For smaller applications, AI can often handle the full cycle effectively. For larger, long-lived systems, though, a coherent domain model and deep context become critical. In the evolutionary modelling approach, building the code is the learning process. The friction, the failed attempts, the moments when a concept doesn’t fit, these aren’t obstacles. They are the essential feedback mechanism that refines your understanding of the domain. The code becomes both the map and the territory. AI breaks that loop. It’s like having someone else read War and Peace for you and then giving you a detailed summary. You might grasp the broad plot, the Napoleonic wars, the fates of Pierre, Natasha, and Andrei, but you miss the emotional depth, the evolving character arcs, the philosophical layers, and the subtle interconnections that give the story its real power. In small applications, those nuances often don’t matter much. In complex systems, they are everything. The accumulated understanding of why things exist, how responsibilities truly relate, and how the domain behaves under pressure lives in those details. When you outsource the building, you also outsource the learning. That’s why the 85%; the structuring, naming, and boundary finding, cannot be fully delegated. It’s not just design. It’s domain acquisition through the act of building.

The “code as learning” point resonates. AI gives you the result, but skips the iterations that normally shape your understanding. That tradeoff is fine for small systems, but becomes risky as complexity grows.