Personalization in email used to mean putting someone's name in the subject line. That stopped being differentiating a long time ago, and the measured effect now is somewhere between small and zero — recipients have seen it ten thousand times and it reads as automation rather than attention.
What AI adds is different in kind. It changes what you offer rather than how you address someone, and it does so by predicting things you do not directly know.
The four things AI can actually predict
What someone is likely to buy next. A recommendation model looks at what a person bought, what similar people bought afterwards, and what is available, and produces a ranked list. This is the oldest and most reliable application, and it works because the underlying pattern is real: purchase sequences are not random.
When someone is likely to open. Per-subscriber send time prediction from their own open history. Small, consistent lift on every campaign.
Who is about to leave. Churn prediction from declining engagement. Useful because it moves the intervention earlier — a subscriber flagged as at risk while they are still opening is far more recoverable than one already dormant.
Which content someone will engage with. Given a set of articles or products, predicting which each subscriber will click. This is what powers the recommended-reading block in a newsletter, and it is a genuine improvement over showing everyone the same three items.
Each of these is a prediction about behaviour, not a fact about a person. That distinction matters for the section below.
The line between helpful and unsettling
There is a real threshold past which personalization stops feeling like service. The reliable way to locate it:
Personalise the offer, not the voice. Showing someone products related to what they bought is helpful. Writing a sentence that narrates their browsing back to them — "we noticed you spent a while looking at the blue jacket" — is not, even though it uses the same data. The first is a recommendation; the second is a report on surveillance.
Use data the person knowingly gave you. They told you their size, their location, what they bought. Using that is expected. Using inferred data — predicted income bracket, inferred life stage, guessed intent — feels different because they never handed it over, and being wrong about it is worse than not using it.
Do not reference the mechanism. "Because you viewed X" is honest and slightly chilling. Just showing X-adjacent products achieves the same thing without the reminder.
Never be specific about something you inferred. A prediction stated as a fact is the classic failure — the retailer that congratulates someone on a life event they have not had. The cost of being wrong is much higher than the benefit of being right.
What the lift actually looks like
Recommendation blocks outperform static promotional blocks fairly reliably, because relevance is doing real work. Predicted send time adds a small consistent gain.
Beyond those two, the returns get thin quickly, and there is a cost people underestimate: every personalised element is a branch that has to render correctly for everyone, including the people with no data. An empty recommendation block, a merge tag with no fallback, or a "for you" section showing default items all read as broken, and broken is worse than generic. See dynamic content and email merge tags for the failure modes.
Where the actual leverage is
The uncomfortable truth is that most of the benefit attributed to personalization comes from segmentation rather than from per-person customisation.
Sending four genuinely different emails to four groups usually beats sending one email with twelve personalised elements, because the four emails can differ in argument and structure while the personalised one can only differ in details. The segment version also fails gracefully — a segment with missing data still gets a coherent email.
Personalization is the layer you add on top of good segmentation. It is not a substitute for it, and adding it to an unsegmented programme is optimising the wrong thing.
A practical order
- Segment by lifecycle stage. Four different emails.
- Turn on send time prediction. Free, automatic, small consistent gain.
- Add a recommendation block to the emails where products or content make sense.
- Add churn prediction to trigger re-engagement earlier.
- Stop. Everything past this is diminishing returns and increasing fragility.
All four are available on every IGSendMail plan including the free one, and if you would rather work through an assistant than a dashboard, the MCP server exposes the same data.
Frequently asked questions
Does first-name personalization still work?
Barely. The measured effect is small and can be negative when the data is dirty, since "Hi ," or "Hi JANE SMITH," is worse than a neutral greeting. Use it only if your field coverage is high and clean.
What is the most effective AI personalization for email?
Product or content recommendations based on past behaviour, followed by per-subscriber send time prediction. Both change what someone receives rather than how they are addressed, which is where the actual lift is.
When does personalization become creepy?
When it narrates behaviour back to the recipient, or when it states an inference as a fact. Showing relevant products is fine; describing how you know they are relevant is not.
Is personalization better than segmentation?
No — segmentation is the larger lever, because different segments can receive genuinely different arguments while personalization only varies details. Personalization is a layer on top of good segmentation, not a replacement for it.
