What actually gets lost when a team works with AI images?

We are building Behind The Workflow and we launch here on 16 September. Before that I would rather ask than pitch.

In our own work the same thing kept happening. An image was approved, the client was happy, and four weeks later nobody could say which prompt, which model version or which reference had produced it. The knowledge lived in one person's chat history, and when that person was away the whole thing had to be redone.

So my question to anyone here who works with AI images or video in a team: what do you actually lose track of? Is it the prompts, the model versions, the references, or is it more the approvals and who said yes to what?

I ask because we had to decide what belongs on the record per asset and what is noise, and I am not sure we drew that line in the right place.