What people expect AI to do
Think of AI like a very fast, very thorough reader. Point it at old software, and within hours it can tell you what every screen does, how the pieces connect, and which parts look outdated. A job that used to take a person two weeks now takes an afternoon.
That’s genuinely useful. It’s also why people assume the hard part is now done.

Why that's not the whole story
Nothing in the code explained this.
It lived in one person’s memory.
What AI is
genuinely good at
Reading fast, mapping how everything connects, and flagging things that look risky or outdated — reliably, at a scale no team can match by hand.
What only
your people know
Why something was built that way — the old client request, the past incident, the workaround nobody wrote down. That lives in memory, not in code.
One simple question to ask
Find a part of your system that looks a little odd. Ask your AI tool why it’s built that way. Then ask your most experienced team member the same question.
A smooth,
confident answer.
Let me
check on that.
If the AI gives you a smooth, confident answer — be a little more careful, not less.
It’s likely guessing well, not remembering. If your team member pauses and says “let me check on that” — pay attention. That pause is your real risk. Not the code. The fact that the reason for it exists in only one or two people’s heads.
AI was never the problem.
Losing the “why” before anyone writes it down is.