How to check an AI draft before sending it.

Give the team an inspectable source.

Use a short fictional procedure everyone can read, such as ordering office supplies. Keep the source and the draft side by side. The exercise is about checking whether the draft preserved the facts and conditions, not whether it sounds confident.

Use invented information in a public demonstration or rehearsal. An exercise does not authorize entering sensitive information into an AI service.

Try a deliberately flawed example.

Fictional source: “The office manager approves supply orders before purchase.” Draft: “Place your order and send the receipt to the manager.” The draft sounds plausible, but it changed approval-before-purchase into notice-after-purchase.

A reviewer should keep the original order of actions. A shorter or friendlier sentence is useful only when it preserves the meaning.

Fictional demo business, demo data. An invented office-supply exercise, not a real policy or client incident.

Use five checks.

• Facts: do names, numbers and dates match the source?
• Conditions: did “only if,” “before” or another meaningful limit survive?
• Authority: did the draft promise something nobody approved?
• Missing information: did it invent an answer rather than identify a gap?
• Ownership: who approves the final version and who sends it?

Correct, escalate or stop.

If the issue is a simple wording error, correct it against the source and review again. If the source is incomplete or contradictory, ask the person responsible for it. Do not let a confident draft settle a policy question.

Give staff a manual fallback and a clear approval owner. A useful training outcome is that someone can explain why a sentence should not be sent—not just produce a better prompt.

Keep the boundary clear.

This is a proposed business-writing exercise. It is not clinical training, a certification or evidence of completed staff instruction. A workshop, tool rollout or automated message workflow needs its own agreed scope and checks.