For twenty years the default answer was buy. New research says that is changing: 35% of enterprises have already replaced SaaS functionality with custom software. Here is what the evidence actually shows, where AI changed the economics, and where it honestly did not.
For two decades, the build-versus-buy question had a default answer, and the default was buy. Custom software meant a long project, a scarce team, and a maintenance tail your IT department would carry forever. Off-the-shelf meant someone else's engineers, someone else's uptime, and a predictable invoice. Sensible companies bought.
But everyone who lived with that answer knows what it cost. You conformed your process to the vendor's model of your business, not the other way around. The workflow that made you different became the workflow the software would not do. “Configurable” turned out to mean consulting rates, and the customizations you paid for became the reason upgrades broke. Meanwhile the per-seat invoice grew with your headcount whether or not the seats logged in.
That trade made sense when building was slow and expensive. The evidence now says building has changed.
Retool's 2026 Build vs Buy report, drawn from enterprise survey data, found that 35% of teams have already replaced the functionality of at least one SaaS product with software they built, and 78% plan to build more this year. Perhaps more telling: 60% report software being created outside formal IT oversight, which means the build is happening whether it is governed or not.
The engine behind it is measurable. A study of more than 4,800 developers at Microsoft, Accenture, and a Fortune 100 firm found engineers using AI coding assistants completed 26% more tasks with no measured drop in code quality. A controlled experiment by Peng and colleagues found a 55.8% speedup on well-scoped greenfield tasks.
And here is the number the vendor content leaves out: a randomized trial by METR found experienced developers were 19% slower with AI on complex, unfamiliar brownfield codebases. Both findings are true, and together they say something precise: AI has not made all software cheap to build. It has made a specific kind of software dramatically cheaper: well-scoped applications built on well-understood foundations.
That “specific kind of software” has a name in every operations meeting: the middle tier. Internal tools. Operational dashboards. The inspection tracker, the crew scheduler, the intake form with an approval chain, the lightweight CRM that only needs to do five things, but exactly your five things. Too important for a spreadsheet, never important enough to justify a custom build at 2019 prices, so they lived in off-the-shelf tools that fit at best eighty percent.
This tier is where the studies' optimistic numbers apply. The requirements are knowable, the patterns are proven, the integrations are standard. Three years ago the honest advice for a mid-market company was “make the SaaS work.” That advice is now often wrong.
Most build-versus-buy content assumes building means hiring a dev team and starting from a blank repository. If your organization runs on Microsoft 365, that is not your situation, and the analysis changes.
You already license a build platform. Identity and access come from Entra. Data governance comes from Dataverse and Purview. Hosting, monitoring, and backup come from Azure. The Power Platform covers the low-code tier, and code apps now put full React and TypeScript applications inside the same governed environment, professional-grade software with platform-grade guardrails. Building here is not “standing up infrastructure”; it is filling in the one layer that is actually unique to you: your logic, your workflow, your rules.
That collapses the classic risks of building. The maintenance tail shrinks because the platform carries patching, auth, and compliance. The bus-factor risk shrinks because the stack is mainstream. And the AI acceleration applies at full strength, because this is exactly the well-scoped, well-founded work the research says AI speeds up.
Honesty matters more than enthusiasm here. Buy when the domain is commodity and deep: accounting, payroll, email. Buy when a product's compliance certifications are the product. Buy when your need is genuinely undifferentiated, because a custom version of a generic thing is the worst of both worlds. And do not let anyone tell you AI has made complex, multi-tenant, safety-critical systems cheap. It has not.
The question to bring to every renewal is narrower and sharper than the old one: is this tool's model of our business close enough to be worth what conforming to it costs us? For the core systems, usually yes. For the middle tier, increasingly, no.
Working through that question, tool by tool, license by license, is bespoke analysis by nature, and it is a conversation we have with clients weekly. If a renewal is coming and the eighty-percent fit is starting to hurt, talk to us.
Working through something like this? We talk shop without a pitch. Bring the problem and we will bring what we have learned in the field.
