Is the Number Real? Four Checks Before You Repeat an AI Answer
An AI answer can sound finished, confident, and specific. That is exactly why it is risky to repeat in a meeting. A figure that appears in a slide or gets read aloud carries the weight of the person who said it, even when that person only copied it from a chat window. Before you repeat any AI answer to your team or a client, run it through four quick checks. Together they take less than a minute and catch most of the errors that cause embarrassment later.
Check one: does it name a source?
Start with the simplest question. Where did this come from? An answer that says "customer churn rose 12 percent last quarter" with nothing behind it is a claim, not a finding. A useful answer points to a specific place: a named report, a table in your data, a particular export, a page on the web. If the assistant cannot tell you where a number came from, treat it as a hypothesis to investigate rather than a fact to pass along.
Be wary of answers that cite a source vaguely, such as "industry reports suggest" or "according to research." Ask which report, which year, which firm. If the answer cannot be narrowed down, the sourcing is decorative.
Check two: can you open it?
Naming a source is not enough. You need to be able to open it yourself. Click the link, find the file, locate the row in the spreadsheet. This check matters because AI systems can produce citations that look entirely plausible but point to pages that do not exist or that say something different from what was reported. A reference you cannot open is not a reference.
When you do open the source, do not just confirm that it exists. Read the surrounding text. A real report can still be quoted out of context, and the figure you need may refer to a different region, period, or definition than the one in your question.
Check three: does the number match?
This is the check most people skip, and it is the one that most often fails. Compare the number in the AI answer against the source you just opened. Does it match exactly, or has it been rounded, combined, or recalculated? Many errors come from small transformations: a percentage that was actually a share of a subset, a total that includes a category the question excluded, a currency that was never converted.
If the number came from calculation rather than direct lookup, rerun the arithmetic yourself on a calculator or in a spreadsheet. Ten seconds of arithmetic is cheaper than correcting a figure after a client has already written it down.
Check four: is the date recent?
A correct number from the wrong year is still wrong. Check when the source was published and what period its data covers. Markets, prices, headcounts, regulations, and product features all change. An answer that is accurate as of two years ago can mislead a decision made today.
Pay attention to phrases like "currently" or "as of now" in AI answers. The assistant may not know the date of its underlying information. Make the date explicit in your own notes so anyone reading later knows what the number describes.
Making the checks routine
These four questions work best as a habit rather than a special procedure. Before a number leaves your hands, ask: source, open, match, date. If any answer is no or unclear, either fix it or label the figure as unverified in the room. Saying "this is from an AI summary and I am still confirming it" is a far better position than presenting an error with confidence.
The goal is not to distrust every answer. Many AI outputs are useful and accurate. The goal is to make sure that the people relying on your words can trust them too. Four checks, done consistently, protect your credibility far more than any single clever answer protects it.
