McKinsey's State of AI global survey for 2026 contains one finding that will be read in every software boardroom this month: 32% of organisations say they have decided against buying at least one software product or feature because they could build the functionality internally with agentic coding tools.
Not evaluated. Not delayed. Decided against.
The industry split says more than the headline. Technology leads at 41%. Healthcare payers and providers are at 39%, professional services and energy and materials both at 38%, financial institutions 36%, media and telecom 34%, pharmaceuticals and medical products 33%. This is not a Silicon Valley affectation leaking into a survey. It is a range of eight to nine points across every major sector, which is the shape of a general shift rather than a pocket of enthusiasm.
The self-selection is sharper still. Among McKinsey's “high performers” — the roughly 6% of respondents who attribute at least 5% of EBIT to AI — close to half report skipping software purchases, against 31% of everyone else. The companies getting the most measurable value from AI are the ones most willing to stop buying software.
The number underneath the number
Here is the part that a vendor's investor relations team should be reading carefully, and the part a buyer should be reading even more carefully.
Agentic coding tools have collapsed the cost of version one. They have not touched the cost of version two through version forty. Maintenance, reliability, security patching, permissions models, workflow integration, data quality, on-call rotation, and the sheer institutional memory of who knows why the thing works — none of that got cheaper. Industry rules of thumb put ongoing maintenance at roughly 15–25% of build cost annually, which over a typical software lifespan makes upkeep a multiple of the original build.
So the licence line item disappears from the software budget and reappears, larger and less visible, in headcount, security and on-call. A saving that moves between cost centres is not a saving. It is a reclassification, and it usually gets discovered at the eighteen-month mark by whoever inherits the system.
Our take: The correct question is no longer “can we build this?” — the answer is now yes for a much wider class of software than it was two years ago. The question is “who owns this in three years, and what is that person's fully loaded cost?” Build where the software encodes something genuinely proprietary about how you operate, because that is where a vendor was never going to fit you well anyway. Buy where the product is commodity infrastructure with a compliance surface — payroll, identity, payments — because you are not buying features there, you are buying somebody else's obligation to keep it working and audited at 3am.
What this does to SaaS pricing
The immediate effect is not mass churn. It is leverage at renewal. A buyer who can credibly say “we scoped building this and it is two engineer-months” negotiates differently from one who cannot, whether or not they ever intend to build it. Expect that to show up first in seat-count reductions and discount depth rather than logos lost.
The medium-term effect falls unevenly. Thin-wrapper products — a form builder, an internal dashboard tool, a lightweight approval workflow — are genuinely exposed, because their moat was always that building them was annoying rather than hard. Products whose value is a regulated data network, an integration estate nobody wants to rebuild, or a liability someone else carries are far more defensible than the 32% headline suggests.
What to watch
- Net revenue retention in software earnings. If this trend is real it appears there before it appears in customer counts.
- Internal engineering headcount at the companies doing the building. If it rises alongside skipped purchases, the cost simply moved.
- Whether “we built it” survives its first security incident. That is the honest stress test, and most of these builds have not had one yet.
- Repeat-buyer behaviour next renewal cycle. Skipping one purchase is an experiment. Skipping the same category twice is a policy.
A third of companies have run the experiment. Almost none of them have yet paid the full bill for it. Both things are true, and the second one arrives later.