More than half of us have already shipped an AI mistake. Here is how to stop being in that half

The report went out at 16:50 on a Friday. The AI-drafted section read beautifully, the numbers were plausible, and everyone was tired. On Monday a client replies asking about the figure in paragraph three, the one that does not appear in any source document, and now the conversation is not about the report anymore. It is about whether your reports can be trusted.

What makes this story so common is that nothing in it involves carelessness. The person who shipped it works hard and checks things, usually. KPMG's numbers say this is the norm, not the exception: most employees have been there, because the tools got fast before the habits did.

The 58% who rely without evaluating are not lazy. Checking an unanchored AI answer is genuinely expensive: find the source, search the document, locate the passage, compare. Under deadline, that cost loses to fluent text that looks done, every time. Fluency reads as competence, and the human brain treats well-formatted output as finished work.

This is why lectures about being careful with AI change nothing. The economics of the check decide whether it happens, and as long as verification costs minutes per claim, the tired version of every one of us will skip it. The fix has to make checking cheaper, not people guiltier.

The first habit is a boundary, not a burden: any AI-supplied claim that will leave your hands, into an email, a report, a decision, gets the 90-second check. Open the source. Compare the claim to the passage. Note what it does not prove. Everything else, brainstorms, drafts for your own eyes, low-stakes summaries, can flow freely.

The boundary matters because it is sustainable. Checking everything is impossible and checking nothing is how the 57% got their scar. Checking exactly what carries your name is a rule you can keep on your worst day, and it is the rule that would have caught the Friday report.

The second habit covers the other direction: when AI edits or produces work you own, its changes arrive as visible proposals you accept one by one, never as silent rewrites. A finished-looking document hides what changed; proposals show you exactly what is new, in context, while your attention is on that passage.

Together the two habits close both doors the mistakes come through: unverified claims flowing out, and unreviewed changes flowing in. Neither habit slows real work down, because both replace a large vague audit with many small cheap decisions.

Tools decide whether habits survive. In Tatsulok, every claim in every answer links to the exact passage behind it, so habit one costs one click, and suggestions on your documents stay visible proposals until you accept them, so habit two is the default rather than a discipline. When your sources do not contain an answer, it says so, which removes the most dangerous category of workplace AI mistake entirely: the confident invention.

If you run a team, the KPMG numbers are your training agenda in two lines: teach the verification boundary, and give people tools where keeping it is nearly free.

This week's drill, no signup needed: recall the last AI output you passed along without checking. Find it, run the 90-second check on its central claim now, after the fact. If it holds, you have lost 90 seconds. If it does not, you have found the mistake before your client does, and you will never again need convincing that the habit pays.