McKinsey's 2026 survey found 32% of organizations declined at least one software purchase because coding agents could build the thing instead. Among the small group of AI high performers it was closer to half.
That same high-performer group reports cost constraints on coding agents about three times as often as everyone else, because they use them most. The people furthest into the replacement have already found its ceiling.
Only about two in ten organizations have coding agents at enterprise scale. Most of the cancelled purchases were decided against a capability the organization has not operationalized yet.
Thirty-two percent of organizations in McKinsey's State of AI survey decided against buying at least one software product or feature because agentic coding tools let them build it in-house. Not considered it. Decided. A vendor did not get the renewal.
That is a procurement number, and procurement numbers behave differently from adoption numbers. Adoption is reversible in a sprint. A cancelled license is a budget line that gets reassigned, a contract that lapses, and a system somebody on your team now owns until it is decommissioned.
The survey went out to 1,719 respondents across 97 nations, fielded 4 May through 8 June 2026 and published on 25 August. Almost every writeup led with the EBIT figure. The 32% is the one that will still matter in three years.
The people doing it most are the people already rationing it
The build-versus-buy split is not evenly distributed. Nearly half of the organizations McKinsey classifies as AI high performers declined a purchase, against 31% of everyone else. High performers are roughly 6% of the sample, defined by attributing at least 5% of EBIT to AI with significant impact. They are the ones who have actually made the technology pay.
Put that beside a different question in the same survey. One in five respondents says their organization is limiting AI use because of operating costs. Broken out by tool, about one in ten reports cost constraints on chatbots, on agents, and on coding agents respectively.
High performers report cost constraints on coding agents about three times as often as others. McKinsey's own explanation is the obvious one: they hit the constraint because they use them most.
Both findings describe the same population. The group most willing to replace a purchased product with an agent is the group that has already found the point where running the agent starts to hurt. That is not a contradiction, and it is not a reason to stop. It is a preview. Everyone else is making the same decision without the data that group already has.
What actually changed hands
A software purchase is a fixed, negotiated, annual price with a cancellation clause. What replaces it is a metered cost with no ceiling, attached to a codebase your team maintains.
Three things move at once when a purchase gets cancelled, and only the first one shows up in the meeting.
The license fee disappears. That is the number in the business case, and it is real.
The run cost appears somewhere else. Tokens, yes, but mostly the boring parts: an on-call rotation, a dependency upgrade every quarter, a security patch when the library underneath has a CVE, an SSO integration when the identity provider changes. None of that lands in the software budget. It lands in headcount and in the operational load of a team that did not grow.
And the liability transfers back. This is the part with the least discussion and the most consequence. When you buy software you also buy somebody else's SOC 2 report, somebody else's uptime commitment, and somebody else's obligation to produce audit evidence when a regulator asks. Build it yourself and every one of those becomes an internal work item with an owner and a due date.
The sector pattern in the survey lines up with that. Building instead of buying is most common in technology and healthcare, then professional services and energy. The regulated corners of the economy move slowest here, and the usual reading is that they are behind. The more plausible reading is that they are pricing the liability transfer correctly and everyone else is getting it for free on the spreadsheet.
The gap between the decision and the capability
Roughly two in ten organizations report scaling software coding agents across the enterprise. At companies above $1 billion in revenue it is about 31%. In any single business function, agent scaling rarely exceeds 10%.
Meanwhile 44% report AI scaling somewhere in the enterprise, up from 38%, and 40% of large organizations are scaling agents in at least one function, up from 27%. Adoption is climbing steadily. Depth is not.
Which means the 32% is running ahead of the delivery capability underneath it. A third of organizations made a durable sourcing decision on the strength of a tool that four out of five of them have not yet put into production at scale. The decision was made with a demo and a pilot. The consequence arrives as a maintenance obligation.
The rest of the survey says the same thing in a different register. Eighty percent report individual productivity gains. Thirty-seven percent report any EBIT impact from AI, unchanged from last year. The 6% high-performer share is flat too. Individual speed is real and it has not converted into enterprise money, which is exactly the profile you would expect from a technology that makes the first version fast and does nothing about the next four years of the thing existing.
One more number worth holding onto: 39% expect headcount to decline in the coming year, against 14% who saw declines in the year just past. Expectation is running about three times ahead of experience. The build decision and the headcount forecast are being made with the same optimism, and they point in opposite directions, because something has to run the software you built.
What this changes on Monday
If you are heading into a renewal conversation, price the run cost before the cancel decision, not after. Not a token estimate. A named owner, a support rotation, a patch cadence, and whoever is going to write the audit evidence when someone asks for it. If those four lines are empty, you have not compared build to buy, you have compared build to nothing.
Put a budget on the agent and watch whether you hit it. The high performers are hitting theirs. If your coding agent spend is not yet constrained, the useful question is whether that means you are efficient or whether it means you have not scaled far enough to find the wall. Those look identical on a dashboard and they are not the same situation.
Separate the greenfield claim from the maintenance claim. Coding agents are strongest on the first version of something and weakest on a large legacy codebase they did not write. A replacement for a purchased product is a greenfield project for about a quarter and a legacy codebase for the rest of its life. Do not let the first quarter's velocity set the expectation for the other twelve.
And be honest about which decision is reversible. Not renewing is cheap to undo in year one and expensive to undo in year three, once the internal build has users, data and integrations. Buying is expensive every year and cancellable every year. The 32% traded a recurring, capped, exit-able cost for an uncapped one with a lock-in of its own design. Some of those trades will be excellent. The ones that go badly will not go badly for another eighteen months, well after the survey that recorded the decision.