Inboxes are more crowded than ever. AI-powered personalization goes far deeper, analyzing purchase history, browsing behavior, engagement patterns, and timing to decide what each individual should receive rather than sending everyone the same message with their name at the top.
That is the promise. The reality is more mixed, and the difference between a program that lifts revenue and one that quietly annoys your list comes down to a few practical decisions. This article covers what actually works, what crosses a line, and what EU law requires.
Why first-name tokens stopped working
“Hi {{first_name}}” was a differentiator fifteen years ago. Today it signals bulk email as reliably as no personalization at all, and it fails visibly when your data is imperfect: “Hi ,” or “Hi JAN,” or “Hi test”.
Genuine personalization changes the substance of the message rather than a token in the greeting. Different products, different offer, different timing, different length, different language.
The four layers worth building
1. Content selection
The simplest and highest-return application. Rather than one product block for everyone, the message shows items related to what this person browsed or bought. For a service business, the equivalent is leading with the service that matches what they read on your site.
2. Send-time optimization
Everyone has a rough window in which they read email. Systems that learn each subscriber’s pattern and deliver accordingly typically produce a measurable lift for very little effort, because the sender does nothing beyond enabling it.
3. Frequency adjustment
The most underused layer. Some subscribers want weekly contact; others unsubscribe at the third message in a month. Letting engagement determine frequency, rather than pushing the same cadence at everyone, reduces list attrition significantly. Fewer emails to the right people is usually more profitable than more emails to everyone.
4. Lifecycle stage
A first-time buyer, a repeat customer, and someone who has gone quiet need different messages. This is closer to good segmentation than to AI, but AI is genuinely useful at spotting the transition between stages, particularly the early signals that a customer is disengaging.
Predictive signals that earn their place
Two are worth the setup effort for most businesses:
Churn risk. Models that identify customers whose behavior resembles those who previously stopped buying let you intervene while there is still a relationship. Reaching someone before they leave is far cheaper than winning them back afterwards.
Replenishment timing. For consumable products, predicting when someone is likely to run out and timing the reminder accordingly converts well, because it is genuinely useful rather than merely promotional.
A third, next-best-product prediction, is popular but frequently disappointing at small data volumes. Below a few thousand customers, simple rules based on category and recency usually perform as well as a model.
Where AI writing helps, and where it does not
Language models are good at producing subject line variants to test, adapting one message for different segments, drafting from a clear brief, and translating between Dutch and English while keeping tone.
They are poor at knowing your customers, judging what is appropriate, and maintaining a distinctive voice across a campaign. Fully automated email generation tends toward the generic, and generic is precisely what personalization was supposed to solve.
The practical division: use AI for production and variants, keep human judgment for what to say and to whom.
The line between relevant and unsettling
Personalization fails when it reveals more observation than the customer expected. The useful test is whether the customer would be comfortable if you explained how you knew.
Comfortable: “You bought this six weeks ago and it typically lasts two months.” Obvious, expected, useful.
Uncomfortable: “You looked at this three times on Tuesday evening.” Accurate, and it announces that you are watching closely.
The same information can be used well or badly. Recommend based on browsing without narrating the browsing. Time the message without announcing the timing.
There is also a category to avoid entirely: inferences about health, finances, relationships, or anything a person might not want inferred. The regulatory risk is real, and the reputational damage when it goes wrong is worse than any lift it produces.
What EU law requires
Personalization at this level is automated processing of personal data, and several GDPR obligations follow.
Transparency. Your privacy statement must describe the profiling in terms a normal person understands, not “we may use data to improve your experience”.
Lawful basis. Marketing email needs consent in the Netherlands. The profiling behind it usually rests on legitimate interest, which requires a documented balancing assessment weighing your commercial interest against the intrusion.
The right to object. Subscribers can object to profiling specifically, not only to marketing generally. Your system must be able to keep someone on the list while switching off personalization for them.
Access and portability. If someone asks what you hold, that includes the behavioural profile, not just their name and address.
Retention. Behavioural data needs a defined lifetime like anything else. Browsing history from four years ago is not improving your recommendations.
Measuring whether it works
Compare against a genuine control group rather than against last quarter. Hold back a random portion of the list from personalization and compare revenue per subscriber. Without a control, seasonality and list growth will convince you of improvements that did not happen.
Watch unsubscribe and complaint rates alongside revenue. A campaign that lifts short-term revenue while accelerating list attrition is borrowing from next year.
Frequently asked questions
How much data do we need before this is worth doing?
Content selection and lifecycle segmentation work at almost any size. Predictive modelling needs meaningful volume, realistically a few thousand customers with repeat purchase history, before it outperforms simple rules.
Does our email platform already do this?
Most mainstream platforms include send-time optimization and product recommendations. The features are frequently present and unused because nobody configured them. Check what you are already paying for before buying anything new.
Is it worth it for a service business rather than a shop?
Yes, though it looks different. Rather than product recommendations, you are matching content to the service someone is interested in and timing follow-up around their engagement. The lifecycle logic matters more than the recommendation engine.
Can we use AI to write in Dutch?
For drafting, yes, but always have a native speaker review before sending. Dutch marketing copy that reads as translated undermines trust in a way that is difficult to recover, and the errors are often tonal rather than grammatical.
A sensible order of work
Get consent and data collection right, segment by lifecycle stage, switch on send-time optimization, add content selection based on browsing and purchase history, and only then consider predictive models. Most of the available return sits in the first four steps, and they require configuration rather than data science.
We build email programs as part of our digital marketing service, and design the AI systems behind them with the privacy constraints handled from the outset. Talk to us if you want a view on what your current data would actually support.
