3 SEP 2026 — McKinsey's State of AI 2026 survey reports that 32 per cent of organisations have decided against buying an off-the-shelf software product because they could build it themselves with agentic coding tools. The figure is being read as the beginning of the end for enterprise software. The same survey says the share of organisations where AI contributes to operating profit is unchanged from a year ago, and that one in five now feels financial strain from what AI costs to run.

What the survey reports

The build decision is not evenly spread. It is 41 per cent in the technology sector, 39 per cent among healthcare payers and providers, and 38 per cent in professional services and energy. Among organisations with more than a billion dollars of revenue, 40 per cent are now scaling agents in at least one function, up from 27 per cent.

McKinsey defines high performers as respondents attributing at least 5 per cent of operating profit to AI. They are 6 per cent of the sample, and nearly half of them are skipping software purchases against 31 per cent of everyone else.

Set against that, the share reporting that AI has contributed to operating profit is 37 per cent, essentially where it was a year earlier. About 20 per cent say AI operating costs are starting to constrain how much they use it.

32%Skipped at least one software purchase to build instead
37%Report AI contributing to operating profit, flat year on year
6%Meet the survey's definition of a high performer
20%Say AI running costs are now constraining their use of it

Two findings from one survey that point opposite ways

A third of organisations have concluded they can build what they used to buy. In the same population, the proportion where AI reaches the profit line has not moved in twelve months.

The two findings do not contradict each other, and the way they fit together is uncomfortable. One is a judgement made at the keyboard about what it costs to produce software. The other is a measurement taken a year later of what that software earned. Production has got faster; the earnings have not moved.

Which is what you would expect if writing the software was never the constraint on what it was worth, and nothing in the survey suggests it was.

The high performer definition sets a low bar

Attributing at least 5 per cent of operating profit to AI is the threshold for the group everyone will want to imitate. That threshold is modest. A company can clear it while 95 per cent of its profit comes from everything it was already doing.

Only six per cent of respondents clear that bar. So when the survey reports that nearly half of high performers are skipping software purchases, it is describing what is happening at roughly 3 per cent of the companies surveyed.

The attribution is also self-reported. No survey of this kind can verify that a company's AI programme caused the profit it is credited with, and the respondents most enthusiastic about AI are the ones most likely to attribute profit to it. We made the same point in July about the reports being cited in the forward deployed engineer market, and it applies to every self-attributed return figure in this category.

Build is a decision about capital cost, not running cost

Agentic coding tools change the price of the first version. Everything after that version costs what it always did, and afterwards is where enterprise software spends its money.

Buy a product and somebody else carries the maintenance, the security patching, the compliance attestations, the compatibility work when the operating system moves, and the contractual obligation to fix it when a dependency breaks. Build it and all of that arrives on your own payroll, at a headcount cost that never appears in the decision this survey measured.

The 20 per cent already reporting strain from AI operating costs is the first sign of this arriving. Inference is metered, an internal tool that calls a model on every request has a bill that scales with use, and a licence that used to be a fixed annual number becomes a variable one that grows with adoption.

What this does not tell you about the software market

Skipping a purchase is a decision about one product at one moment. The survey asks whether an organisation has done it at least once, which is a low threshold to cross and says nothing about how much of the portfolio moved.

A company that built its own internal expenses tool and kept its enterprise resource planning system, its customer relationship platform and its data warehouse has answered yes. So has one that rebuilt everything. Those are different worlds and the question does not separate them.

The sector split hints at which is more common. Technology firms lead at 41 per cent. That is the sector with the most engineers and the most confidence about maintaining what it writes, which suggests the trend is less about which software is replaceable and more about who can absorb the running cost.

The question worth asking before the next build

Price the second year. What will this cost to run at full adoption? Who patches it? Who is on call for it? What happens when the person who prompted it into existence changes jobs?

A first version that arrived in a fortnight answers none of those. The evidence in this survey is that the organisations answering them well are a small minority, and the majority's operating profit has not yet noticed the difference. We looked at the same gap from the task side, where adoption is broad and shallow at once. This is the procurement version of it.