Reconcile the totals before you ask for the summary, not after

If you're handing an AI tool a spreadsheet or a pasted table of figures to summarise, don't assume it's going to notice the totals don't reconcile unless you specifically ask it to check. A model asked to "summarise this" will summarise what's there, including a subtotal that's out by £400 because of a formula that didn't pick up the last row - it has no built-in instinct to sanity-check the source data before describing it, because you didn't ask it to.

Ask explicitly, as a separate step: "before summarising, check that the subtotals sum to the total, and flag anything that doesn't reconcile." That one extra sentence changes the task from "describe what's here" to "verify what's here, then describe it" - and it's the difference between catching a broken formula before it reaches a report and finding out about it when someone else spots it in the board pack.

This is worth doing even when you're confident the sheet is right, because the check costs a few seconds and the failure mode it catches - a summary that faithfully describes numbers nobody actually checked - is exactly the kind of error that looks most credible once it's dressed up in a well-written paragraph.