What should I investigate first when an AI tool gives the wrong answer?
When an AI tool gives a wrong answer, check what information it was allowed to use before you rewrite the instruction. Follow one disputed claim to where it came from and confirm who owns it, its date and its status. A clearer instruction can't make an expired policy current.
Teams often reach for the instruction first. The source deserves a look before that.
Look for a newer version of the document, and for conflicting information in connected folders or reference material. Then ask whether the tool could reasonably tell an approved rule from a draft. If that distinction lives only in someone's memory, make it visible in the source system. A filename nobody maintains won't do it.
Fix the source through its responsible owner and repeat the original task. Check related outputs that may have leaned on the same information, because a local correction can leave other uses unchanged.
Only after that, decide whether the instruction or the review process needs work. Going in this order keeps you from building ever more complicated prompts around a basic information problem, and it gives the broken part of the workflow a clear owner.
What to try this week
Take one wrong AI answer your team has seen and follow it to its source, and confirm who owns the document along with its date and status. Fix it through the owner, repeat the original task, and leave the instruction alone until then.