The build-versus-buy question for AI is not the same question it was for software, and treating it as such is how good companies waste a year. The underlying models are a commodity you should almost always buy. The data, the workflow, and the evaluation around them are where your advantage lives, and those you should almost always build. Most expensive mistakes come from inverting that — building a model nobody needed, or buying a workflow that buried the one thing that made the company different.
Start with a sharper question than 'build or buy': what is the durable advantage here, and does it come from this capability being yours? If a vendor can sell the same thing to your competitor next week, owning the build buys you nothing but cost and maintenance. If the value comes from your proprietary data, your specific process, or a quality bar only you can define, then buying the generic version will quietly cap your ceiling. The decision is not about technology; it is about where the moat is.
Our default counsel is to buy the model and the infrastructure, and build the thin, high-value layer that is yours: the retrieval over your data, the agent that runs your workflow, the eval that encodes your quality standard. That layer is usually smaller than leaders fear and more defensible than vendors admit. It is also the part that, done well, compounds — each system you build makes the next one faster.
There is a third answer that good frameworks make room for: wait. For some use cases in 2026, the honest read is that the capability is six months from being a reliable build and is not yet a trustworthy buy. Naming that explicitly is a service to the client. We would rather tell a leader that the right move is to sequence a use case behind two others than to bill them for a system that will be embarrassing by the time it ships.
The framework we hand clients fits on a page: separate the commodity from the advantage, buy the former, build the latter, and be willing to defer when neither is ready. It is not glamorous, but it is the difference between an AI program that shows up in next year's numbers and one that shows up only in next year's regrets.