Every team building with AI is making the build-versus-buy decision whether they know it or not. The teams that get it right are not more cautious. They are running a structured process, and the process has three stages that only work in sequence.

The stages are a sort, a set of columns, and a series of gates. The sort determines what a component is. The columns determine where it belongs. The gates determine whether it is allowed to get there with AI involved.

Stage One: The Sort

Every component in a stack faces three questions in a fixed order. Does it differentiate the product? Does it fail recoverably? Is there a named owner when it breaks?

The answers are not independent. A component that does not differentiate does not get credit for failing recoverably. A component that fails irrecoverably does not get credit for having an owner. The sort is a sequence, and the sequence is the point.

Three questions, answered in order, produce a profile. That profile is what the next stage acts on.

Stage Two: The Three Columns

The profile lands the component in exactly one of three columns. Build with AI. Build without AI. Buy hardened.

Build with AI is the surface. Components here differentiate, fail recoverably, and have a named owner. Iteration is cheap, failure is survivable, and speed compounds. This is where AI earns its keep.

Build without AI is the differentiated foundation. Components here differentiate and fail irrecoverably — worth owning, not worth learning on. The team writes the logic, reviews it, and signs off with no generation step in between.

Buy hardened is everything else. Components that do not differentiate, or that fail irrecoverably without a named owner, are components someone else has already built, tested, and is accountable for. A solved problem rebuilt from scratch is not innovation. It is paying tuition for a lesson the industry already funded.

The columns are a map. They tell a team where a component belongs. They do not yet tell a team whether it is allowed to arrive there.

A solved problem rebuilt from scratch is not innovation. It is paying tuition for a lesson the industry already funded.

Stage Three: The Gates

Before any component reaches AI-assisted building, it passes through four gates. A component that fails any gate is disqualified from every gate after it. The order is not negotiable.

Gate one asks whether the component differentiates. Failing here sends it toward the buy column without further AI consideration. Gate two asks whether failure is recoverable. Failing here disqualifies the component from AI-assisted building even if it differentiates and has an owner.

Gate three asks who owns the component when it breaks. If the answer is a named team with the expertise to fix it, the component passes. If the answer is “whoever wrote it,” it fails.

Gate four asks whether AI is allowed to originate the logic. Components that cleared the first three gates can be built with AI assistance. Components that cleared gates one and three but failed gate two cannot — they are worth owning, but they must be written and reviewed by the team with no generation step.

The gates are what separate a decision from a guess. Every component that reaches the build-with-AI column has passed all four. Every component that reaches the build-without-AI column has passed three and failed exactly one.

How the Three Stages Work Together

The sort determines what the component is. The columns determine where it belongs. The gates determine whether it is allowed to get there with AI involved. Skip the sort and the columns are arbitrary. Skip the columns and the gates have nowhere to send anything. Skip the gates and the whole process collapses into an opinion.

A component that fails the sort never reaches the columns. A component that reaches a column without passing the gates is not a decision. It is a default, and defaults are how the same weaknesses get rebuilt in slightly different forms.

The Question the Framework Answers

The question is not whether the code can be generated. It is which of the three columns the component has earned, and whether it cleared every gate on the way there.

AI removed the cost of producing code. The sort, the columns, and the gates are how a team decides what that code is allowed to touch — and how they know the answer was reached on purpose.