It's easy to add more code, but it also increases the risk, especially if you are the only one who understands how it works. As more people on the team understand it, the risk decreases. It becomes less likely that the code interacts in an unexpected way with other parts of the system, and it also becomes more likely that the existence of the code is justified because noone felt the need to replace it.
If you fire the team and replace it with another, you now have a lot of code and zero understanding. That understanding first has to be mined before you can close the asset/liability gap.
In the age of LLMs, it becomes easy to produce 10x as much code as before, but I don't think you can put the same multiplier on how fast you can increase the understanding. It may even be <1 if you are not being mindful about your LLM use.
The thing I keep seeing is greenfield AI projects built by ill-conceived, self-styled software factories. They grow incredibly quickly, and then reality sets in and the pace slows. Demonstrating this is difficult. There are real productivity increases to be had here no doubt, but the long tail is still expensive.