The system can keep running long after the reasoning behind it has left.
You can’t test it, because the only place it runs is production, so you read the logs and you sit with someone while they do their job with it and build up a picture of what it does from the outside. The customer said it pulls from the billing platform on a schedule, but the logs say it pushes on a trigger, the connection string points at a host that isn’t on the network diagram, and the job everyone calls the nightly sync runs four times a day. At one point the worker does a step that makes no sense, and when you ask why, the answer is that’s just how it’s always done. It takes time and the picture you end up with is nothing like the description, but you get there.
Why any of it is that way is the part that doesn’t come back. You can see it pushes instead of pulls without knowing what broke the one time someone did it the other way, or which of those four daily runs is load-bearing and which is a workaround for something that went wrong years ago and never got written down. It isn’t in the code or the comments or the four-line wiki page, and the ticket history starts two years after the decision was made. The person doing the step every day can’t tell you either. It was in the head of whoever built it, and they left.
The notes can tell you what needs doing without telling you why it ended up that way.
This is most of why migration is slow rather than the reading. Repointing that connection string at the host the diagram says it should have been using all along is exactly how you cause the outage you were brought in to avoid, so you move one change at a time, each one a bet that the thing you’re removing was decoration and not critical.
I’ve done the departed-creator version of this my whole career, where the system outlives the person and you reconstruct the reasoning from outage history and whoever’s still in the building. A prompt-built system is the same shape with the person taken out of it from the start. Someone makes a prompt and you have a system. There was no point at which anyone decided it should push instead of pull for a reason you could go and ask about, because the shaping happened inside the model in a single step and what came out the other side is just the system, running, doing what it does.
The legacy estate is what a system looks like after the why has drained out of it over a decade. A prompt-built one looks like that on the day it ships. Nobody kept the prompt, but the prompt was never the point.
So the worry that gets attached to all this is the wrong worry. When someone leaves and the concern is that output craters and you have to hire four people, that’s the visible loss, and it was mostly visible already. What actually left was the why, and with the old systems the why was usually gone long before the person was. AI doesn’t introduce that problem. It just builds the system that way from the start, with no version of the why that anyone could have gone and recovered.