I think part of the problem is that current AI tooling massively lowers the cost of making infrastructure and product changes, but doesn’t equally improve consequence reasoning.

Generating changes is becoming cheap. Predicting second and third order effects across large systems is still hard.