Enterprise buyers spent 2025 answering the build-versus-buy question with a number: 76%. That is the share of enterprise generative AI solutions purchased rather than built in-house last year, up from a roughly even 53%/47% buy/build split in 2024, according to Menlo Ventures' annual survey of enterprise AI adoption (Menlo Ventures). The swing happened inside twelve months. Most of the "should we build this ourselves" debates founders are still having internally were already settled by their peers before the debate started.
The money followed the decision
Enterprise generative AI software spend reached $37 billion in 2025, up 3.2 times from $11.5 billion in 2024, split roughly $19 billion on applications and $18 billion on infrastructure (Menlo Ventures). Applications, the layer closest to "buy," grew to be nearly as large as the entire infrastructure layer combined. Adoption backs this up at the ground level too. McKinsey's global AI survey found 72% of organizations reported regularly using generative AI in 2025, up from 33% in 2024, a jump the firm attributes largely to teams adopting off-the-shelf gen AI tools rather than standing up custom models (McKinsey).
What buying actually solved
In a separate survey of 100 enterprise CIOs, Andreessen Horowitz found 90% said their organizations are now testing third-party generative AI applications for use cases like customer support, rather than building the capability internally (a16z). The case studies buyers cite explain why. Intercom committed $100 million to replatform its business around AI and shipped its Fin AI agent within months of GPT-3.5's release instead of training a model from scratch; Fin now resolves millions of customer service queries per month (OpenAI). Moderna used ChatGPT Enterprise, not a custom-built system, to cut a core analytical step in drafting Target Product Profiles from multiple weeks to hours (OpenAI). Neither company needed a new model architecture. They needed the workflow solved fast, and a purchased product got there before an internal build could.
Buy and build are converging, not competing
The line between the two categories is blurrier than the headline ratio suggests. The same a16z survey quotes a CTO at a high-growth SaaS company describing what "build" now looks like inside a company that still calls itself a builder.
Nearly 90% of the company's code is now AI-generated through tools like Cursor and Claude Code, up from just 10 to 15% a year earlier.
(CTO at a high-growth SaaS company, cited by Andreessen Horowitz.) That company would show up as a "builder" in any survey. Its build process is now mostly assembling AI tools rather than writing logic from scratch. Buying a finished application and building a feature on top of a foundation model API are converging on the same behavior: assemble from existing components, do not originate them from zero.
The decision rule for this week
The evidence points to a filter, not a philosophy. Buy when the workflow is common enough that a vendor has already been paid to solve it for dozens of other companies: support deflection, coding assistance, document drafting. Wire directly to a model API and build the last mile yourself when the workflow touches proprietary data, a regulated process, or the specific judgment your product is supposed to be uniquely good at. Notice that nobody in this evidence set built a foundation model. Intercom and Moderna both integrated an existing one and spent their engineering budget on the last mile around it.
This week: take the AI features sitting in your product backlog and sort each one into two piles using that filter, common workflow versus proprietary edge. For every item in the common-workflow pile, get a vendor quote or a trial before a single engineer writes code against it. The 76% of the market that bought instead of built in 2025 was not moving slower for it.