How to check AI-generated numbers before you use them

AI-generated numbers can look right and still be off. A short verification routine catches most errors before they cause problems, especially in finance, reporting and customer-facing work.

Step 1: Sort numbers by risk

Not every figure needs the same scrutiny. Start by sorting what the AI produced:

• High risk: money, tax, doses, legal or compliance figures, anything published externally.
• Medium risk: percent changes, averages, totals in internal reports.
• Low risk: rough estimates clearly labeled as approximate.

Step 2: Ask for the working

Prompt the model to show each step of the calculation. A visible chain of steps makes errors easier to spot, such as a dropped decimal or a tax rate applied twice. Showing the working improves transparency, but it does not guarantee the answer is correct.

Step 3: Recompute with a reliable tool

Re-run the math in a spreadsheet, a calculator or code. For a total like $347.60 plus 13% tax, the exact result is 392.79. A near miss such as 392.97 is easy to overlook without this step.

Step 4: Check counts directly

For word counts, letter counts or item counts, use a word processor or a script rather than trusting the model's tally. Language models read tokens, not individual letters or words.

Step 5: Add a sanity check

Ask whether the result is plausible. Does the total equal the sum of its parts? Does a percentage fall between 0 and 100? Does the figure match last quarter's order of magnitude?

Step 6: Build the habit into the workflow

For repeated tasks, the best approach is to have the system do this automatically: route math to a calculator or code, and recheck totals before output. See how AI thinks for more on model behavior, and the Responsible AI-by-Design Framework for how Silverberry approaches trustworthy use.

Verification takes minutes. A wrong figure in a published document can take far longer to fix.