5 AI Workflows to Test in Your Email Program, and What Each One Needs to Work
Last updated October 2026 by Lily Newman
The AI worth testing in email right now is narrow: send-time optimization on campaigns, drafting with a human editor, predictive segments built on your order data, churn-risk winbacks, and AI as a quality check on your own emails. Each one has a data requirement that decides whether it will work for you. Here are the five, what they need, and how to test each.
1. Send-time optimization, for campaigns, at scale
Send-time optimization picks a time of day for each recipient. It is not the same as choosing how long to wait after a customer abandons a checkout.
Klaviyo's version, Smart Send Time, is available for email campaigns only. The test phase needs a recipient list of 12,000 or more, and it works by sending at random times first and then clustering around the best time it finds. You cannot A/B test it, and Klaviyo advises against using it for time-sensitive content or around major holidays (Klaviyo). One more caution: tools that read open data can be skewed by Apple Mail Privacy Protection unless the vendor accounts for it (Litmus), so ask yours how it does.
In flows, the customer's action sets the clock. Test the delay, such as 30 minutes against an hour after checkout, not the time of day.
How to test it: Run it for a few sends against a fixed send time. Compare click rate and revenue per recipient.
2. Drafting subject lines and copy
Until AI gets better at nuance and really understanding your brand, use it for variety and volume and let a person choose. Give AI tools your best-performing subject lines and a few lines on your voice, ask for twenty options, and expect to keep two. Left alone, it drifts toward generic phrasing, and it will happily invent a discount or a claim you cannot make. A human checks every price, date, and promise.
How to test it: Run AI-assisted variants against your human-written control. Look at opens directionally. Judge on clicks and revenue per recipient.
3. Predictive segments from your order data
Klaviyo's predictive analytics estimate each customer's historic and predicted lifetime value, churn risk, average time between orders and expected date of next order. They need at least 500 customers with orders, 180 days of order history and orders in the last 30 days, and some customers with three or more orders (Klaviyo).
Two uses are worth the setup. Historic value can build a VIP segment. Expected date of next order can start a second-purchase nurture for one-time buyers. Klaviyo itself cautions against counting down to the expected date for repeat customers, because it can drive unsubscribes.
How to test it: Send the same offer to a predicted-value segment and to your current VIP segment, and compare revenue per recipient.
4. Churn-risk winbacks
Klaviyo's churn risk is based on how many orders a customer has placed and how often. That is a better foundation than "no open in 30 days," which Apple's preloading makes unreliable. Build winbacks on purchase behavior. Customers with a light lapse get content and a reason to return. Customers with a deep lapse get the incentive. Offer the discount last, not first.
How to test it: Run a flow with risk-tiered branches against your single-track winback. Measure reactivated buyers and margin after discount, not opens.
5. AI as your quality check
Before a send, paste the email into an assistant and ask for a two-sentence summary. Gmail is rolling out AI summaries of long email threads, free, starting in the U.S. in English (Google), and if your offer is missing from the summary, your opening is buried. Then ask it to list every date, price and claim in the email, and check each one against your promotional calendar yourself. It catches typos too, though it cannot click your links, so test those by hand.
Keep customer data out of any tool your company has not approved. Paste in the email, not the list.
Where a person stays in the loop
Prices, offers and product claims
Consent language and unsubscribe handling
Anything that goes out under a founder's name
Segment definitions that decide who gets a discount
How to run any of these tests
Pick one hypothesis, hold back a control group, and run it for at least two full send cycles before you decide. Judge it on revenue per recipient, click rate, and unsubscribe rate. If the result is a tie, take the cheaper option.
Frequently asked questions
What is send-time optimization? A feature that picks the time of day each recipient is most likely to engage, instead of sending everyone at once.
Can AI write marketing emails? It can draft them. A person still needs to edit for voice and verify every claim and offer.
What data does predictive analytics need? In Klaviyo, at least 500 customers with orders, 180 days of order history and recent orders, plus some repeat buyers.
Is AI worth it for a small email list? Often not yet. Klaviyo's send-time test needs 12,000 recipients and its predictions need 500 customers with orders. Below that, put your time into flows.
Want an honest read on which of these your program is ready for? Talk to Newly Marketing. For the bigger picture, see how AI is changing lifecycle marketing.
Sources
About the Author
Lily Newman is the Founder and Principal of Newly Marketing, a digital marketing consultancy specializing in lifecycle, email, mobile, CRM, and retention marketing. A fractional DTC e-commerce marketing leader, Lily advises high-growth startups and Fortune 500 brands in retail, CPG, health and wellness, and technology, drawing on 15+ years leading acquisition and retention for Levi's, Peet's Coffee, Mighty Leaf Tea, Four Sigmatic Foods, and more.
