When AI in email marketing is mentioned, most examples start with a newsletter subject line. It is easy and effective. In real-world operations, however, creating one sentence is rarely the most expensive part. Time disappears in data preparation, segment selection, content assembly, link checks, testing and evaluation.
I discussed the topic in an interview for MladýPodnikatel and worked on it in the DreamROI project. Since then, the tools have changed, but the core experience remains: it makes sense to automate a repeatable process, not responsibility for the offer and customer relationship.
Three different things hide under the term AI
Generative AI
It creates proposals for subject lines, copy, call-to-action variants, or turns one article into an email. It is fast, but can invent facts, change the meaning of an offer or produce an interchangeable tone. It is suitable as a co-author of the first draft and a critic, not as the final approver.
Rules-based automation
It sends a message after registration, purchase, inactivity or a status change in the CRM. It need not contain any model. Well-designed rules often have a greater effect than text generation because they respond at the right moment.
Prediction and personalisation
A model can rank products or articles for a specific recipient, estimate interest or recommend a send time. Services such as Amazon Personalize work with interaction history and a catalogue. But that means data, infrastructure and ongoing evaluation. Without sufficient signal, the model is only guessing confidently.
Start with one specific bottleneck
I would not start with the question “How do we put AI into email marketing?” I would start where the process regularly loses time or money:
- an editor manually selects articles for several segments every week,
- salespeople forget to send a follow-up message after someone downloads material,
- an e-shop sends everyone the same product order,
- the team creates versions in several languages and cannot keep up with checking them,
- no one can say which automation brings orders and which only sends.
Then you can choose the simplest intervention. Sometimes it is a new automation branch. Other times, a better data export. A model comes only when its benefit outweighs the cost of implementation and control.
When text generation makes sense
A generative model helps me mainly with variants and transforming finished source material. From one approved article, it can propose a short newsletter, three subject lines and copy for different stages of the relationship. The brief, however, must contain facts, audience, objective, prohibited claims and examples of the brand voice.
A practical process:
- A person selects the offer, audience and one email objective.
- They put approved source material into the brief, not a vague request to “write something salesy”.
- The model creates several distinct variants.
- A person checks facts, price, terms, links, tone and continuity with the page.
- The selected variants are tested against a predefined metric.
- The result is stored as a learning, not merely as a winning subject line without context.
I describe a more detailed workflow in the article AI and text creation in a company.
Personalisation needs data, not enthusiasm
For content recommendations, you need a reasonably described catalogue and interaction history. For predicting send time, you need enough past sends and opens. For example, Mailchimp’s send-time optimisation explicitly works with prior account data.
There is no single number of visits or contacts at which AI will automatically start working for everyone. It depends on event frequency, variety of the offer, seasonality and the goal. With a small list, precise manual segmentation may be cheaper and more reliable than your own recommendation system.
How to evaluate an experiment
Email opens alone are not enough. Technical changes in email clients have also made them a less reliable metric. Depending on the objective, track clicks, completed purchases, qualified enquiries, unsubscribes, complaints and the recipient’s long-term value.
For a larger intervention, retain a control group. If everyone receives the personalised email, you cannot tell whether a standard mailing would have achieved the same result. A/B-test one material hypothesis, not the subject line, offer, time and list all at once.
What must remain under human control
- Strategy and offer: the model does not know the commercial priority unless it is given one.
- Facts and conditions: a responsible person checks prices, dates, guarantees and legal text.
- Exceptions: a sensitive customer group or crisis situation needs its own rules.
- Brand voice: the tool can follow a pattern, but the brand must know how it wants to come across.
- Consent and data: do not casually send personal, confidential or client data to a public AI tool.
- Switching off: the automation must have an owner and an easy way to stop it.
The simplest pilot for a smaller company
Choose one regular newsletter. Measure how much time preparation takes today, how many errors recur and which commercial metric you track. Let AI prepare variants from an approved article and a review checklist. After four to six sends, compare time, quality, clicks and negative reactions.
If time has merely moved from writing to corrections, the pilot has failed. If the process became faster without a loss of accuracy and people received more relevant content, you have a reason to continue. Only then would I address more complex personalisation or integration.
When not to use AI
I would not use it as an autonomous sender for sensitive offers, where approved source material is lacking, or where one error could harm a customer or reputation. Likewise, it makes no sense to build a prediction layer on top of a broken CRM and inconsistent consent records.
First, put the basics of email marketing and automation in order; you can find practical work with one system in the ActiveCampaign guide. If you want to choose the first pilot based on the actual process and data, write to me through contact.