When to Build vs. Buy AI (The Framework I Actually Use)

Most build vs. buy decisions in AI are made wrong — not because the analysis is bad, but because the framework is.

Companies apply SaaS logic to an AI decision. They compare vendor features, run a demo, check pricing. That framework works for software that does a defined thing. AI doesn't do a defined thing. It does a probabilistic thing, shaped by data it's never seen before.

The gap between 80% accuracy and 95% accuracy in a SaaS tool is a product annoyance. In an AI product, it's the difference between something you can ship and something you quietly kill.

Is this infrastructure or a weapon?

Before anything else (before cost, before timeline) I ask one question: is AI in this area a competitive differentiator, or is it infrastructure?

Infrastructure means everyone in your category will have it eventually. Meeting transcription. Email classification. Document parsing. Commodity capability. Buy the best vendor and move on.

A weapon means the model behavior is where you actually compete. Buying a vendor solution means your competitors can buy the same thing tomorrow — you're not building a moat, you're renting one.

Three filters after that

Once the differentiation question is answered, I run three more:

Data advantage. Do you have proprietary data (the kind a vendor doesn't have access to) that a custom model would learn from? If yes, building gives you a compounding return over time. If your data looks like everyone else's, a vendor trained on a billion more examples than you'll ever collect is probably better.

Control requirements. How much do you need to understand why the model makes the decisions it makes? Vendor models are black boxes. A vendor update can silently break something in your product, and you'll find out from a customer complaint. In regulated industries, or anywhere explainability matters, that's a non-starter.

Discovery vs. scale. Trying to validate whether AI can solve this problem at all? Buy cheap, run the experiment fast. Already know the use case works and ready to scale it? That's when building starts to pay off.

What I actually see in the field

Most companies should buy more than they think, at first.

Companies that over-build early spend 12 to 18 months on infrastructure a vendor already solved, before they've validated whether the capability delivers any value.

Companies that over-buy end up with a portfolio of vendor solutions that each own a different slice of their customer experience, with APIs that were never built to work together. Their AI stack becomes a collection of bolt-ons that were never meant to be a system.

The pattern that works: buy to validate, then build selectively where differentiation is real. You don't have to decide once. The right answer changes as the product matures.

One question I always ask last

What happens if this vendor gets acquired, doubles their price, or quietly changes their model?

If the answer is "we'd have a bad quarter," you need more optionality than a single vendor gives you. Either build an abstraction layer that lets you swap models, or start developing internal capability in parallel now.

The companies that treat AI as a core capability rather than a software line item are the ones with a durable edge two years from now. The ones who bought their way to AI are managing renewal negotiations, not compounding an advantage.