AI is often trying very hard to be helpful. Occasionally, too hard.
If the information is incomplete, it can be tempted to fill the gap and give you something that sounds perfectly sensible. A very useful instruction is simply: if the evidence is not there, say you do not know.
We use versions of this when analysing reports, documents and operational information. Do not infer a number that is missing. Do not pretend the document answers a question it does not answer. Tell us what you need next.
That small change makes the answer much more useful because uncertainty becomes visible. “I don’t know yet, but I need these two things” is often far better than a beautifully written guess.