Create a verification request: make the AI double-check its own numbers. Return JSON {"data":"...","claim":"...","checkHow":"..."}: data — context; claim - what numbers are we checking (list them); checkHow - how exactly to double-check (recalculate in a different way, show intermediate steps, check related numbers with each other).
AI generates plausible text, and the number “920” is as easy for it to write as “960”: arithmetic is not sacred to it, it does not count, but guesses what is similar to the truth. Therefore, any numbers from AI are a draft until they agree with each other. Levers: ask to see intermediate steps, not just the outcome; recalculate in the second way; check the related numbers. Insider: use convergence. Revenue must be equal to the average check multiplied by the number of checks: 920 × 1,630 ≈ 1,499,600, but they told us 1,200,000 - a gap of almost 300 thousand, this is not rounding, but a real mistake. Important: a small gap of a couple of percent is a normal rounding of the average bill, and not a cause for alarm; Sound the alarm precisely at a large gap, like here. If three numbers do not fight with each other, at least one is lying - and this can be seen without a single formula, at the level of “converges or not.” Second trick: check boundaries and common sense - the share cannot be “34%” and at the same time “the majority of sales.” A typical mistake: copying fancy numbers into a report to your boss without checking them. The AI will make a mistake maybe once out of twenty, but this is the time it will end up in the presentation.
Unlock access to submit solutions for instant AI review.