Next time an AI gives you a number, don't take its word for it. Find the source it cites, open that record in the original tool, and confirm the figure or date matches. It sounds like extra work, and for a single throwaway question it sometimes is. But for any number that will end up in a client email, a budget review, or a decision, this one habit catches most of the quiet errors that AI answers produce.
The reason is simple. A language model is very good at producing text that sounds like an answer. It can read a question about last quarter's revenue, pick up a plausible-looking figure from somewhere in its context, and present it with the same confident tone it uses for correct answers. The tone does not change when the number is wrong. So the tone cannot be the thing you check.
What "the source" actually means
When an AI tool is connected to your business systems, a good answer should point to a specific place: a CRM opportunity, an invoice record, a row in a help desk ticket export, a line in a dashboard. That pointer is the useful part. A vague reference like "based on your sales data" is not a source. You want the record name, the identifier, and the field.
Once you have that, the check is mechanical. Open the same record in the tool where it lives. Look at the field the AI named. Compare the value. If the invoice says it was due on the 14th and the AI said the 12th, you have caught something, and now you know to look at how the date was read, whether it came from a stale copy, or whether two records were blended together.
Where the mismatches usually come from
In practice, a few patterns account for most of the wrong numbers.
The first is stale data. A system that exports yesterday's file will return yesterday's total, and nothing in the sentence will tell you the file is a day old. Checking the source record in the live tool shows you whether the figure has moved.
The second is the wrong entity. Two customers with similar names, two deals on the same account, two invoices that share a reference prefix. The model picks one, and the number belongs to the other. Opening the source lets you see which record was actually used.
The third is unit or period confusion: monthly versus annual, net versus gross, a fiscal quarter that starts in a different month than you assumed. These errors are hard to spot from the answer alone, because the number looks reasonable.
Making the check fast
Verification only survives if it is quick. A few habits help. Ask the assistant to name the record and field every time you request a figure you intend to reuse. Keep the source tool open in a second window for the numbers that matter. For recurring reports, note which record is the authoritative one, so that you are not guessing at the start of each week.
It also helps to decide in advance which numbers warrant the check. A rough sense of the team's mood does not need a citation. A dollar amount going to a customer does. Matching the amount of checking to the cost of being wrong keeps the habit sustainable instead of exhausting.
Why this matters more as tools get better
It is tempting to think the check becomes unnecessary once the models are accurate enough. But accuracy is a rate, not a guarantee, and the errors that get through are exactly the ones that look most plausible. A better model makes fewer mistakes and makes each mistake harder to notice. The verification step does not depend on the model being bad. It depends on the fact that the record in the source system is the thing that is actually true, and the answer is only a claim about it.
Treat every AI number as a claim to be checked against its origin, and most of the risk goes away. The habit is small. The payoff is that you stop finding out about wrong figures from a client who has already acted on them.