35%, 18.6% or 76%: build vs buy depends on who you asked

Have you ever dropped a stat into a deck, only to have someone ask where it came from and realised you weren't totally sure? You're in good company, and this year's build vs buy numbers show how it happens. Three of them crossed my feed, and from the quote cards alone you'd assume they were arguing. One says a third of teams have already swapped a SaaS tool for something they built, another says fewer than one in five developers build to replace a vendor, and the third says three quarters of enterprise AI is bought rather than built. I don't think any of them is wrong, though, because they asked different people about different things, and somewhere between the report and the repost that detail fell off, even though it's what tells you how to read them.

Three numbers, three rooms

The most shared of the three is 35%, from Retool's 2026 Build vs. Buy Report, which says 35% of teams have already replaced at least one SaaS tool with a custom build, and 78% expect to build more custom internal tools in 2026. From what I saw it travelled furthest among builders and the "SaaS is dead" crowd, which makes sense, since it tells them they're winning.

The second is 18.6%, from the Stack Overflow Developer Survey 2026, which asked developers what they build. Of 13,338 respondents, 18.6% picked replacing paid vendors, and Stack Overflow's own page rounds that to "only 19%", which the skeptics I follow shared with what looked like a bit of relief.

The third is 76%, from Menlo Ventures' State of Generative AI in the Enterprise, where 76% of AI use cases are purchased rather than built internally. That one did the rounds in buyer and investor circles. So depending on whose feed you were reading, everyone is building, almost nobody is, or everyone is buying... and the only way to sort that out is to look at who was asked.

Reading the sample

Retool's survey covered 817 Retool customers and builders in late 2025, so a company asked its own customers and builders whether they build software. I don't think that's a scandal, honestly. Companies survey their customers all the time and it's still useful, but it gives you a picture of people who already build rather than of companies in general.

The wording matters too, since "at least one SaaS tool" counts one replaced tool the same as ten. The same page also carries a headline that reads "35% of enterprises have already replaced SaaS with custom software", so somewhere along the way teams became enterprises and one tool became SaaS. That's the kind of drift I usually watch happen on social media, except this time it happened before anyone hit share.

Stack Overflow asked a much broader group, and the question was multi-select, so the answers don't add up to 100. The full spread is worth a look: 77.3% build small scripts and automations, 63.0% build team-level internal tools, 54.4% build customer-facing products, 46.0% build cross-team or platform tools, and 18.6% build to replace paid vendors. Replacing a vendor is the least common answer, so developers build a lot, but mostly small things close to their own work.

Menlo asked a third group entirely, 495 US enterprise AI decision-makers surveyed from November 7 to 25, 2025, who are the people who actually sign the contracts. Menlo also compares the 76% with the year before, when 53% of AI solutions were purchased, and it looks like a striking jump until you read the nouns. Last year counted "AI solutions" and this year counts "AI use cases", and since those may not be the same thing, I'd treat the trend as a direction rather than a precise swing. I could be wrong on how much that matters, but it's worth checking.

Before a stat goes into a deck or a budget conversation, I try to ask three quick things. The first is who was asked, because a vendor's own customers, developers in general and enterprise buyers will give you three different answers to the same question, so the sample shapes the result before the headline exists. The second is who paid for it, not because sponsored research is bad, but because every publisher has a room it knows best, and it usually surveys that room. The third is what exactly was counted, since "replaced at least one tool", "builds to replace vendors" and "use cases purchased" sound like the same idea but measure quite different things. None of this takes digging, because for all three numbers the answers sit a scroll or two below the headline, and hats off to the publishers for putting the method right there.

My own read is that build and buy aren't really competing for the same ground. Developers are building lots of small things (scripts, automations, tools for their own team), and some teams go further and retire a SaaS tool, while the large companies buying AI are mostly buying it, which is roughly what I'd expect for the heavy layer underneath that's hard to make yourself. So building is growing at the edges, near the work, and buying is growing underneath, and the argument online only sounds like a fight because each side is quoting the room it happens to stand in.

That edge, the focused tool for one team's process, is where we spend our days at Workbricks, but those three questions work either way. Next time a percentage shows up with a confident caption, try scrolling down to find the sample before you share it, because it's usually the most interesting part of the whole report.