How to Implement Human-in-the-Loop for AI Emails

System prompts telling an AI to "ask for permission" are fundamentally unreliable because LLMs can rationalize their way past a text-based instruction. For high-stakes actions like sending emails—where a wrong tone or a factual hallucination can damage a professional relationship—you need a structural gate, not a prompt. A hard data dependency ensures the sending function is physically unreachable until an external system returns an approved status.

This approach transforms the AI workflow from a risky "fire and forget" system into a controlled pipeline: the agent drafts, a human audits/edits, and only then does the code execute the send command.

The Logic Flow

Instead of a direct call to an SMTP server, the agent follows this sequence:

1. Draft Generation: The agent produces the content.
2. External Holding: The draft is pushed to a review system (like Impri) via API.
3. Human Intervention: A reviewer reads the draft, makes necessary edits to the body or subject, and hits "Approve."
4. Execution: The system polls for the decision; if approved, it sends the final edited version, not the original AI draft.

Real-world Implementation: Node.js & TypeScript

Here is a practical tutorial on how to build this approval gate using nodemailer and a review API.

import nodemailer from "nodemailer";

const IMPRI_KEY = process.env.IMPRI_API_KEY!;
const IMPRI_BASE = "https://api.impri.dev";

async function sendWithApproval(opts: {
 to: string;
 subject: string;
 body: string;
 expiresIn?: number; // seconds; default 3600
}): Promise<"approved" | "rejected" | "expired"> {
 // 1. Push the draft to review
 const push = await fetch(`${IMPRI_BASE}/v1/actions`, {
 method: "POST",
 headers: {
 Authorization: `Bearer ${IMPRI_KEY}`,
 "Content-Type": "application/json",
 },
 body: JSON.stringify({
 kind: "email.send",
 title: `Email to ${opts.to}: ${opts.subject}`,
 preview: {
 format: "markdown",
 body: `**To:** ${opts.to}\n**Subject:** ${opts.subject}\n\n---\n\n${opts.body}`,
 },
 editable: ["preview.body"], 
 expires_in: opts.expiresIn ?? 3600,
 }),
 });

 if (!push.ok) throw new Error(`Push failed: ${push.status}`);
 const { id: actionId } = await push.json();

 // 2. Poll until human decision
 let result: { status: string; decision?: { final_preview?: { body: string } } };
 for (;;) {
 const poll = await fetch(`${IMPRI_BASE}/v1/actions/${actionId}`, {
 headers: { Authorization: `Bearer ${IMPRI_KEY}` },
 });
 result = await poll.json();
 if (result.status !== "pending") break;
 await new Promise((r) => setTimeout(r, 10_000));
 }

 if (result.status !== "approved") {
 return result.status as "rejected" | "expired";
 }

 // 3. Send the human-approved version
 const approvedBody = result.decision!.final_preview!.body;

 const transporter = nodemailer.createTransport({
 host: process.env.SMTP_HOST,
 port: 587,
 auth: { user: process.env.SMTP_USER, pass: process.env.SMTP_PASS },
 });

 await transporter.sendMail({
 from: process.env.SMTP_FROM,
 to: opts.to,
 subject: opts.subject,
 text: approvedBody,
 });

 // 4. Update audit log
 await fetch(`${IMPRI_BASE}/v1/actions/${actionId}/result`, {
 method: "POST",
 headers: {
 Authorization: `Bearer ${IMPRI_KEY}`,
 "Content-Type": "application/json",
 },
 body: JSON.stringify({ status: "success" }),
 });

 return "approved";
}

By treating human approval as a mandatory API response rather than a prompt suggestion, you create a fail-safe LLM agent deployment that is actually production-ready.