Sales and follow-up by email with an AI agent
An AI sales agent writes to prospects from its own inbox, schedules follow-ups, and answers replies in the thread. The risk in this job runs outward: an agent with access to company data can put a password, a credential or customer data into a message, and sent mail cannot be taken back.
Updated
How an AI sales agent sends outbound email
An AI sales agent, often called an AI SDR, sends a message, schedules the follow-up, and replies when the prospect answers.
- Give the agent its own inbox.
- Decide who approves outgoing mail. With
send_approvalatfirst_contact, the first message to any new recipient waits for a human. - Send now, or schedule a time 30 seconds to 30 days ahead.
- When the prospect replies, read the thread context and reply in the thread.
await mail.inboxes.update(riley.id, { sendApproval: "first_contact" });
await mail.messages.send(riley.id, {
to: "taylor@lumen-outfitters.example", subject, text,
sendAt: new Date(Date.now() + 4 * 86_400_000), // screened again when it goes
});{
"name": "send_message",
"arguments": {
"inbox_id": "inb_4x8kq2m9tn7vb3hcd5ze",
"to": ["taylor@lumen-outfitters.example"],
"subject": "Pilot for Lumen Outfitters",
"text": "Hi Taylor, thanks for Tuesday's call...",
"send_at": "2026-11-24T17:00:00Z"
}
}The send_at value is a sample time: use one from 30 seconds to 30 days ahead. A scheduled message is screened when it is due, through the same checks as mail sent now. Refused then, it returns to drafts.
The follow-up and the message with a password
The job has two sample messages, both outgoing: a follow-up, which got a deliver verdict, and a login message, which got reject. The verdicts below are the ones the live engine gave on November 20, 2026.
Both are from Riley, the agent, at riley@trotta.email, to Taylor Reed at Lumen Outfitters. The follow-up, "Pilot for Lumen Outfitters", thanks Taylor for Tuesday's call, says the pilot plan is ready, and asks about a walkthrough on Monday at 10.
The other, "Your pilot login", says the pilot is ready and tells Taylor to sign in with a team account. The login name and the password are written out in the body.
| Message | Verdict | Summary, as the engine wrote it |
|---|---|---|
| Pilot for Lumen Outfitters | deliver, level safe | Nothing sensitive found in this message. |
| Your pilot login | reject, level danger | The message contains a password or credential. |
The login message carries the signal "Outgoing mail contains a secret". For outgoing mail, reject means the message is not sent: the call returns 422 outbound_blocked with the verdict, and the message is stored as blocked.
In a mailbox that sends first and checks after
A mailbox that sends first and checks after delivers the login as written. The password is in the prospect's mailbox, and in every place that mailbox is copied or forwarded, before any report mentions it. A secret sent by mail is exposed for good, so the repair is to change the password.
Replies from prospects are inbound mail from strangers, and in a Trotta Mail inbox they are screened before delivery like any other message. The outbound screen also looks for replies that an injected instruction is steering to an address the inbox has never written to. See data exfiltration through agents.
Setting up a sales agent well
A sales agent needs its own inbox, a human on approval where you want one, and no secrets in its context.
- Claim the agent before it prospects. An agent that signed itself up cannot start a conversation with a new outside address until a human claims it.
- A human on first contact.
first_contactholds the first message to each new recipient aspending_approvaluntil a human approves or discards it.alwaysholds every send. - Or drafts alone. With
agent_writeatdrafts, the agent writes and a human sends. The draft is screened with the agent rules when it is sent, whoever presses send. - No secrets in the agent's context. The outbound screen looks for secrets and sensitive data, and no filter catches everything. An agent that never holds a password cannot send one.
- Your own domain, set up fully. Your own domain is part of the paid plans. Publish its SPF, DKIM and DMARC records before you send from it. See connect a domain.
Questions
Can an AI sales agent send email from its own inbox?
Yes. The agent sends, schedules and replies through the API or the MCP tools, from an address on trotta.email or on your own domain on a paid plan. Outgoing mail is screened before it leaves.
Can a human approve an AI SDR's emails before they send?
Yes. Set send_approval to first_contact to hold the first message to each new recipient, or always to hold every one. Held sends wait as pending_approval until a human approves or discards them.
What happens if the agent tries to email a password?
In the saved sample the message got a reject verdict with level danger. A rejected send returns 422 outbound_blocked and the message does not leave.
Can an AI agent schedule follow-up emails?
Yes. Add a send time 30 seconds to 30 days ahead. The message is screened when it is due, and it can be cancelled before then.