AI in email marketing means using AI models to decide who gets an email, write it, pick when it goes out and predict what each customer will do next. Writing works from day one. The predictive features need data first: hundreds of buyers, months of orders or very large sends.
For a seller on paid traffic, the useful order is: AI drafts, a person checks the offer and the claims, and a split test judges the email on orders per recipient.
I'd start with the abandoned checkout email, because it sends every day and the result shows up in sales within a week.
Most guides to AI in email are written for brands with a big list and years of order history. A lot of paid-traffic sellers have neither. They have an offer that launched in the spring, a list built from checkout and lead forms, and a Q4 calendar that already looks crowded.
I spent nine years as CTO of a direct-response company, and the line I wrote about it on our about page still sums up email for me: "Sales that didn't reach the CRM. Recovery calls that didn't fire." None of the leaks in that line is a copy problem. Each one is a buyer who should have heard from you at the right moment and didn't. AI helps with the copy, and it helps even more with getting the right email to the right person.
What AI does in email marketing
AI shows up at four points in an email program. At each one it does the fast part, and something stays with you.
| Step | What AI does | What you still check |
|---|---|---|
| Who gets it | Turns "bought twice, nothing in 60 days" into an audience | Who should be left out: refunders, recent buyers, staff |
| What it says | Drafts subject lines, body copy, variations | The offer, the price and every claim |
| When it goes out | Send time optimization from past engagement | Whether your list is big enough for it to learn |
| How it did | Predictions such as next order date or churn risk | Orders and revenue per recipient from a real send |
The first row is where I'd put the most attention. A good email to the wrong audience loses money twice: it annoys people who just bought, and it trains inbox filters that your mail gets ignored. Describing an audience in a sentence used to need a data person. Now it takes a minute, which makes it the cheapest win AI gives a small team.
What each AI email feature needs before it works
Every predictive feature learns from your own history. The tool makers say how much they need, and it is more than many new offers have.
The table compares AI features inside two common email tools, by what each needs from your account. The rows overlap in practice: a brand big enough for predictions usually also uses AI drafts.
| Tool and AI feature | What it does | What it needs |
|---|---|---|
| Klaviyo predictive analytics | Predicted lifetime value, churn risk, expected date of next order | At least 500 customers with orders, 180+ days of order history, some customers with 3+ orders (Klaviyo Help) |
| Klaviyo Smart Send Time | Tests send hours, then sends at the best one | Campaigns of 12,000+ recipients, and no A/B test on the same campaign (Klaviyo Help) |
| Mailchimp Send Time Optimization | Picks a send time within 24 hours of your date | Standard plan or higher, enough past sends; not for automated emails (Mailchimp Help) |
| Mailchimp Email Content Generator | Drafts subject lines, headings and paragraphs, three options at a time | Standard plan or higher, the new builder; listed as a beta (Mailchimp Help) |
Picture a supplement offer that launched in June on Meta and has 300 buyers by Black Friday. Predicted lifetime value won't switch on for it yet, and send-time testing won't run on its list. Paying extra for an "AI tier" in that position buys features the account can't feed.
What that seller can use today is AI for writing and for describing audiences. Neither needs history. Both save hours in a busy quarter.
Writing emails with AI without losing the sale
AI copy is only as specific as the brief. Give a model the product name and it writes the average email for that category, which is the email your competitors are already sending.
A brief I'd hand to any model, for a reminder to people who left the checkout:
Write 3 subject lines and 1 short email for buyers who left our checkout. Product: [name], [price], [guarantee terms]. Their words (from reviews and support emails): "[quote]", "[quote]". Main reason people hesitate: [shipping time / price / doubt it works]. Rules: no claim we can't prove, no fake urgency, plain text, under 120 words, one link back to the checkout.
- Feed it your buyers' words. Reviews, refund reasons and support emails hold the phrases that sell. The model can't find those on its own.
- Keep the offer yours. Let AI write around the price and the guarantee, and never let it invent a discount or a deadline you don't run.
- Check every claim. The FTC's CAN-SPAM compliance guide says "The subject line must accurately reflect the content of the message." A model chasing opens will happily write a subject line the email doesn't back up.
- Ask for variety on purpose. Ask for drafts with different angles (the problem, the proof, the risk reversal), so a test compares ideas and not adjectives.
