# How to Use AI to Improve Your Email Marketing

By Justin Novak · https://www.tminus.ventures/playbook/adversarial-copy-playbook

Your sent folder is a dataset. Every campaign you've ever sent holds a lesson about what makes your list open, click, and book a meeting. No human is going to read 123 campaigns and cross-reference every send. AI will, in an afternoon. Then a second set of agents tries to tear every finding apart, so only the real patterns survive.

## Why it works

Think about what's actually in play in one email:

- the name of the sender
- the subject line
- the time of day it went out
- the day of the week
- the length of the email
- a call to action versus no call to action
- a picture versus no picture
- a link versus no link, and how that link is phrased

That's eight dimensions on a single send. Across 123 campaigns, it's a matrix with thousands of cells, joined to who actually opened, clicked, and converted. None of this was feasible before AI. Not because the data didn't exist. Because nobody was ever going to do those joins by hand. Too much for a person. Trivial for a model.

We ran exactly this on a client's outbound program: 123 campaigns, over 210,000 per-recipient delivery rows, joined at the person level to CRM outcomes. Out the other side came a ranked playbook of 18 copy rules.

## The adversarial step

Here's the part most people skip. A first-pass analysis will lie to you, because it confuses the list with the copy. Your best subject line probably went to your warmest list. Of course it won.

So after the mining pass, we set up separate verifier agents with one job: attack every finding. "That was the list, not the copy." "That was the send time, not the subject." Every rule had to survive cross-examination against the data. Most didn't. The ones that did are the playbook.

## What survived

Three of the surviving rules, measured in conversions per 1,000 emails sent:

![Bar chart of three copy patterns and their measured lift across 123 campaigns](/playbook/img/adversarial-copy-playbook/findings.svg "Three of the 18 surviving rules, conversions per 1,000 sent")

A first line about the reader (their company, their world, not yours) converted at 1.21 per 1,000 against 0.30 for a generic opener. One single link phrased as a calendar invite beat a bare link 0.92 to 0.40. Afternoon sends to cold lists beat mornings 1.25 to 0.29. No guessing anywhere. It all came out of data we already owned.

One more that changes how you schedule follow-ups: 90% of conversions landed within 24 hours of a send. If a campaign is quiet on day two, it's done. Send the next one.

## How to run it

1. Export everything: campaigns, subject lines, send times, body copy, and per-recipient outcomes from your email platform and CRM.
2. Have AI classify every campaign across the dimensions above and join the results at the person level. That's hours of agent time, not weeks of analyst time.
3. Run the adversarial pass. Separate agents argue against every finding. Keep what survives.
4. Write the survivors down as the playbook your next campaign has to follow. Keep feeding new sends back in so the rules sharpen over time.

The target metric is whatever a conversion means to you. For us it was registrations that became booked meetings. Swap in demos, replies, or signed proposals. The method doesn't care.