fielded May 4 through June 8 across 1,719 respondents in 97 countries. Buried inside a report mostly about the gap between AI adoption and financial return is a number that should matter a great deal to anyone tracking whether agentic coding is actually changing how software gets made, not just how it gets written: 32% of organizations report they decided against buying at least one software product or feature because they could build it internally with agentic coding tools instead.
That’s not a productivity statistic. It’s a procurement statistic, and procurement decisions are a lagging indicator that something has structurally shifted — nobody skips a software purchase on a whim.
The build-vs-buy number is the real story#
For two years, the agentic-coding pitch from Anthropic, this blog included, has rested on a claim that’s easy to make and hard to verify: AI agents are good enough now that the old make-or-buy calculus for internal tooling has changed. McKinsey’s survey is the first large, methodologically serious data point that puts a number on that claim rather than a vendor case study or a vibes-based trend piece.
Nearly a third of enterprises walking away from a SaaS renewal or a build proposal because Claude Code, Cursor, Copilot, or whatever their agent of choice can produce the equivalent in-house is a genuinely disruptive signal for the software industry’s business model, not just for how engineering teams spend their day. Every category of software with thin moats and high per-seat pricing — internal dashboards, workflow automation, bespoke integrations, low-differentiation SaaS — is now competing against a marginal cost that keeps falling toward zero.
The adoption curve behind that number is climbing fast, too: 40% of respondents at large organizations (annual revenue above $1 billion) report scaling AI agents in at least one function, up from 27% in last year’s survey. Software coding agents specifically are already being scaled by roughly three in ten larger enterprises — not piloted, not evaluated, scaled. That’s a 13-point jump in a single year on the “are we actually doing this at scale” question, which is usually the slowest-moving number in any enterprise technology survey.
The honest counterweight: EBIT impact didn’t move#
Here’s the part a vendor blog post would bury and this one isn’t going to: the survey’s headline financial number is flat. Only 37% of respondents attribute any EBIT impact to AI use at all, unchanged from 2025. Just 6% qualify as “AI high performers” — organizations attributing 5% or more of EBIT to AI with a clearly significant effect — also flat year over year.
Put those two findings next to each other and you get an uncomfortable but honest picture: adoption and behavioral change are accelerating (more scaling, more build-vs-buy substitution, more agents doing real work) while the bottom-line financial proof hasn’t caught up at all. That’s not a contradiction, it’s a timing gap, and it’s the same gap this blog has flagged before — Stanford’s 26% productivity gain running headlong into METR’s 19% slowdown finding on complex codebases, or the 5.5% financial-ROI figure from McKinsey’s own 2025 survey. Individual and team-level productivity gains are real and increasingly well documented. Enterprise-level P&L impact is still mostly a promise.
One more data point worth sitting with: only 14% of respondents say AI contributed to an overall workforce decline over the past year, versus the 32% who expected workforce reductions in last year’s survey covering the same period. The mass-layoffs-via-agentic-coding narrative that dominated tech Twitter for most of 2025 hasn’t shown up in this year’s actual outcomes data. Teams appear to be redeploying capacity into more building, not just doing the same building with fewer people — which is consistent with the build-vs-buy number above.
Why this matters more than another benchmark#
This blog spends most of its time on model releases, Terminal-Bench scores, and changelog entries — the technical layer of whether Claude Code, Cursor, or Copilot got objectively better this week. McKinsey’s survey is a reminder that the more consequential question is organizational, not technical: are the people who control software budgets actually changing their decisions because of what these tools can now do?
The answer, per the largest AI-adoption survey McKinsey runs annually, is yes, and it’s happening faster than last year. The caveat is that “we decided to build it ourselves” and “it showed up as measurable profit” are two different claims, and only the first one has moved this year. Anthropic’s own sales pitch — Claude Code turns engineering orgs into build shops rather than procurement shops — has real survey evidence behind it now. Whether that translates into the EBIT line is next year’s number to watch, not this year’s.
What to watch for: whether McKinsey’s 2027 survey shows the EBIT-impact number finally moving in step with the build-vs-buy and scaling numbers, or whether the gap between “we’re doing more of this” and “it’s making more money” persists into a third year.
Sources: The state of AI in 2026: On the road to ROI — McKinsey & Company (August 25, 2026); prior coverage referenced: The Economics of AI-Assisted Development (this blog, June 13, 2026).
