What does AI actually do in email marketing?

Sending an email to a list resolves into a stack of separable decisions. Six of them are where the software clusters. What to send. Who gets it. What it says. What it looks like. When it goes. What happens next.

“AI email marketing” is a label stretched across all six. It is worth pulling apart, because they are not equally automated.

Vendor help documentation is precise, quantified and occasionally blunt about the mechanical jobs. On decisions it goes quiet. That pattern holds across platforms, which makes it useful when you are evaluating any of them.

Which jobs does the automation actually reach?

Five of the six. Timing, copy, design, audience and flow logic all have documented features across these five platforms. The sixth — deciding what a campaign should be about — has none.

The last row is the finding.

Where does the automation stop inside a flow?

Brevo draws the line explicitly rather than leaving it to be inferred: the generator builds the wiring of a flow and stops at the email content and the segmentation filters, which you add by hand before activating.

Aura generates an automation from a description, and the documentation lists what it configures for you: “Wait and delay steps: duration and unit set automatically from your described timing” · “Trigger and entry settings” · “Step labels” · “A/B split percentages” · “Update contact attribute steps” · “Add to list and remove from list steps” · “Webhook steps” · “SMS steps”.

Then it names the exclusions:

“Email step content and segmentation filters are not yet generated automatically. After Aura creates your workflow, you need to add your email templates and configure any conditions or filters manually before activating.” — Brevo Help Center, Create an automation with Aura

Read the two lists against each other. Everything in the first is plumbing: how long to wait, what to call a step, how to split a test, which webhook to fire.

Both exclusions are judgment. What the email says, and which customers should receive it.

That is not a criticism. Wiring a flow by hand is tedious and worth automating. But it locates the boundary, and the boundary sits earlier than generating an automation from a description sounds like it should.

Do all vendors stop the automation at the same point?

No. The line is not standard: Klaviyo’s flow generator writes the email content and the segmentation filters that Brevo’s leaves to you. Klaviyo’s Composer writes flow message content: “Composer will first reflect the structure of the flow that you are requesting and then generate the content for each of the messages within that new flow.” It builds audiences too — “Create a segment by describing the audience you want in plain language.”

A third does not document the job at all. Omnisend’s AI documentation lists nine features across content, forms, segments and analysis, and no flow-generation feature appears among them.

So one vendor calls email content and segmentation the manual part, another generates both, and a third’s AI documentation does not list the job. That split is worth carrying into any evaluation, because the phrase AI-built automations does not tell you which half of the work arrives finished.

Does the tool that writes the content also send it?

No. Klaviyo’s Composer publishes its own boundary, and it sits at execution rather than at content. It “recommends, drafts, audits, and QAs”, and “does not send, publish, schedule, or change your account settings on its own”, “does not apply fixes from an audit for you”, and “does not replace your judgment”. The drafting is automated. Approving it is not.

What do the design tools generate?

Design AI is documented mostly as asset application rather than composition.

Omnisend’s brand tooling “applies your brand logo, fonts, and color palette automatically to email templates”, and needs a connected store to pull them. Of its brand kit, Mailchimp says “We’ll use AI and the assets in your brand kit to quickly generate stylish, on-brand graphics and layouts for your emails, social posts, automation flows, and more.” Applying that kit automatically to template previews is the part gated to a paid tier — “available for accounts with an Essentials plan or higher.”

What does a generated design leave unfinished?

The parts that carry the offer. Klaviyo’s Email AI generates an email section from a description, then states what will not arrive finished: “Certain elements (like product blocks, images, links, and buttons) are not fully configured by the generation tool”.

It needs “a paid Klaviyo account”, and its own example of a prompt is “Create a sale reminder with urgency to order before midnight, and include an image at the top and a button right below”. Other block types, including tables, are “not supported.” Usage caps at 99 generated sections a day.

So the layout arrives and its commercially important parts — the product block, the link, the button — still need configuring. The tool covers the mechanics and stops at the part carrying the offer.

What send-time thresholds can you check before buying?

A recipient floor and a plan tier. Klaviyo’s Smart Send Time will not run below an audience of “12,000 people or greater”, and the same vendor’s per-profile model is gated on plan tier rather than on volume. One category, two mechanisms from the same vendor.

It publishes its method instead of keeping it opaque — “Instead of determining your business’ send time through hidden formulas, Klaviyo uses a robust testing framework” — plus a table of how many test campaigns each list size needs.

Why does every platform call send-time AI something different?

The category shares no vocabulary: Send Time Optimization at Mailchimp, Optimize send time for individual contacts at HubSpot, Send at best time at Brevo, Smart Send Time and Personalized Send Time at Klaviyo. Underneath, each reads different data.

