The honest answer is almost never buy or build. It is buy these layers, build that one, and partner where you lack the muscle but need someone accountable. This is a framework for getting it right, layer by layer.
Treat AI as one decision, and you rebuild plumbing you should have bought or hand a vendor the one thing that was your edge.
Decompose it by layer, default each layer, and decide build, buy, or co-build for each one.
If a capability underpins your competitive position, build it, because buying it means competitors can buy the same thing.
If the capability is undifferentiated, buying is faster, cheaper, and frees your team.
Cost and ROI is consistently the single most overriding factor in these decisions, and the one teams model worst.
Break the capability into infrastructure and models, platform and orchestration, application and workflow, and data and domain logic. You are deciding four things, not one.
Buy infrastructure and platform, blend application, build and own data and domain logic. The teams that lose do the reverse: they reinvent infrastructure while using generic applications that do not fit. Then look for reasons to deviate.
For anything you are tempted to build, ask: is it a real advantage, sensitive, or deeply integrated? If none, buy it.
For each layer you will build, ask whether a partner co-build would raise your odds.
The market has settled the commodity layers: buy them. The differentiated layer, your data, your domain logic, the last mile, is yours to own.
Almost never. Foundation models and infrastructure are commoditized and improving fast. Build the data and domain logic on top, not the model underneath, unless extreme scale or regulation forces self-hosting open models.
Statistically, no. Fully internal builds succeed at about half the rate of vendor co-builds, largely for lack of external accountability. You can keep ownership and control while co-building the hard parts.
In the differentiated 30%. Build does not have to mean build entirely from scratch with your own people, which is the riskiest path. Co-build with a partner who brings production experience and external accountability while you keep the IP, the data, and the result.
No. The 76% is buying the commodity layers, which is right. The differentiated ~30%, your proprietary data and workflow logic, is where building still wins. Buying that away is buying away your advantage.
Building down the stack and buying up it: reinventing infrastructure while using generic applications that do not fit. Do the reverse.
