AI email drafting with a mandatory human send step

Drafting and sending are different authorities. An AI step can propose text while the workflow prevents transmission until a person reviews the exact recipient, subject, body and attachments.

Keep the send gate closed

The automation lab creates a synthetic draft in a pending state.

from automation_cases import email_approval_case

result = email_approval_case()
assert result["state"]["draft_created"] is True
assert result["mandatory_review"] is True
assert result["state"]["send_enabled"] is False
assert result["external_message_sent"] is False

No mailbox or messaging service was contacted.

Review the complete message

Generate a draft from authorized source data and mark uncertain claims. Apply recipient/domain allowlists, data-loss checks, link and attachment validation, tone/policy rules and required disclaimers. The approval UI must show the actual send payload, not a summary.

Bind approval to a digest and reviewer identity. Edits after approval reopen review. On approval, use a narrow send credential and an idempotency key; record provider message ID. Decline, timeout and unavailable reviewer should stop or route according to written policy.

Test reply-all, hidden recipients, stale drafts, injection in quoted email, attachment substitution and retry after an ambiguous timeout. Avoid storing entire email bodies in broad workflow logs.

The Generative & Agentic AI course connects human approval to agent tool boundaries.

Exercise

Build a local draft-and-approval state machine. Mutate the recipient after approval and prove the send stub rejects it; replay the approved request without duplication.

Continue learning

This article is part of the Business automation with AI sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Automate support categorization with a fallback queue.
• Continue with Create a document intake workflow with duplicate detection.

Reference: Zapier AI tool approval documentation.