Mailchimp pools behaviour across its whole customer base and says so: “Mailchimp customers send a lot of email to a lot of people, which gives us data about individual engagement patterns” — gated to “the Standard plan or higher.” HubSpot reads each contact’s own last 90 days, a Beta feature requiring a “Marketing Hub Enterprise subscription”. Brevo starts from your account’s own averages and sharpens as you send.

What happens when there is not enough data to time a send?

Three of the four per-contact products publish a fallback, no two alike, and the fourth declines the job rather than guessing.

HubSpot sends contacts “without enough engagement data… at the beginning of the sending range” — a fixed time, not an optimised one. Brevo falls back twice: to your own past campaigns if you have any, and for an account that has never sent, to “the Brevo average”, drawn from “emails sent by other users”. Klaviyo also falls back twice, but stays inside your account — “send-time patterns from similar profiles in your account”, then “what’s generally best for your account.”

Mailchimp declines the job: “If there’s not enough data for Mailchimp to find an optimal time, click Cancel and choose another option.”

Which makes the labels less informative than they look. Brevo’s cold start reaches for the same cross-customer pool Mailchimp draws on — though Mailchimp still needs “enough data from your sent emails” to run.

That matters on a growing list. Every new subscriber arrives with no history, so some share of every send runs on a fallback rather than that person’s behaviour — or, on one platform, not at all.

What has to be true before predictive features work?

Predictive scoring gates on order history rather than list size. Klaviyo’s gate is four conditions, and the first asks for 500 customers who have actually placed an order.

Klaviyo publishes its gate as four conditions: at least 500 customers who have placed an order, an ecommerce integration or API sending placed-order events, 180 days of order history with orders in the last 30, and some customers with three or more orders. It is precise that the first is not a list-size number: “This does not refer to total profiles, but rather the number of people who have actually made an order with your business.”

Together those describe a business trading six months with repeat purchasing.

Is list size the number that gates predictive scoring?

No — the gate counts purchasers, not subscribers. A store with 40,000 subscribers and 300 purchasers fails on the first condition, which asks for 500 who have placed an order. List size is the obvious number to quote when asking whether a tool suits you. This gate does not read it.

What do the predictive scores actually measure?

Klaviyo states its predictions “work best when averaged over many customers and are not expected to be exact for any single individual”, illustrating with a predicted order count of 1.43 — a figure that cannot describe a person.

Omnisend labels customers “at risk” using a published calendar for stores with “fewer than 100 returning customers or less than 5–7% returning customers”: “Last purchase 90–180 days ago.” That is a date subtraction, and a reasonable one. It is a rule rather than a model, which is worth establishing before assuming a score was trained on anything.

What do the segment and copy tools hand back to you?

Different things, and only one of the two publishes a caveat about it. Klaviyo’s segment builder “accepts natural language inputs… and converts them into a Klaviyo segment” and attaches the warning itself: “you are still responsible for the final segment definition.”

Its own examples show the scope: “Engaged profiles last 30 days”. These map to conditions that already existed. The tool writes the filter, not the strategy — and a tool that turns your sentence into conditions can turn it into the wrong ones, with a result that looks equally confident.

Neither of these two jobs comes with a published volume threshold, and the gates fall by job rather than by vendor. Omnisend puts its copy tools on “all Omnisend plans (including Free)” and its segment builder behind a connected store; Klaviyo’s needs a paid account.

Does the copy tool read your data?

The answers differ, and not neatly by vendor. Two subject-line tools are documented as taking a description. But Omnisend’s separate body-copy tool “generates texts in your brand voice based on prompts and past campaign content.”

Two generation features, one help page, different inputs.

Does AI-written copy perform better?

Mostly unmeasured, and where it has been measured the answer is mixed rather than flattering. None of the five platforms’ help documentation carries a controlled test.

One exception is worth reading, because the result does not flatter the vendor who published it. GetResponse compared more than 16,000 emails sent through its platform in 2024:

Read whole, it is mixed. AI-generated emails earned more clicks per delivered email and more per open. Fewer opened them. Marginally more unsubscribed.

GetResponse states the caveat itself: treat the data “only as a sample and not as a direct guide to making decisions in your strategy.” It is also observational — users chose which emails to generate — so not a controlled test.

Which job does none of the documentation automate?

Set the six side by side and one has no entry: deciding what a campaign should be about. No feature in any of this documentation makes that call.

Which offer, aimed at which moment in a customer’s relationship with you, judged against what the last one did — that question sits upstream of everything these tools generate. Segment builders translate a description you have already formed. Flow generators, whether or not they write the messages, start from a goal you supplied. Klaviyo’s subject-line assistant asks you to “describe the context”. Composer starts from you “describing your goal in plain language.” In the help documentation, every one takes a decision as its input.

