the 10-second check that catches ai mistakes before they ship.

elisabeth hitz · july 14, 2026 · 4 min read

you never ship ai output unverified, and the check is faster than you think: spot-check the two or three facts the whole thing rests on, against the source, before you hit send. you do not re-read everything. you find the load-bearing numbers and confirm those. ten seconds of that catches the mistake that would have cost you a client's trust.

the trap is that ai output looks finished. it is formatted, confident, and fluent, which is exactly why people forward it without a second look. fluent and correct are not the same thing, and the gap between them is where the embarrassing errors live.

why is confident-looking output the dangerous kind?

because your brain reads confidence as competence. a clean paragraph with a specific number in it feels checked already. it is not. the model produces the same polished tone whether the number came straight from your file or got quietly reconstructed wrong along the way. this is the core reason you cannot tell an ai is wrong just by reading it, the wrong answer and the right answer wear the same suit.

the check: three moves, ten seconds

you are not auditing the whole thing. you are pressure-testing the parts that carry weight. here is the whole habit.

• spot-check the load-bearing facts. pick the two or three numbers or claims the output actually rests on, the price, the total, the date, the name, and check each one against the source it came from. not everything. the ones that matter.
• ask the model to cite where each came from. say "for each number, tell me which row or line you pulled it from." if it cannot point to a source, that is your flag. a real number has an address in your data.
• sanity-check the totals. do the parts add up to the whole? does the margin make sense? does the sum roughly match what you expected? a total that is off by an order of magnitude is the easiest error to catch and the worst one to send.

that is it. three moves, and none of them require re-reading the full output. you are checking the load-bearing beams, not inspecting every nail.

why "ask it to cite" works so well

a number that came from your data can be traced back to a row. a number the model invented cannot, because there is no row. asking for the source forces the distinction into the open. when you say "which line did this come from," a real figure gets a clean answer and a fabricated one gets a vague one. you do not have to know the data cold to feel that difference.

this pairs with clean input. if you fed it a mangled pdf or an unlabeled dump, you gave it more room to go wrong in the first place, which i covered in the sibling piece on stopping ai mistakes on your own data. good input reduces the errors. the ten-second check catches the ones that get through. you want both, because most of the time the ai is not the thing that is broken, the process around it is.

make it a reflex, not a debate

the reason this fails for most people is not that the check is hard. it is that they decide each time whether the output "looks like it needs checking." it always does. the whole point is that you cannot tell by looking. so you take the decision out of it: nothing goes to a client until the load-bearing facts are confirmed. every time. it becomes as automatic as checking your mirrors before you pull out.

ten seconds, on a document you were about to send anyway, is the cheapest insurance you will ever buy against looking careless in front of someone paying you.

the three questions people ask

doesn't this defeat the point of using ai?

no. the ai still did the drafting, the formatting, and the heavy lifting. you are checking three facts, not redoing the work. the time saved is still enormous, you are just not shipping blind.

what if i don't know the source well enough to check?

then that is the real problem, and the citation move surfaces it. if you cannot verify a number and the model cannot point to where it came from, you do not send that number. you go find the source or you cut the claim.

can i just ask the ai to check its own work?

it helps a little and it is not enough. the same system that made the error can miss it on review. asking for sources is more reliable than asking "are you sure," because a source is checkable and reassurance is not. the final confirmation stays with you.