The point is that it is hard to prove that that 5 days code will never break. In the eyes of a customer a feature ready in one morning and in production before evening is a truly great thing and they know that even before AIs developers were shipping a lot of bugs. Some of them were evident only weeks, months or even years after the commit that added them. A customer of mine got a few hours of downtime a few Saturdays ago because some code from 2021 did some potentially long operation in the threads serving requests instead of doing them in a background job.

So what can one do when they expect that you use an AI to speed up development and they manually check features before they are deployed? You develop with an AI, check that it works, deploy to preprod, let the customer do its tests, business as usual.

There are a lot of variables here. Small features I would say of course. Big features or the foundation of another app or codebase is a lot more risky. So it comes down to weighing the tradeoffs and risks.

Even with small features I would say there is some kind of "entropy tax" that AI is adding, but of course this is yet another tradeoff. If companies find value in iterating quickly and then having to throw everything away just to iterate quickly again, then sure. Some businesses are setup to do this. Others not so much. You can't afford to write critical software over and over again. It has to be done right and having that extra care for a few days is the best tradeoff in these scenarios.