A research paper reveals that AI models themselves are just a small part of what makes AI products work—the bulk is unglamorous operational software that handles security, access, and user experience.

What happened: A research paper from April found that Claude Code's codebase is only about 1.6% AI decision logic, with the remaining 98.4% being operational infrastructure. The finding underscores that the actual AI model is cordoned off from most of the work, which is handled by deterministic computer code.

Why it matters: As AI tools move from research curiosities toward real consumer products, the orchestration layer—the non-AI operational tasks that let the technology function in the real world—has become as important as the model itself. Consumers expect products to simply work without thinking about security or access control, and that requires far more software than the AI component alone.

What to watch: The balance is shifting: while a very good AI model is still necessary for great results, the infrastructure being built around the models is becoming the actual differentiator. Companies building consumer AI products will need to invest heavily in the operational software, not just the models, to compete.