Sort incoming mail and draft replies from templates

What it is

A two part workflow that watches an inbox, decides what each message is about, and writes a first draft back using wording you already approved. You keep the categories and the canned text in a spreadsheet. The automation reads that sheet, matches the message, and drops the draft into your mail tool for a human to send.

The sheet is the brain. Columns for category name, a short description of what belongs in it, and the template reply. Add a row and the automation picks it up on the next run. No code changes, no redeploying anything.

When to use it

Shared inboxes where the same six questions arrive all day. Support queues, sales enquiries, a bookings address, a general contact form that forwards to mail. If most replies start with the same three sentences and differ only in the details, this saves the typing.

Skip it when every message needs judgement from scratch, or when the reply must include figures pulled from a system the automation cannot see. A wrong draft costs more time than a blank one.

How to set it up

1. Build the sheet first. One row per category. Keep the description field honest about edge cases, since the matching step leans on it. Write templates with clear placeholders like the customer name or the order number.
2. Connect the mail account and add a trigger that fires on each new message in the target folder or label. Filter out your own sent mail and obvious spam before the next step.
3. Add a step that fetches the sheet rows so the categories are available downstream.
4. Send the message body plus the category list to a language model step. Ask for one category back, plus a confidence note if the match is weak. Instruct it to return a fixed fallback value when nothing fits.
5. Add a branch. On a confident match, fill the chosen template with the details from the email. On a fallback, leave it untouched and flag it.
6. Create the draft in the same thread, never send it. Set the recipient and subject from the original message.
7. Log the message id, the chosen category and the timestamp in a second sheet. That log is how you spot categories that keep getting misread.

What a good result looks like

A draft sitting in the thread within a minute, addressed to the right person, with the right template and the specifics filled in. Your team reads it, adjusts a line, and hits send. The fallback cases land in a separate view so nothing gets answered with the wrong script.

After a week, check the log. If one category keeps winning when it should not, tighten its description. If a category never appears, its template is probably too broad to be chosen. Two rounds of that and the drafts need almost no editing.