Human in the loop
How to let AI reply to emails with human approval.
Let AI reply to emails with human approval by giving the agent an API key with an approval hold: it reads each thread waiting on an answer and writes the reply, and the reply waits for a person to approve it before it is sent.
The agent does the reading and the writing. A person keeps the send, from one queue, and the approved reply goes out in the original thread.
202 {
"status": "pending_approval",
"thread_id": "9b2e41c0-…",
"approval_id": "c71d08aa-…"
}Approve sent in the original thread
Reject never sentThe loop
Five steps, two of them a person's.
A reply arrives
It joins the thread of the message it answers, and the thread is flagged as awaiting a reply. Out-of-office replies and bounces never set the flag.
The agent finds what is waiting
list_threads with awaiting_reply, skipping any thread with pending_reply, whose answer is already drafted.
The agent drafts the answer
It reads the thread with the quoted history removed and replies in place. On a key with requires_approval the reply is held and comes back as pending_approval.
reply_to_message, on a held key202 { "status": "pending_approval", "thread_id": "9b2e41c0-…", "approval_id": "c71d08aa-…" }A person approves or rejects it
Under Approvals in the dashboard, or from the API with a key that has no limits of its own. A draft nobody decides expires after 72 hours.
The reply goes out in the thread
With the right threading headers, so it lands under the customer’s message. Suppression and the plan allowance are checked again at release, because a hold can last days.
Rules that keep it honest
What the agent cannot do, by design.
Approve its own draft
The drafting key is refused, and so is any key with a limit of its own and any OAuth sign-in.
Retry its way past the hold
A held send is a 202, not an error. Sending again only queues a second draft for the same person to reject.
Lift the hold
A key may not edit itself, and key management has no MCP tool, so the hold stays until a person changes it.
Start every agent on a hold. Once a month of its drafts has gone out unedited, move it to a daily cap and an allowlist instead, and keep the hold for the agents that write to customers.
Why not just Resend?
Resend has an MCP server too. The difference is what happens after the send.
The reply is a thread, not a webhook
SendRaven: Joined on Message-ID to the message it answers. Resend: a received-email event; you fetch the message and set reply headers yourself.
Per-key daily cap and recipient allowlist
SendRaven: daily_send_limit and allowed_recipients on the key, checked on every send. Resend: keys scope to a domain and to send-only; no cap or allowlist.
Human approval hold on a key
SendRaven: requires_approval drafts the message and waits for a person. Resend: not offered.
From resend.com, last checked 16 September 2026. The full comparison, including where Resend is the better choice.
Questions
Can AI reply to my emails but wait for my approval?
Yes. Connect the agent with an API key that has requires_approval, and each reply it writes is held as a draft under Approvals until a person releases or rejects it; released replies go out in the original thread.
How do I approve an AI agent's email replies?
In the SendRaven dashboard under Approvals, or through the API with a key that carries no limits of its own. The agent's own key cannot approve anything.
What happens if nobody approves the AI's reply?
It expires after 72 hours and is never sent. The thread stays awaiting a reply, so the agent can draft again on its next run.
Why does my AI agent draft the same reply twice?
Because a thread stays awaiting a reply until a reply is actually sent. Skip threads where pending_reply is true, which means an answer is already held or scheduled, and derive the Idempotency-Key from the message being answered.

The agent writes. You press send.
Free includes 3,000 emails a month, inbound included, and is never charged without an upgrade.

