An approved AI draft should not be sent after its content or underlying facts changed. Bind review to the exact revision and check freshness at the sending boundary.
Draft approval concerns specific content
A reviewed draft is a proposed message with a particular recipient, content, attachments, and business meaning. Approval of an earlier version does not automatically authorise a later edit, even when the subject line remains the same.
For a proposed model, store draft identity, revision, recipient scope, relevant source versions, approval evidence, and intended send operation. Use a stable content representation or equivalent version control appropriate to the system.
Do not let a background sender read whichever draft text happens to be current while relying on an old approval flag. The approved revision and executable payload must remain connected through trusted state.
Reproduce a fictional quote-message change
An AI drafts an email saying a service request is awaiting scheduling. A manager approves it. Later, another process changes the draft to say a technician will arrive Thursday. The old decision must not authorise that stronger commitment.
A second fixture leaves the wording unchanged but changes the appointment state: the time is no longer confirmed. The send gate should follow the defined freshness policy and prevent a misleading message even though the draft revision did not change.
| Stale-draft condition | Expected behaviour |
|---|---|
| Content edited after review | Fresh approval where required |
| Recipient changed | Recheck scope and decision binding |
| Attachment replaced | Invalidate or re-review under policy |
| Underlying commitment changed | Revalidate factual claim |
| Duplicate send trigger | Same logical send operation |
Inspect the exact delivered artifact in a fake mail sink. A database row showing “approved” is not proof that the sent message matched what the reviewer saw.
Preserve useful feedback and history
When a draft becomes stale, retain reviewer comments and source references so the next revision does not lose context. Explain why sending stopped and who owns the update. Avoid silently discarding a time-sensitive customer response.
Strict binding can cause repeated review when many edits are harmless. Define which changes are consequential and make the policy explicit. Do not permit cosmetic-change exceptions to become a route for unreviewed promises or recipient changes.
This sending design is proposed. The voice integration case study describes related workflow experience. Pair it with approval withdrawal.
Stale AI draft FAQ
Is a Boolean approved flag enough?
No. It must identify the content, recipient, scope, and revision the decision covered.
Can unchanged text become stale?
Yes. The underlying appointment, policy, or customer facts may have changed.
What proves safe sending?
Verify decision binding, current facts, recipient permission, and the actual delivered content under controlled fixtures.