TL;DR: according to McKinsey's "The State of AI in 2026" survey, published on August 25, 2026 and based on 1,719 respondents across 97 countries, 32% of organizations have decided against buying at least one software product or feature because they could build it internally with AI coding agents.
The figure made the rounds of the business press on September 1, 2026 under the "build vs buy" label. It deserves more than a headline. It says something precise about the price of software, about what AI has made easy, and about what it has not made easy at all. For an SME, the right question is not "should we build?" but "build what, and who keeps it running afterwards?".
Key takeaways
- McKinsey surveyed 1,719 participants in 97 countries between May 4 and June 8, 2026: 32% report having skipped a software purchase in favour of internal development assisted by coding agents.
- The pattern is strongest among "high performers" (the 6% of respondents who attribute at least 5% of EBIT to AI): nearly half of them have skipped a purchase, against 31% of their peers.
- SaaS prices are rising at the same time: Vertice's SaaS Inflation Index reached 16.4% in June 2026, measured on more than $75bn of managed spend.
- The counterpoint exists: according to MIT NANDA's "The GenAI Divide" report (2025), AI initiatives built in-house succeed 33% of the time, against 67% for those run with an external partner.
The facts: what the McKinsey survey says
The 2026 edition of McKinsey's "The State of AI", subtitled "On the road to ROI", rests on an online survey run from May 4 to June 8, 2026 with 1,719 participants spread across 97 countries. Two results frame the one that matters here: 80% of respondents see individual productivity gains, but only 37% report any impact on their company's EBIT, a share that is flat year on year.
In between sits the new data point. Nearly a third of respondents (32%) say their organization decided not to buy one or more software products or features because it could build them internally with agentic coding tools. According to the analysis published by Forkast News on September 1, 2026, the practice is most widespread in technology (41%), then among healthcare payers and providers (39%), in professional services and energy (38%), and in financial institutions (36%).
Size matters too. Among large enterprises (more than $1bn in revenue), the share scaling AI agents in at least one function rose from 27% to 40% in a year. Among smaller organizations, that share is flat at around one in five according to the same survey. Lieven Van der Veken, senior partner at McKinsey, sums up the shift in the same article: "Leaders are asking what their organisations need to build AI tools themselves. The rise of software coding agents and in-house development is one clear sign of this broader shift."
Why are companies building instead of buying?
Because two curves crossed in 2026: the cost of producing software is falling, the cost of renting it is rising. On the first curve, coding agents have changed the order of magnitude for one specific category of needs: internal tools, connectors, small line-of-business applications, features a vendor sells as a paid add-on. These are programs with a clear scope, running on data the company already owns.
On the second curve, the numbers are stubborn. Vertice's SaaS Inflation Index, computed on more than $75bn of managed spend, reached 16.4% in June 2026, against general inflation of roughly 2.7% across the G7. In other words, the software bill of a company that changes nothing grows six times faster than its other costs. The deferred purchases McKinsey measured are not a fad; they are a cash-flow response.
McKinsey is not alone, either. Retool's 2026 "Build vs. Buy" report, published on February 17, 2026, already found that 35% of surveyed enterprises had replaced at least one SaaS tool with software built in-house.
What the figure leaves out: the cost of keeping it running
Skipping a purchase is easy to measure; what follows is much less so. An analysis by The Daily Brief dated August 31, 2026 recalls the rule every IT manager knows: writing the code accounts for 10 to 40% of a program's lifetime cost; the remaining 60 to 90% is maintenance, patches, security updates and support. Those expenses do not show up in the budget line of the avoided purchase. They show up elsewhere, later, often once the person who commissioned the tool has moved on.
The strongest counterpoint comes from MIT NANDA's "The GenAI Divide" report (2025): among the AI initiatives studied, those built in-house succeeded 33% of the time, those run with an external partner 67%. The report does not say buy rather than build; it says that a project carried alone, without method or a named owner, fails twice as often. Gartner adds its own caution, quoted in the Forkast article: more than 40% of AI agent projects could be cancelled by the end of 2027.
