AI in Email Marketing: Send Time, Subject Lines and Segments

Features are easy to switch on and hard to judge

Every email platform now has a panel of AI features: predicted send times, generated subject lines, churn risk, product recommendations, predicted lifetime value. They cost little to enable, and the dashboard reports sound encouraging. Very few businesses check whether any of it made a difference.

Here is how we would rank the common uses of AI in email marketing for a typical small or mid-sized business, and how to check each one without trusting the vendor's own attribution.

Where AI in email marketing actually helps

Note how the top of the table is about deciding who gets what and when, and the bottom is about writing. That matches our experience across AI projects: the decision problems pay more reliably than the generation ones.

What send time optimisation really does

Send time optimisation estimates when each subscriber is most likely to open, based on when they have opened before, and holds their copy of the email until then. It sounds clever and it is reasonably simple underneath.

Its limits are worth knowing. Subscribers with little history get a default time. Apple Mail Privacy Protection pre-loads images and inflates opens, which muddies the signal the feature relies on. And the effect on revenue is typically modest, because a good email read at 7pm instead of 9am is still read.

It is worth switching on for large lists where you are not sending time-sensitive content. For a flash sale ending at noon, send everyone at once.

Testing AI subject lines properly

Generated subject lines can find angles a tired marketer would not. They can also drift into the clickbait register that trains subscribers to ignore you. The only way to know is to test on the metric you care about.

1. Write your own subject line as the control
2. Generate several alternatives with clear brand voice instructions, then pick one a human would be happy to send
3. Split a random sample of the list, large enough to detect a meaningful difference, and send both
4. Judge on clicks and revenue per recipient, not opens, given privacy features distort open rates
5. Keep a log of which kinds of subject lines win over months, because single results are noisy

Also watch unsubscribe and spam complaint rates for the AI variants. A subject line that wins clicks today and raises complaints costs deliverability for every future email.

Segments and triggers: the unglamorous winners

The strongest email results usually come from sending the right message when a customer does something, not from better copy on a broadcast. AI and machine learning help here by predicting which customers are near a decision, and by choosing products to include.

• Replenishment reminders timed to each customer's typical reorder gap
• Browse and basket abandonment with products the customer actually viewed
• Lapse prevention for customers whose gap since last order is unusually long for them
• Post-purchase guidance that reduces returns and support contacts
• Win-back for previously valuable customers, excluding those who buy anyway

Our guide to customer segmentation with machine learning covers how to build the segments, and Shopify email and retention covers the flows for stores on that platform.

How to measure without fooling yourself

Platform attribution usually credits an email with any purchase within a few days of an open or click. Many of those customers would have bought anyway. The cleanest check is a holdout: keep a random slice of eligible subscribers out of a flow or feature for a period and compare revenue per subscriber between the groups.

For small lists this takes patience, and sometimes the honest answer is that a feature has no detectable effect. That is a reason to turn it off and simplify, which often improves deliverability as a side effect of sending less.

When to go beyond the platform's built-in AI

Built-in features are fine for most businesses. A custom model makes sense when your data lives outside the email platform, for example in an ERP, a subscription system or a B2B order portal, or when you want predictions such as lapse risk used consistently across email, sales calls and the website.

In those cases SpiderHunts usually builds the prediction in the data layer and syncs scores to the email tool, which keeps one version of the truth. That is typically part of broader marketing automation work, and it is overkill for a list of a few thousand subscribers.

Frequently asked questions