Reviewing AI drafts
Approve, edit, and give feedback on AI-generated inbox drafts before auto-send.
Reviewing AI drafts
When Handlet generates a draft reply, you stay in control. Your approvals, edits, and correctness feedback build Stage 2 training evidence per intent before auto-send can unlock for that intent.
Step-by-step workflow
- Review the draft — read the full message and check the detected intent label (for example Quote, General enquiry, Follow-up).
- Approve or edit — send when ready using Approve & send, or change wording with Edit and send your version.
- Optional correctness feedback — after send succeeds, look above the composer for Was this correct? Confirm or reject accuracy, or choose Not now if you are busy.
Send is complete after step 2. Step 3 never undoes a sent message.
Primary actions
| Action | When to use |
|---|---|
| Approve & send | The draft is ready to go to the customer as-is. |
| Edit | Change wording or assumptions, then send when ready. Edits help Handlet learn what to improve. |
| Was this correct? | After you send, confirm whether the draft was accurate (optional Not now if you are busy — send is never undone). |
What counts toward training (Stage 2)
Each action below counts as one validated message for the draft’s intent when calibration progress is enabled:
| UI action | Calibration signal | Counts? |
|---|---|---|
| Approve & send | Approved and sent | Yes |
| Edit then send | Edited before send | Yes |
| Yes, looked correct | Correctness confirmed | Yes |
| No (optional reason) | Correctness rejected | Yes |
| Not now | Dismissed — no answer recorded | No |
Negative feedback (No) still counts as training evidence — it shows you reviewed the draft, not that it was correct.
Intent labels
Draft review surfaces show the detected intent so you know which type of message you are judging. Training progress is tracked per intent — quote approvals do not count toward follow-up training, and vice versa.
How feedback is used
Your actions per intent build trust signals Handlet uses before conditional auto-send can unlock for that intent. See Calibration and confidence bar for minimum counts and progress bars on Channels & Tone.
Dismissals (Not now) are tracked separately from yes/no answers so Handlet knows when you skipped feedback without treating it as confirmation.
When auto-send can still be blocked
Even after Training complete for an intent, auto-send may block when:
- The master switch or per-intent toggle is off
- Confidence is below your threshold (minimum 85%)
- Pricing, booking, or complaint guardrails fail
- Platform rules veto the intent or conversation
Check the Action Queue for blocked-send reason codes.
Related articles
Related policies
- Conditional autonomy — human-in-the-loop defaults until calibration is sufficient.
- Auto-send governance — platform guardrails that always apply.