For health and supplement offers, the claim check is the step I'd least want to hand off. An email gets forwarded and screenshotted, and a promise that looked fine in a draft sits in someone's inbox forever.
Testing AI email copy: count orders, not opens
Most AI subject line tools score drafts on predicted opens, and the open is the weakest number in email. Apple Mail Privacy Protection loads messages in the background, so emails to those users "register an open regardless of whether the recipient has read the email message or not" (AWS Messaging blog, 2021). That makes the open rate a poor judge of AI copy.
Here is a subject line test on 20,000 recipients, split evenly, with a $59 order (illustrative numbers):
| 10,000 recipients each | Your subject line | AI subject line |
|---|---|---|
| Open rate | 40% | 44% |
| Orders | 100 (1.0%) | 120 (1.2%) |
| Revenue | $5,900 | $7,080 |
| Revenue per recipient | $0.59 | $0.708 |
The open rate gap is real on this list: the sample size calculator needs about 2,400 recipients per side to detect 40% against 44%. The order gap is the one that pays, and it needs about 43,000 recipients per side to tell 1.0% from 1.2% reliably (95% confidence, 80% power). On 10,000 each, twenty extra orders can easily be luck.
So on a list this size, I'd test bigger differences: a new angle, a new offer in the email, a different send for refunders. Those gaps are wide enough to show up in orders. And when a test does run, keep a small holdout group that gets nothing, so you can see whether the email sold anything beyond what those buyers would have bought anyway.
Sending more in Q4: the inbox rules still apply
The real risk of AI in email is volume. When an email costs a minute to write, it is tempting to send one every day of Black Friday week, to everyone.
Gmail's email sender guidelines ask senders to keep the spam rate in Postmaster Tools below 0.10% "and avoid ever reaching a spam rate of 0.30% or higher". Anyone sending more than 5,000 messages a day to Gmail also needs domain authentication and one-click unsubscribe. The full list is in our post on email marketing campaigns.
- Send more to fewer people. Extra Q4 sends go to recent buyers and recent clickers. The people who haven't opened anything since summer get one or two.
- Watch complaints daily in November. A send that sells well and doubles complaints is borrowing from December.
- Write early. Draft the whole BFCM sequence with AI in early November, then spend the saved time on the audiences and the claim check.
Paid-traffic lists carry one extra risk. Addresses typed into a checkout are sometimes mistyped or fake, and every bounce counts against your deliverability. Validating addresses when they're collected costs less than cleaning a list after a bad send.
AI audiences and email automations in ElasticFunnels
Audiences in ElasticFunnels cover the first row of that table: getting the right buyers into an email without exporting spreadsheets. An audience is built from your own orders and page data, so it already knows who bought, who took the order bump, who refunded and who left the checkout.
- Describe it in a sentence. The AI builder takes requests such as "bought at least twice in the last 90 days", "abandoned checkout, then bought within 7 days" or "bought 3-6 months ago, nothing since", and can leave out internal email domains. It uses AI credits. The rules builder does the same without them.
- It stays current. The audience refreshes on its own as people buy and refund, and its member table exports to CSV.
- Emails on real events. Automations start on a new purchase, an abandoned cart or a refund, then send an email through SendGrid, Mailgun, Mailrelay or Maropost, or create or update the buyer in a connected email tool such as Klaviyo or Mailchimp.
- Test two emails on real buyers. The Split Test step in an automation sends each version to part of the flow, so you can compare an AI draft against your current email.
If you already send through Klaviyo or Mailchimp, keep their AI for drafting and let the funnel decide who gets what, since the funnel is where the order data lives. The automations page shows the triggers and steps. Every feature is on every plan.
- Spend AI on the audience before the copy. A plain email to buyers who are due to reorder usually beats a clever one sent to everyone.
- Skip the "AI tier" of an email tool until your account meets its data requirements. Under 500 buyers, you're paying for a switch that stays off.
- Never let a model pick the discount. It will choose the number that gets clicks, and you'll pay for it on every order in the quarter.
- If a subject line test is judged on opens, ask what it did to orders before you roll it out.
AI made the draft cheap, and that helps in a quarter when everyone sends more. What it didn't make cheap is knowing which buyers should get which email, and telling a real result from a handful of lucky orders. Put the time AI saves into those two jobs, and the email program gets better instead of just louder.
