A Third of Companies Now Build the Software They Used to Buy — and McKinsey's New Survey Explains Why the Bill Doesn't Shrink

For two decades, enterprise software lived on a simple proposition: your problems are common enough that someone else has already solved them, and renting their solution is cheaper than building your own. McKinsey’s 2026 State of AI global survey, published August 25, contains a number that puts that proposition on notice. Nearly a third of respondents — 32 percent — report that their organizations have decided against buying one or more software products or features because they could be built internally with agentic coding tools.

That single sentence has travelled further than anything else in the report, and for good reason: it is the first large-scale, credibly sourced data point showing that AI coding agents are not just accelerating developers but reshaping how technology budgets get allocated. When a company declines a vendor’s quote because a coding agent can produce the functionality in-house, the deal that dies is not a lost discount — it is the assumption underneath an entire industry built on selling software by the seat.

The number, and the fine print

The survey drew 1,719 responses across 97 countries between May 4 and June 8, weighted by each nation’s contribution to global GDP. The headline finding, in the report’s own careful wording: “Coding agents are emboldening companies to build their own software rather than purchase it. Nearly one-third of respondents (32 percent) report that their organizations have decided against purchasing at least one software product or feature because they were able to build the functionality in-house using agentic coding tools.”

Two pieces of context matter before anyone redraws a roadmap. First, the unit is one product or feature, at least once. A company that built a single reporting dashboard instead of buying a seat licence counts exactly the same as one that replaced its customer platform. Second, McKinsey provides no prior-year figure for this question, so there is no trend line yet — only a baseline that is already remarkably high.

The distribution is the more telling part. Among AI “high performers” — organizations that attribute at least 5% of EBIT to AI and describe the impact as significant — nearly half have declined a purchase to build instead, against 31% of everyone else. The behavior concentrates where AI capability is deepest, and it shows up most often in technology and healthcare, followed by professional services and energy.

Building rose. Benefiting did not.

Here is where the survey stops being a procurement story and becomes something more uncomfortable. Every measure of doing more with AI went up year over year: enterprise-wide scaling rose from 38% to 44%, large organizations scaling AI agents jumped from 27% to 40%, and individual productivity gains are now reported by 80% of respondents. But the one measure of getting paid for it stayed still: 37% of organizations attribute any EBIT impact to AI, which McKinsey describes as “essentially unchanged from a year ago,” and the high-performer share is flat at about 6%.

McKinsey’s own summary is unusually direct: “Organizations’ conviction in AI is growing faster than the immediate financial returns they can attribute to it.” The productivity gains employees feel at their desks “have yet to translate into broad financial impact for organizations.”

The build-instead-of-buy decision sits precisely inside that gap. Declining to buy a product because you can build it is an act of capability, not of return. Whether it pays depends on what happens next: whether the thing you built replaced a real cost, whether it changed how work is done, and whether anyone maintains it in year two. The survey cannot see any of that, and to its credit, it does not claim to.

What the 6% do differently

The report devotes its longest section to the small group of companies actually making money from AI, and four differences stand out — none of which is a technology choice.

They redesign the workflow, not just the tooling. High performers “fundamentally redesign workflows that are enabled by AI rather than insert AI into existing ones.” Building a tool to fit an unchanged process is the pattern the flat 37% describes; changing the process and then building for it is the pattern the profitable 6% describes.

They build more, and buy less. High performers are twice as likely to be scaling coding agents and 2.7 times more likely to be scaling other agentic AI. Nearly half have declined a software purchase because they could build it — the build decision correlates with returns, but only when it follows the redesign.

They aim past efficiency. Around 80% of both groups pursue cost reduction, but most high performers also use AI to pursue growth or innovation, and they are 3.3 times more likely to intend a fundamental business transformation within three years. A tool built to cut a cost has a ceiling; one built to sell something new does not.

They pay for it, and manage the risk. High performers are more than twice as likely to spend over 15% of their technology budget on AI, twice as likely to have visibly committed senior leaders and defined measurement processes, and much more likely to be actively working on AI-driven vulnerabilities and unintended actions.

The cost arriving next

The survey also flags the constraint that will define the next chapter of this story. About one in five organizations say AI-related operating costs — which McKinsey specifies as including token costs — have already constrained their use of AI. The report calls this “a meaningful consideration, but not yet a widespread constraint,” with the share broadly consistent across company sizes and industries.

The irony is sharp: high performers report cost constraints on coding agents about three times as often as everyone else, because they use them most. The heaviest builders are the first to discover what vendors were pricing into their licences all along — maintenance, security review, and the migration burden of owning what you run. A purchased product has a price that is known in advance and mostly fixed. A built one has a token bill that scales with use, indefinitely.

That does not argue against building. It argues for pricing the run cost of a built feature the way a vendor would have — before the decision rather than after. Meanwhile, 28% of organizations now spend more than a tenth of their technology budget on AI, and 60% still expect to increase investment next year. Conviction, as McKinsey notes, is running ahead of returns.

The SaaS question nobody can answer yet

For software vendors, the 32% is the warning shot. If a third of your customers can already decline a purchase because an agent can replicate a feature — and nearly half of the most sophisticated ones can — then the moat has to live somewhere other than feature surface area. Compliance, uptime, edge-case coverage, integrations, and liability allocation are the things a coding agent still cannot generate in an afternoon.

For buyers, the useful rule emerging from the high-performer data is deceptively simple: build where the workflow is yours and purchased software would force it into someone else’s shape; keep buying where the function is generic and the vendor’s scale is the actual product — payments, identity, email delivery. And when the temptation is a single feature trapped behind an enterprise tier, build the feature against the vendor’s platform rather than replacing the platform itself.

The bottom line

The 32% figure will be quoted for a year as evidence that software procurement is changing, and it is. But it should be quoted next to the 37% — the share of organizations seeing any earnings impact from AI, which has not moved in twelve months. Capability has become the easy half of the decision. The companies McKinsey holds up as the model did not start by asking what they could build. They redesigned a workflow, set a growth goal rather than only a cost goal, committed budget and leadership, and then found that coding agents let them build exactly what the redesign required. Declining the vendor’s quote was the last step in that sequence. Taken first, it is just a cheaper way to change nothing.