In that case, it should be documented at least that the process is driven by LLM and hence, it's not known exactly how it makes a decision...but I guess that would not look good to most users so they just leave it out.