Getting an unmanned and unsteerable probe out into space to take photos of Saturn's rings requires some serious planning. After launch and maybe a few minor corrective thrusts here and there, the probe just flies where the laws of physics tell it to. Get the calculation wrong, and the precious object crashes into Jupiter instead. If that happens, the lesson is that you have to plan better next time.
And that's also the (wrong) lesson that project managers learned when it came to software projects: End result doesn't solve the business problem? Must have not planned hard enough. Next time we'll write a better specification and then those pesky engineers will get it just right.
Of course that's not what happened. The way to a successful software project was to realize that a good product isn't built as much as it's grown, in small iterations, from tiny seeds. Like a plant growing toward the light, the product grows toward customer needs, which are met and discovered in a never-ending cycle.
The space probe needs upfront planning because it cannot be steered, and it permits upfront planning because there are no surprises and the governing laws are well understood.
The software product is the exact opposite. It doesn't permit upfront planning because there are boundless surprises in a poorly defined and understood environment. I fail to see how "but AI writes the code faster" makes the environment less uncertain.
If anything, the apparent speed benefits of AI should allows us to be more relaxed about pre-specifying things. If you were previously worried that you had to spec out everything in exhaustive detail because development would take so long and you wanted to make sure you got everything right, then what are you worried about now?
Build. Measure. Learn. As fast as possible, as often as possible. A lengthy spec just gets in the way of that.
