Safe AI email drafts: a cheat sheet for keeping a person in charge

Make it drafts-only (n8n Gmail Tool)

1. Resource opens on Message, and the operation is send. Choose Draft. The operation switches to Create.
2. Under Add option, set Thread ID to {{ $json.threadId }} so the draft sits under the customer's email.
3. Add To Email set to {{ $json.From }}, so you can just hit send after you read it.
4. Keep the email type on text. Plain text reads more like a person.

Give it facts and rules

Rules:
You cannot send email. You can only create drafts.
If you don't know something, say someone from the team will follow up. Never make up prices, dates or services.

From the Greenline system message. The rule helps, but choosing the Draft resource is what makes sending impossible.

Keep private data out

• Before you paste a thread into ChatGPT, leave out anything private, like a card number.
• Scan for ID numbers, birth dates, account and card numbers, and health details.
• The test: only the details this reply needs. Write a short facts list instead of pasting whole records.

Watch for text that gives orders

• Prompt injection: an email the AI reads can hide instructions, like "ignore your rules." The model can't always tell data from commands.
• Give an agent only the tools its job needs. Drafts, not sends.
• Wrap outside text in clear delimiters, like the <email> tags in the lesson's prompt.

Before you click Send

• Read every draft for dates, prices and promises.
• Ask ChatGPT for a table of every promise and the fact behind it, and fix anything flagged.
• Review drafts daily. A pile of unread drafts helps no one.
• If the agent keeps handing off the same question, add the answer to its facts.