“My judgment is the bar.” How an AI email assistant drafts replies without sending them.

Why do important emails get buried?

Important emails get buried because they arrive mixed in with everything else. In my outbound mailboxes, that’s out of office replies, bounces, auto replies, and warmup email, the automated traffic that keeps a sending mailbox in good standing. At a benefits agency, it’s carrier notices and system mail piled on top of the one client asking about an ID card.

The busier the inbox, the longer a real question waits.

Why not let AI send email replies on its own?

AI shouldn’t send most email replies on its own because most emails need judgment the AI doesn’t have. The predictable ones are a short list, and the answers to those get worked out ahead of time. Everything else depends on the person, the history, and what I know that isn’t in the email.

Training a reply agent to my standard takes a long time, too. So I take away its send permission. It can read, and it can write drafts. That’s all it can do.

What do most agencies try first?

Most agencies try inbox rules and saved replies. Filters move carrier notices into a folder. A saved reply answers the common question. Both help, and both break on anything that doesn’t match the rule exactly. The other common try is the AI drafting built into the email app, which usually works one message at a time, when you ask.

Wanna fix it?

A buried, risky inbox gets fixed with an AI email assistant that checks the mailbox on a schedule, throws out the noise, sorts real replies by what the sender wants, drafts an answer, and waits for a person to approve it. Here’s how mine works on my outbound mailboxes. It runs on Claude.

1. Check every 5 minutes, touch nothing. It reads 10 mailboxes every 5 minutes. It never marks a message read, moves it, or deletes it.
2. Filter the noise first. Automated traffic gets dropped. Out of office replies, auto replies, bounces, and “this person left the company” notices get sorted out of the way.
3. Read what the sender wants. Each real reply gets tagged as interested, questioning, objection, rejection, referral, or unclear, with a confidence score and a one line summary.
4. Draft by the rules. Plain text, two to five sentences, my sign off. A banned phrase list (“circle back,” “touch base”). It may never invent a meeting time, a price, a case study, a client name, or a result. A rejection gets a gracious close with no pitch.
5. Wait for a click. Every draft lands in a queue. I approve it, edit it, or hit Redraft with a note like “shorter.” Only my click sends it, from the same mailbox, in the same thread.

Unknown senders get added as new contacts, so nothing falls through. When someone replies, their email sequence stops.

What does it look like once it’s running?

Once it’s running, the inbox sorts itself and I only see the messages that need me. In its first 28 hours, it set aside 1,019 automated emails, plus 33 system messages and bounces, without me opening a single one.

When a real reply comes in, it lands in the queue with a draft already written. For the few set issues, the template does most of the work. Everything else is a starting point, and I rewrite it until it’s something I’d send.

For a benefits agency, I’d point the same setup at the service inbox: templates for the few questions that come in every week, drafts for the rest, and an account manager’s click on every send.

Can you do this yourself?

Yes, most of it. Set up filters for carrier notices and auto replies. Write saved replies for your five most common questions. Turn on your email app’s AI drafting and treat every draft as a first pass you edit.

It stops working when you have several mailboxes, when a real reply sits for hours under the volume, or when you want replies sorted by intent across a team. One inbox, filters and saved replies. Several inboxes and a team answering them, build the assistant.

Quick answers

What is an AI email assistant?

An agent that reads incoming email, filters the noise, sorts replies, and drafts answers for a person to approve.

Should an AI email assistant send replies on its own?

For most email, no. Chris Cordon’s can read and draft, and only his click sends.

How often should it check the inbox?

Mine checks every 5 minutes.

What should it filter out?

Automated traffic, out of office replies, auto replies, bounces, and notices that a contact left.

What should AI email replies never include?

An invented meeting time, price, case study, client name, or result.

Can it handle more than one mailbox?

Yes. Mine reads 10.