Automation

Building an Automated Email and Document Workflow

Published Nov 9, 2026

Generate an email draft and Markdown record from the same checked data without sending or accessing a real mailbox.

Overview

For learners who want to complete this practical exercise with inspectable sample data and explicit checks.

Separate AI drafting, content checking, and artifact generation.
Create a structured MIME email with the standard library.
Organize output files with stable record IDs.

You will build: DEMO-1.eml and DEMO-1.md containing the same checked response.

Prerequisites

Python 3.10+ and an editor; optionally a chat assistant for drafting.

The script uses no SMTP, OAuth, or email account. An optional assistant may have limits/fees; all example records are fictional.

Step-by-step tutorial

Conceptual practice workflow; not measured results.

Step 1. Use a sample request instead of a mailbox

Use a simple request for a getting-started guide. The fictional policy only says that the guide is under review and has no confirmed completion date. Static input makes the workflow reproducible without granting mailbox access.

Step 2. Draft with AI using an explicit policy

PROMPT · INSTRUCTIONS

Request: "Please send the getting-started steps."
Fictional policy: the guide is under review; no completion date is confirmed. Draft a polite 1–2 sentence response. Add no link, attachment, or promise of later delivery. Do not send anything.

Read and correct the draft, then place your accepted text in approved_reply below. This field records your own exercise-level content check; it does not fabricate an external reviewer or workflow approval.

Step 3. Generate both artifacts from one source

Save this as email_workflow.py. EmailMessage handles header/body structure, and example.com addresses are for illustration only. The .eml is a local artifact; there is no email-sending operation.

PYTHON

import json
from email.message import EmailMessage
from pathlib import Path

RECORDS = [{"id": "DEMO-1", "subject": "Onboarding guide",
            "request": "Please send the getting-started steps.",
            "approved_reply": "Thank you. The guide is being reviewed; no date is confirmed."}]

def build(records, out):
    out.mkdir(parents=True, exist_ok=True)
    seen = set()
    for record in records:
        ident = record["id"]
        if not ident.startswith("DEMO-") or not ident[5:].isdigit() or ident in seen:
            raise ValueError("invalid or duplicate id")
        seen.add(ident)
        message = EmailMessage()
        message["From"] = "training@example.com"
        message["To"] = "recipient@example.com"
        message["Subject"] = "Draft: " + record["subject"]
        message.set_content(record["approved_reply"])
        (out / (ident + ".eml")).write_bytes(message.as_bytes())
        document = ("# " + ident + "\n\nRequest: " + record["request"] +
                    "\n\nApproved draft: " + record["approved_reply"] + "\n")
        (out / (ident + ".md")).write_text(document, encoding="utf-8")
    return len(seen)

if __name__ == "__main__":
    print("Created", build(RECORDS, Path("drafts")), "draft pair; nothing sent")

SHELL

python3 email_workflow.py

Expected result

Created 1 draft pair; nothing sent

Step 4. Inspect output and rerun

Open drafts/DEMO-1.md and inspect the .eml as text or as a draft if your mail app supports it. Compare subject/body and check that no completion date was promised. Rerunning overwrites the same ID pair instead of duplicating it. This is not transactional storage: a failure can leave a partial pair that requires a corrected rerun.

Step 5. Plan real integration one permission at a time

For a real mailbox, choose Email Trigger (IMAP) or another organization-approved connector, configure separate credentials, filter a test folder, and persist message IDs for deduplication. Keep draft creation separate from sending; do not attach SMTP merely because a model suggests it. Check retention policy before sending email content to AI.

Verification

Both artifacts use exactly the same approved_reply and make no network call.
Rerunning does not create additional IDs.

Troubleshooting

Email line breaks or Unicode look wrong

Cause: The MIME message was parsed by naive string splitting.

Fix: Use email.parser or a MIME-aware client and inspect the decoded body.

An old output remains after a batch fails

Cause: The exercise has no batch transaction or automatic cleanup.

Fix: Use a dedicated test directory and design validation/error handling before replacing real outputs.

Exercises & extensions

Add two requests, preserve unique IDs, and verify the output-pair count.
Write a test parsing .eml and comparing its decoded body with Markdown.

The exercise does not read mail, send mail, or create Google Docs. A local .eml is not evidence of email delivery; real integration requires separate checks.