Your name, your check
The two-minute check
Before anything AI-assisted leaves your desk, check five things. Facts: is each claim true, with a source if it matters? Numbers: do totals match the spreadsheet? Names and dates: right people, right spellings, real meetings? Promises: has it committed you to anything? Confidential detail: has anything slipped in that shouldn't go to this reader?
Tone and fit
AI drafts tend to be longer, more upbeat and more formal than people write. Cut the first and last sentences and see if anything is lost. Remove phrases you'd never say. Check it fits the relationship: a reply to a worried customer and a note to a long-standing colleague shouldn't sound the same.
Ask the AI to attack its own work
A useful habit: paste the final version into a fresh chat and ask, "What would a sceptical reader challenge here? Which claims need evidence? What might be misread?" A new conversation doesn't share the original's assumptions. It isn't a replacement for your own read, but it often finds something.
When to say AI helped
Follow your employer's policy first. Many workplaces expect disclosure for client deliverables, published content and anything legal, medical or financial. Even where it isn't required, being open about AI use in reports and analysis builds trust. If you'd be uncomfortable with the reader finding out, treat that as a sign you should tell them.
If a client found a mistake in something you sent, would "the AI wrote it" make the conversation better or worse?