In 2026, AI does four things well in email marketing: it writes a different message for each person, picks the angle and the timing, follows a brand voice and vocabulary, and learns from results. It does three things badly: it invents facts if you let it, it can't set strategy, and it replaces neither consent nor deliverability. Much of the rest is hype. Below you'll find the four things that work, each with an example and a way to verify it, the difference between real and fake personalization, how it all connects to Shopify and Klaviyo, and the risks to check. If you are weighing AI for email marketing, this is the checklist to start from.
The four things AI does well, and how to verify them
Writing a different text for each person. In a supplement store with a fixed template, every abandoned cart gets the same email. With AI, someone who has already bought three times gets a text that starts from their history; a first-time visitor gets a text that explains shipping and returns.
Choosing the angle and the timing. The same customer might respond to a review, a product detail, the shipping service or genuine scarcity. AI can assign different angles to different customers and adapt the wording to the time the email will arrive: the flow decides the window, the copy adjusts to it.
Following a tone and a vocabulary. A clothing brand that uses a casual voice and never says "unmissable" can enforce that on every text, as long as the tool actually checks it.
Learning from results. Every email produces an outcome: an order or no order. A system that records which angle each customer received can shift sends toward what converts.
In AchillesAI's flows, for example, the angles are called communication angles and are tested per customer and per store, while the learning also draws on the Network: connected stores contribute aggregated, anonymous data, and no store's data is visible to another.
How to verify it works. Run an A/B test per person, not per campaign. Randomly split the customers who enter the flow: half get the generated texts, half get your usual template. Compare orders per recipient across the whole flow, not opens, which are inflated by the privacy protection in mail clients. Comparing October with September is not a test: season, traffic and promotions all change.
Real personalization and fake personalization
Fake personalization is the name in the {first_name} field: the text is identical for everyone, and only one word changes.
Subject: Hi Emily, did you forget something?
Hi Emily,
you left some items in your cart. Complete your order now before it's too late!
[Back to your cart]
Real personalization works on the customer: history, product, moment and language. Here is the same abandoned ring at a jewelry store, for two different people.
Subject: The silver ring you were looking at
Good morning Sarah,
the band ring you picked last night is still in your cart. It's your first order with us, so two things worth knowing: shipping is insured, and you have thirty days to swap sizes at no cost.
[Pick up your order]Subject: A ring to go with your earrings
Good evening Daniel,
the ring in your cart is from the same line as the earrings you bought in March: same silver, same finish. If it's a gift, we can ship it in a sealed box with a card.
[Complete your order]
Neither contains an invented fact: history, product line and return terms all come from the store. The AI chooses what to say to whom.
Tone of voice and banned words: why you need a rewrite rule
Writing "don't use the word unmissable" in the prompt is not enough. A model follows instructions almost every time, and "almost" across thousands of emails a month means off-brand messages. Vocabulary should be treated as a control, not as advice.
The method that works has three steps: clear instructions on tone and words to avoid; an automatic check on the generated text before sending; and, if the check finds a banned word, a mandatory rewrite of that sentence followed by a new check. In AchillesAI the store sets banned words and a tone, and if a text contains them it is rewritten before it goes out. If a vendor answers "the model knows", there is no control.
What AI should not decide
AI writes and chooses among options you have approved. It doesn't decide strategy: which products to push, how much to discount, how many emails to send in a month, which customers to stop writing to. Those depend on margins, inventory and positioning, which no model knows better than you.
The same goes for discounts, which need rules that are yours and strict: a code valid only on a first purchase shouldn't be offered to someone who has already bought, and an incentive shouldn't go in the first email (we explain why in our article on discounts in abandoned cart emails).
Will AI replace email marketing? No. It replaces mass writing, meaning the same text sent to everyone. Strategy, consent and domain reputation remain human work. Salesforce's 2026 State of Marketing (4,450 professionals in 26 countries) points the same way: personalizing content at scale is among the top uses of AI, but almost half of respondents haven't yet figured out how to adapt their strategy, and the obstacles are mostly in the data.
How it connects to Shopify and Klaviyo, and which risks to contain
The setup is almost always the same: events → text generation → sending from the platform. Shopify records the event (product viewed, cart, checkout) and passes it to Klaviyo. The Klaviyo flow calls the AI tool with that customer's data, and the tool writes the text. The email goes out from Klaviyo, with its own send times, filters and unsubscribes. AchillesAI, for instance, receives events from Klaviyo flows via webhook and leaves the sending in Klaviyo. That way consent and deliverability stay where they are already managed, and if you switch the AI off, the flow keeps running with a fixed text.
There are three risks. Hallucinated prices and availability: the model should never write a price or a stock level from memory; this data comes from the catalog and goes in as fields, not as generated text. Off-brand tone: contain it with the vocabulary check described above and a spot review of sent emails. Dependence on a single vendor: flows, lists and consent should live in your own account on the sending platform, so changing tools doesn't mean rebuilding everything.
| Task | Does AI do it well today? | What to check |
|---|---|---|
| A different text for each customer | Yes | That it uses history, product and language, not just the name |
| Choosing the persuasive angle | Yes, if it's tested | That scarcity is real and results are measured per customer |
| Copy that fits the arrival time | Yes | No "good evening" in the morning, no "today" three days later |
| Following tone and banned words | Yes, with a rewrite rule | That an automatic check runs before sending |
| Learning from results | Yes, with enough volume | A test with a control group, not a month-to-month comparison |
| Prices, availability, deadlines | No | That they come from the catalog as data, never from generated text |
| Strategy and discount policy | No | Rules you write, applied without exceptions |
| Consent and deliverability | No | Authenticated domain, clean lists, sending from your own platform |
Where to start: six questions for any vendor
Before switching on an AI tool for your emails, ask:
- Is the text written for the individual customer (history, product, language, moment) or for a segment?
- Do prices, availability and discount codes come from store data, or does the model write them?
- What happens if the text contains a banned word: is it rewritten before sending?
- How do I measure the result: is there a per-customer test with a control group?
- What data does it learn from, and can other stores see my store's data?
- If I stop using it, do my flows, lists and consent stay in my account?
Then start with a single high-volume flow, usually abandoned cart recovery, and run the per-customer test for a few weeks before extending it. For the timing of the sequence, read when to send abandoned cart emails; to see how copy written per customer changes, see the abandoned cart email examples.
