Most writing about AI in email marketing is about wording: draft the subject line, shorten the copy, produce three variants. That is the smallest part of the work and the part that was never the bottleneck.
For this guide we looked at what German retailers actually sent in October 2026, four full weeks from our newsletter panel. The result is awkwardly clear. The mechanic with the second-biggest payoff is already in most inboxes, and often in the wrong emails. The mechanic that would cost nothing is missing almost everywhere.
So this guide separates three things that get conflated in every conversation about AI and email, shows from market observation which of them is actually running, and gives the order to start in. The underlying rule was checked on 6 November 2026 against Anthropic's published guidance on building effective agents.
Three layers that have nothing to do with each other
"We use AI in our newsletter" can mean three entirely different things. They differ in effort, in payoff, and in whether they need data or just an account.
| Layer | What gets automated | What it needs | What it returns |
|---|---|---|---|
| Copy | Subject, preview text, paragraphs, variants | An account and a stored brand voice | Speed, no revenue |
| Selection | Which items go in which email, in what order | Catalogue, availability, purchase history | Noticeably more per email |
| Trigger | Who gets what, and when | Events in the account and the basket | The largest effect |
The layers do not build on one another, they just cost different amounts. The copy layer is set up in two days and stays a writing aid. The selection layer needs data about the catalogue and about what someone has already bought. The trigger layer needs an event from the system rather than a slot in the content calendar.
Nearly every team starts with copy, because it has no prerequisites. And many stay there, because the next step looks like a project rather than a tool.
What the market actually does
The layers are easy to tell apart in panel data, because each leaves a trace in the email itself. What we see is what was sent, not how it performed: nobody has other shops' opens, clicks or revenue. And we do not name the shops, because the point is the pattern, not the sender.
The selection layer has arrived. Roughly one email in three carries a recommendation block, with five items on average, and close to three in four shops use one at least once a month. That is the layer with the second-largest effect, and it has become standard practice.
But it sits in the wrong emails. More than half of pure content emails show recommendations, against only about one in five promotional emails. The reverse would be correct. In a promotional email the intent to buy is already high, the discount is doing its work, and the extra item costs no further margin. Running recommendations only in the quiet emails spends them where they matter least.
Direct address is almost entirely absent. Only about one email in fifty addresses the recipient in the subject line at all, and that comes from roughly one shop in seven. A name in the subject has been a merge field in every sending platform for twenty years. It costs nothing and needs no model. Its absence is not a technology problem.
Dynamic content stays decorative. About one email in ten references the season or the weather, and nearly always editorially: "cold days", "autumn weather". Only a handful of shops work with real values, meaning temperature and chance of rain at send time. That is the oldest available form of dynamic content.
Send frequency leaves room for quality. The median shop sent seven emails in four weeks, fewer than two a week. At that volume, drafting speed is not the constraint. Which points at the real one: not production, but the decision about what belongs in which email.
The four jobs worth doing
Not every job is a good first one. The useful ones recur often, produce a result you can check in two minutes, and fail visibly.
Subject and preview text in variants. The classic entry point, and it qualifies because you judge the output instantly. It becomes genuinely useful with a stored brand voice and a list of banned phrases instead of long prompts, and the median subject line in the market runs just under forty characters, so there is little room to go wrong. How to store that context instead of retyping it is in the guide to Claude for marketing.
Item selection per segment. This is the half-used lever. The job is not "suggest some products". It is "take the category of the last order, exclude what was bought, exclude what is out of stock, and sort by margin". That is a rule rather than a model problem, which is exactly why it works. The mechanics behind it are in the guide to cross-selling in e-commerce.
Triggers instead of calendar slots. Three flows are missing or dormant in most shops: after a first purchase, after a gap longer than the usual repurchase interval, and on return to a category. The repurchase interval per customer is arithmetic, not a prediction, and needs no AI at all. The technical build is covered in the guide to marketing automation for e-commerce, and the step beyond it in the guide to hyper-personalization in CRM.
The weekly review. Every Monday, the same question: what went out last week, here and across the market, which mechanics are clustering, what do we plan from that. This is the best first piece of automation, because it recurs, because the output is a text, and because a mistake triggers nothing. It writes nothing into the shop.