The table below sums up the three paths a company actually has in front of it, and what each one costs beyond the first month.
| Path | Visible cost | Hidden cost | Right when |
|---|---|---|---|
| Buy a SaaS | Subscription, rising 16.4% a year per Vertice | Unused features paid for, data held by the vendor, processes bent to fit the tool | The need is standard and a market tool covers 80% of the case |
| Build in-house with agents | An employee's time, AI tool licences | 60 to 90% of lifetime cost in maintenance, security and handover when the person leaves | A technical team exists and can own the tool over time |
| Have it built to order | Fixed fee or quote, then contractual maintenance | Dependence on the provider if code and documentation are not handed over | The process is a differentiator and no market tool fits |
What does this change for a French SME?
In practice, an SME with 10 to 100 employees almost never has the technical team that makes the second path viable. McKinsey's 32% mostly describes organizations with more than $1bn in revenue and an IT department. For an SME, the question reads differently: the same coding agents that let a large group do without a vendor let a service provider deliver a custom tool at a price that would not have been possible three years ago. The falling cost of building does not only benefit those who build themselves; it benefits those who have things built.
Three uses draw directly on this shift. First, the paid add-on: when a vendor charges 30 to 50% more for a reporting module, a connector or a client portal, a small application sitting next to the existing tool often costs less than a year of the option. Second, the glue between tools: the software that 32% of companies stopped buying is often exactly this kind of brick, and a well-scoped n8n automation replaces it. Third, the line-of-business tool that does not exist on the market because the process is specific to the company: no SaaS will ever fit there, and that is the case where custom software has always been justified, AI or not.
Judgement, on the other hand, does not change. If a market tool covers 80% of the need, take it and adapt the remaining 20%, not the reverse. If volume is low (a few cases a week), a well-kept spreadsheet remains unbeatable. And if nobody in the company can describe the process to be tooled on one page, the problem is not software.
How this shows up in my own work
Both sides of the McKinsey survey turn up in my assignments. On the "buy" side: for Horus Condition Report, a bilingual inspection company, the right answer was a Pipedrive configured in French and English, not a custom CRM; the need was standard, the market tool covered it, and custom software would have cost more for the same result. On the "build" side: for the 3018 helpline, France's national number against cyberbullying, the counsellors' process (five channels, telephony, hosting in France, sovereign AI) existed in no market CRM, and the Emma CRM was co-built to order for twelve counsellors and managers.
In between, my own back office illustrates the logic of the 32% at the scale of a one-person company: leads, logos, indexing and articles are handled by a tool built with coding agents, where I would once have stacked three subscriptions. The hidden cost exists for me too: it is maintenance, and it is budgeted.
A third of organizations have stopped buying what they can build. The useful question for an SME is not whether to join that third, but to know, tool by tool, which of the three paths in the table costs least over three years. If you have a subscription where you use a single feature, or a process no software covers, a 30-minute call is enough to settle it: book a slot, or read what a custom web app covers.
Frequently asked questions
What exactly does McKinsey's 32% figure measure?
It measures the share of respondents whose organization decided not to buy at least one software product or feature because it could build it internally with AI coding agents. The "The State of AI in 2026" survey ran from May 4 to June 8, 2026 with 1,719 participants in 97 countries. The figure does not say how many tools were replaced or what they cost to maintain.
Should an SME build its software instead of buying it?
Not as a rule. A market tool that covers 80% of the need remains the right choice, and maintenance accounts for 60 to 90% of the lifetime cost of software that is built. Custom software is justified when the process is specific to the company, when a paid add-on costs more than a small application, or when no tool properly connects the existing systems.
Why are SaaS prices rising so much in 2026?
Vertice's SaaS Inflation Index reached 16.4% in June 2026, roughly six times general inflation across the G7. Vendors pass on the cost of AI features, restructure their tiers and charge for more add-ons.