One wrong total in a client email costs more than an hour of checking. The risk is not that AI writes nonsense. It writes something that looks right, and that is harder to catch.
Before any ChatGPT draft leaves your desk, run these seven checks:
- Add up any total yourself. Models find rows well and do arithmetic less reliably.
- Read every promise. A date, a discount or a "we will call you" is a commitment in your name.
- Read every rule it wrote. If it says refunds are always allowed, check your policy.
- Look for extras. Things you did not ask for, like an added service or a made-up detail, often sneak in.
- Compare numbers with the source. Check each number against the sheet, invoice or email it came from.
- Treat a green check as it ran, not as it is right. A workflow that finishes can still produce a wrong answer.
- Read the whole thing once more before it leaves your hands, since your name goes on it.
Where to put them
Ask the model to list every number, date and promise it used. Then run checks 1, 2 and 5 against that list. That turns a vague worry into a two-minute task.
The free lesson walks through the checks with a real example, an n8n chat that found the right rows and still added them up wrong:
https://learn.ynteractive.com/content/blog/check-ai-output-before-you-use-it
Disclosure: I run Stepthrough, which publishes these free lessons.
Top comments (0)