AI tools write fluently, and that is the trap. A draft can read perfectly and still contain a wrong figure, a made-up reference or a date that is a year out. The tool does not flag these. It sounds just as sure when it is wrong as when it is right.
The fix is not to stop using it. It is a short check before anything leaves the building.
The two-minute check
- Numbers: totals, prices, percentages. Check each one against the source, not against what looks plausible.
- Names and dates: people, companies, deadlines. These are the details AI most often gets slightly wrong.
- Sources and quotes: if it cites a report, a clause or a regulation, open it and confirm it says that. AI can invent references that look real.
- Promises: anything that commits you to a price, a date or a scope. Read it as the client would.
- Tone: does it sound like you? A polished paragraph that does not match how you normally write stands out.
Before sending your next AI-drafted email, run the five checks above. Time it. Most people find it takes under two minutes, and it usually turns up at least one thing worth changing.
Match the check to the risk
Not everything needs the same scrutiny. An internal note to a colleague needs a quick read. A client proposal, a quote or anything legal or financial needs every line checked by someone who knows the subject.
A simple rule works: the further it travels and the more it commits you to, the more carefully it gets checked.
Make the sender responsible
AI does not take responsibility for what it writes. The person who sends it does. Make that the team rule, and the check stops feeling optional.
Use tools that show their working
The safest AI tools point to where each answer came from, so checking takes seconds rather than minutes. If a tool cannot show its source, treat its facts as a first draft.