The prerequisites are not in the model
Three of the four jobs above fail on access rather than on the model. Four things have to be in place before anything beyond a writing aid pays off.
The catalogue has to be queryable, including availability, not exported last week. Purchase history has to be readable per customer, at least the category and date of the last order. Segments have to live in the sending platform rather than in a spreadsheet. And one named person signs off while anyone is still checking.
Miss any of those and you get correct sentences about a generic catalogue. That is why so many projects stop at the copy layer: none of it is required there. The rules for those connections, which data may enter which tool and where reading ends and writing begins, are in the AI policy template for marketing teams.
An order for the next four weeks
| Week | What you set up | How you know it landed |
|---|---|---|
| 1 | Check merge fields and segments in the sending platform, store the brand voice | The subject can carry a name without anyone typing it |
| 2 | Move the recommendation block into promotional emails, not just content emails | Every promotional email carries a second reason to buy |
| 3 | Switch on one trigger that fires on the repurchase interval | Emails go out without appearing in the plan |
| 4 | Automate the Monday review across the market and your own sends | The review is ready before the meeting starts |
The order is not arbitrary. Weeks one and two cost hours and work immediately, week three costs days and keeps working, week four returns time every week. Starting at week four automates a decision nobody has made yet.
There is a good reason for that sequence from the people who build these systems. Anthropic's guidance on effective agents recommends finding the simplest solution possible and adding complexity only when it demonstrably improves outcomes. Predefined workflows, it notes, offer predictability and consistency for well-defined tasks, while agents earn their place where the number of steps cannot be predicted in advance. For email that means selection and triggers are rules, not an agent problem. Where the line falls is covered in the guide to AI marketing agents.
Where the limits are
Two decisions do not belong in the automation on the first pass: the size of the discount, and which segments receive a promotion at all. Both move margin, neither is checkable in two minutes, and a mistake shows up in the monthly accounts rather than in the email.
Labelling is the other limit. As soon as generated images go into the template, the same rules apply as on any other ad surface, and the place for the label is the image template, not the approval round. What applies in your own channel is in the guide to AI disclosure in ads.
Then measurement. Turnaround time per email, rework rate and weekly usage hold up; estimated time saved does not. How to judge individual campaigns separately from that is in the guide to campaign performance.
Where Reyo sits
Reyo supplies the outside view: what the market is sending. Reyo MCP is the interface for your own AI client, for example Claude, ChatGPT or Claude Code. Your client reads your competitors' campaigns and emails, creates entries in the Campaign Planner and writes comments, always only within your own account. Access is included from the Professional plan, and the MCP guide shows the steps per client.
That solves the fourth job above, the Monday review, with one connection: the client queries the market week, groups by mechanic, compares against the previous period and files the draft in the planner. One read connection, one write connection inside your own account, none pointing outward. Where the observations come from is shown by competitor monitoring, and which mechanics are common in the market by the collection of newsletter examples.
Five mistakes that are easy to avoid
- Starting with copy and staying there. Faster sentences are not a result. Selection and triggers are.
- Recommendations only in the quiet emails. The market does exactly this. In a promotional email the intent is higher and the extra item is free.
- Treating direct address as done. It is a merge field, not a project, and it is missing almost everywhere.
- Using dynamic content decoratively. "Cold days" in the copy is not dynamic. A value at send time is.
- Automating before it is reliable. A flow that is not clean by hand will only send the mistake faster.
Conclusion
AI in email marketing pays off least where it is most visible. The copy layer is set up in two days and stays a writing aid. The lever is in the selection, the biggest one is in the trigger, and both are mainly rules applied to data you already hold.
The market shows the pattern plainly: recommendation blocks have become standard but sit in the wrong emails, and the cheapest personalization of all is missing at most shops. If you do one thing this week, check whether your next promotional email carries a recommendation block and whether your subject line knows the name. Those are two checkboxes, not a transformation programme.
As of 6 November 2026. Market observations from four full weeks (5 October to 1 November 2026) in Reyo's German newsletter panel, aggregated at email level. The rules on simple solutions and on the line between workflows and agents come from Anthropic's published guidance on building effective agents, checked on 6 November 2026.