Do the vendors hide where the automation stops?

No. Every exclusion on this page comes from the vendors’ own help documentation, and the requirement that you arrive with a prompt runs through all of it. What it does is invert the pitch, and the pitch is worth quoting rather than characterising. Omnisend’s AI page: “Omnisend AI writes your copy, suggests subject lines, builds segments, recommends products, and predicts which customers are about to churn. You review, adjust, and send.” Klaviyo’s: “K:AI handles who to reach, what to say, and when to send it—personalizing every message with the right products.”

The documentation says AI formats, times and wires them, then waits for you to say what they are for.

Does one vendor’s product page match its own documentation?

No. Klaviyo publishes a pitch and a contradiction of it. Its marketing page promises “AI that creates, resolves, and optimizes— autonomously”, says Composer will “uncover opportunities”, and describes agents that “don’t wait to be told what to do” and work “without prompting.” Its help centre documents that same product starting from a goal you supply.

Both pages are current and both are the vendor’s own — which is the argument for reading the help centre before the demo rather than after.

Which parts of the work have actually been measured?

Two surveys, six years apart, and the answer is the same both times: everything except deciding what a campaign should be about.

One survey put hours on it, in 2017. A sample of 3,500 marketers ranked six email production tasks by average time: graphics and design 4.1, coding and development 3.8, copywriting 3.0, data pulls 2.4, testing 2.3, analytics 2.1. Copywriting — a job this generation of tools targets directly — came third by that measure, behind design and code. A reader poll on the same page, 43 responses, put copywriting first instead; two instruments, two answers, and the smaller one is nowhere near large enough to overturn the larger.

What the analyst publishing the figures noticed is the part that lasted. They saw what the list omitted, and asked in print:

“But do you notice a task that is missing? Where is the planning task? All these tasks are important, but how do you know what to write or design for the email?”

Those hours are from 2017 and should not be read as current. The durable finding is structural: a survey built to time email production timed six tasks and never counted the deciding.

What did the newer research measure?

The production cycle rather than the tasks — and it left out the same job the 2017 survey did. Litmus’s 2023 State of Email Workflows report, surveying “over 440 email marketers worldwide”, timed the cycle rather than the tasks: “The email production cycle is one week for 21% of email marketing teams. But for 62% of email marketing teams, it takes two weeks or more to build an email.” Weeks per email, not hours per task, so it does not update the 2017 figures.

Its blocker question bears more directly. Ranked, the obstacles are “Building (41%), Designing 40%, and Testing (39%)”, then “Collecting feedback (35%), Content creation (34%), and Getting buy in from all stakeholders (32%).”

Six blockers, and every one is downstream of a decision already taken. Choosing what a campaign is for appears in neither list, just as planning did not appear in 2017.

What has never been measured?

Planning. Neither the 2017 task survey nor the 2023 Litmus report counted the time it takes to decide what a campaign is for, so that figure has no published source.

What the vendors sell against is the other half — the tasks both surveys did count. Omnisend promises “Write and personalize faster”; Klaviyo offers to “Go from prompt to campaign in minutes” because “What used to take hours can now take minutes.” Writing, designing and building are exactly where those hours were measured. The pitch lands on the half that has numbers.

Which makes it a question about your own programme. The surveys say what the industry counts, not what your week goes on.

How do you read any AI claim in this category?

Four questions, all answerable from the documentation before a trial: which of the six jobs it touches, what the vendor says it will not do, what has to be true before it runs, and whether it is a model or a rule.

1. Which of the six jobs does it touch? Timing, copy, design, audience, flow logic, or the decision. The first five have documented features. Claims about the sixth deserve scrutiny.
2. What does the vendor say it will not do? An exclusion list carries more information than a feature list, and a vendor publishing one is easier to evaluate than one that does not.
3. What has to be true before it runs? A published threshold — 12,000 recipients, 500 purchasers, a connected store, an Enterprise tier — tells you when a feature works. “Enough data” does not.
4. Is it a model or a rule? A trained probability and a 90-to-180-day date subtraction can carry the same label.

What does this mean for a smaller programme?

Several of these features are gated on volume or plan tier, so a small store will find parts of the category unavailable, and an AI feature list is a weak tiebreaker at that size.

The more useful reading is where your hours go, not which platform to buy. These tools have gone furthest into the jobs that were already mechanical, and stop — by their own documentation — where a judgement starts. If most of a week goes into deciding what the next campaign should be, that is the half the documentation stops short of.