How to use AI to improve your email marketing performance

Most email programmes are underperforming quietly. Open rates look acceptable. Click rates are not catastrophic. Nobody raises it in the Monday meeting. But the revenue attributed to email has been flat for two years, the unsubscribe rate creeps up each quarter, and the team is spending three days a week producing campaigns that feel identical to the ones they sent six months ago.

The mistake most organisations at this stage make is treating email as a content problem. They hire another copywriter, test a new subject line formula or invest in a better template. None of it moves the needle because the real problem is a data and personalisation problem - one that no amount of creative iteration will fix.

AI changes the economics of email personalisation fundamentally. Not by writing better copy for you, but by making it commercially viable to treat each subscriber as an individual rather than a segment. This article explains where AI creates genuine, measurable lift in email performance and how to apply it without overbuilding.

Why segmentation alone is not enough

Most email teams have segmented their list. They have a VIP tier, a lapsed cohort, maybe a category-affinity split. That is a reasonable starting point, but it is not personalisation - it is grouping. Everyone in the "lapsed" segment gets the same win-back email, even though some lapsed customers stopped buying because of price sensitivity, others because of a poor delivery experience and others because a competitor launched a better product.

Static segmentation applies the same logic to every member of a group. AI-driven personalisation applies different logic to each individual based on their specific behavioural signal.

A mid-market fashion retailer with a database of 400,000 subscribers illustrates this clearly. Their existing segmentation produced twelve audience buckets. When they trained a propensity model on purchase history, browse data and email engagement signals, they discovered that send time alone had a measurable effect on conversion for over sixty percent of their list - and that optimal send time varied by more than nine hours across the database. Sending each subscriber at their individually predicted optimal time, with no change to creative, lifted attributed email revenue by seventeen percent in the first quarter.

The lesson is not that segmentation is wrong. It is that segmentation sets a ceiling. AI removes it.

Where AI creates the most reliable lift

Three areas of email marketing respond well to AI intervention. Apply them in this order, because each builds on the previous.

Send-time optimisation is the lowest-effort, highest-confidence starting point. The data exists, the models are well-established and most enterprise email service providers now expose this as a native feature or API. If your platform does not, a lightweight predictive model built on your engagement history will do the same job. Start here because it requires no changes to creative, content strategy or approval workflows.

Subject line and preview text prediction comes next. This is not about generating subject lines with a language model and hoping for the best - that is a content experiment, not an AI programme. The right approach is to train a model on your own historical send data: which subject line characteristics, emotional registers, length brackets and topic signals have predicted opens for your specific audience. A consumer electronics business, for example, might find that urgency-framed subject lines drive strong open rates on Black Friday but suppress engagement in March, while curiosity-framed lines perform consistently across the year. Your data will tell you something your competitor's data will not.

Personalised product and content recommendations is where the revenue impact compounds. Rather than surfacing the same "you might also like" block for every subscriber in a category cohort, a recommendation model weights products by the individual's recency, category affinity, price tolerance and seasonal behaviour. For an ecommerce business with a catalogue of more than a few hundred SKUs, this is not achievable manually. AI makes it routine.

A fourth lever - predictive churn and re-engagement timing - is worth noting. Sending a win-back email thirty days after last purchase because that is your standard trigger is a rough heuristic. A churn propensity model tells you which subscribers are genuinely at risk right now, and which are simply in a low-purchase season. The difference matters: misfired urgency erodes trust and accelerates unsubscribes.

Building versus buying: a decision framework

Marketing directors at this scale face a real build-versus-buy question. The answer depends on three factors.

Data volume. Personalisation models trained on thin data produce unreliable outputs. If your active email subscriber base is below 50,000 and your purchase frequency is low, off-the-shelf AI features in your existing platform will outperform a custom model until you accumulate more signal.

Catalogue complexity. If you sell fewer than 200 distinct products or content types, the marginal return from a bespoke recommendation engine is limited. Standard collaborative filtering within your ESP or CDP will be sufficient. Above that threshold, a custom model starts to justify the investment.

Strategic dependency. If email is a top-three revenue channel - which it is for most direct-to-consumer businesses - the case for owning your models rather than renting them from a platform vendor becomes commercially significant. Platform-native AI features are black boxes. You cannot audit their logic, retrain them on your commercial priorities or port the models if you change provider.

A practical starting position: use platform-native AI for send-time optimisation and basic recommendation blocks while you build the data infrastructure to support custom models. Do not attempt to build everything at once. The organisations that get the most from AI in email are the ones that deploy narrow, well-scoped models incrementally - not the ones that announce a full personalisation transformation and deliver it eighteen months late.

Getting the data infrastructure right first

This is where most AI email projects fail, and it fails early. The model is the easy part. The hard part is having clean, accessible, joined-up data before you ask a model to learn from it.

At minimum, AI-driven email personalisation requires: a unified customer identifier that connects email engagement to onsite behaviour and purchase history; event-level data (not just aggregate metrics) from your email platform; and a product or content catalogue with consistent attribute tagging.

Many mid-market businesses have all three of these data assets but they sit in three different systems with no reliable join key. A customer who clicks an email link from their mobile, browses on desktop and purchases through an app may appear as three separate records. The model will learn from noise.

Before commissioning any AI work for your email channel, run an audit of your identity resolution capability. A business spending £40,000 a year with an email agency while its customer data sits in siloed systems is spending money on creative polish over a broken foundation.

Tools like a CDP (customer data platform) address this, but a CDP implementation is a significant project in its own right. A faster path is often to work with a data consultancy to build a lightweight identity spine in your existing data warehouse - enough to support joined-up modelling without a full platform overhaul.

Measurement: knowing whether it is working

The final failure mode is measuring AI email performance against the wrong metrics. Open rates, in particular, are now heavily distorted by Apple Mail Privacy Protection and similar features across other clients. A model optimising for opens may be optimising for a phantom signal.

Measure what the email channel is actually supposed to do: revenue per send, conversion rate on clicked sessions and list health over time (deliverability, unsubscribe rate, complaint rate). If you are running AI-driven personalisation, you need a holdout group - a randomly selected portion of your list that receives the standard, non-personalised version. Without a holdout, you cannot attribute lift to the AI intervention versus seasonal effects or campaign quality.

Set this up before you go live. Retrofitting a holdout group is technically possible but commercially uncomfortable - you will face internal pressure to give every subscriber the "better" experience before you have confirmed it is actually better.

The email channel is not dying. It is bifurcating. Programmes built on genuine individual-level intelligence will widen their lead over programmes built on segmentation and manual creative. The gap between those two positions is not closing - it is growing, quietly, every quarter.

The right next step is not a platform upgrade or a creative brief. It is an honest assessment of where your data infrastructure currently sits and what it would take to support the models that drive real performance. That is a two to three week piece of work, not a twelve-month transformation.

Rodan runs paid diagnostic engagements that produce exactly that assessment - a clear view of your current data state, the specific AI interventions most likely to move your email metrics and a sequenced roadmap for delivering them. If your email programme has plateaued and you suspect the problem runs deeper than creative, book a diagnostic with Rodan.