Never let the model originate a number you didn't give it

When you ask an AI tool to summarise or write up financial figures, every number that comes back in the response should trace to a number you actually gave it - a total it added up from your data, or a percentage it calculated from figures you provided. The moment a number appears that you didn't supply and can't immediately see the working for, treat it as invented until you've checked it, because a model producing a plausible-sounding figure that was never actually calculated is one of the most common and hardest-to-spot failure modes in this kind of work.

The tell isn't that the number looks wrong - it's that it looks completely reasonable, sitting in the right range, formatted correctly, next to real numbers it didn't invent. That's what makes it dangerous in a way an obviously broken number wouldn't be: nobody double-checks a figure that already looks right.

The practical fix costs almost nothing: ask the model to show its calculation inline, not just the answer - "total = £X + £Y = £Z" rather than just "£Z" - so there's something to actually verify rather than a number to take on faith. If it can't show working for a figure, that's your signal it wasn't calculated from anything you gave it, and it doesn't go in the report until you've worked it out yourself and confirmed it.

This matters more here than almost anywhere else on this site, because a wrong number in a board pack or a customer invoice doesn't get quietly corrected later - it gets acted on.