AI coding tools extract more leverage from codebases with clear, common patterns than from proprietary or inconsistent systems. Popular stacks like React, TypeScript, and Python give models millions of training examples, while private frameworks and legacy Java codebases force developers to teach patterns through limited context windows — burning tokens and increasing output variance.

Two workflows illustrate the gap. In a consistent codebase, the model reads the spec, recognizes established patterns, and generates. In an inconsistent one, it must first infer conventions from messy history, then read supplementary docs, then attempt the task. Each extra context step adds cost and degrades reliability.

This reframes the economics of rewrites. Modernizing isn't just about tech debt — it's about restructuring around patterns AI already understands. Teams maintaining internal DSLs or idiosyncratic architectures spend cycles teaching the model what Next.js or Rails developers get for free.

The compounding disadvantage isn't merely velocity. Output quality diverges because models operating in familiar pattern space produce fewer hallucinations and better architectural decisions. Rewrite ROI calculations must now account for AI compatibility as a first-class factor.