AI Email Marketing: How to Use AI Without Making Your Emails Sound Robotic

Most guides about AI in email marketing are written by people who have never had to send a campaign that flopped. They describe what AI is, list the features, and gesture at "personalisation at scale". That is not useful.

What is useful is knowing which specific parts of a campaign AI does well, which parts it quietly ruins, and how to tell the difference before you hit send.

This guide is about that second thing. It is written from the position that AI is genuinely transformative for email production, and that the failure mode is not bad grammar. It is sameness.

What actually goes wrong when AI writes your email

The common fear is that AI output sounds robotic. That is the wrong fear, and chasing it leads people to add exclamation marks and em dashes. The real failure is subtler: when the same model writes a hundred emails from the same kind of prompt, the outputs converge.

They converge because the model is doing exactly what it was asked. Given a brief with no distinguishing detail, the most probable next token is the most generic next token. Ask for "a friendly re-engagement email for lapsed users" and you will get the platonic re-engagement email: warm opening, "we miss you", a soft offer, an upbeat sign-off.

It is competent. It is also the same email your competitor generated, because they used a similar brief.

The damage compounds through the funnel:

• Subject lines cluster. "We miss you" and "Still thinking about us?" are what the model produces when it has nothing specific to work with. Your deliverability and your open rate both suffer when a whole batch of sends shares a subject-line shape.
• Supposed personalisation is not personal. Inserting a first name and a city into otherwise identical copy is not personalisation. Recipients read the difference between "Hi Sam" and a sentence that could only have been written to Sam.
• Weak claims get strengthened. Asked to make copy "more compelling", models add superlatives. "Fast delivery" becomes "lightning-fast delivery". That is a claim you now have to be able to substantiate.
• The voice drifts. Run five teammates' prompts and you get five different registers, unless you constrain it.

None of these are model failures. They are briefing failures. The quality of AI email output is almost entirely determined by the specificity you put in, not by which model you picked.

A working split: what to automate and what to keep

The most useful mental model is not "how much AI" but "which decisions". Sort the work into three buckets.

Automate fully

These are tasks where the model has enough context to be correct, and where variety is not required:

• First-draft subject line variants. Generate twenty, then pick. This is a breadth task and the model is good at breadth.
• Restructuring messy copy. You wrote 700 words as one block. Ask for it as a scannable structure with headings and short paragraphs. The model is not inventing content, just organising yours.
• Preheader text. Mechanical, needs to complement the subject, low stakes.
• Alt text for images. Accessibility win, tedious by hand.
• Translating existing copy. With a human check if the market matters.
• Summarising a long article into a teaser. The source material does the work.

Automate with a real brief

Here the model can produce something usable, but only if you supply specifics it could not invent:

• Body copy for a defined segment. Only if you tell it what that segment did, what they bought, and what changes for them.
• Personalisation blocks. Only with real merge data and real logic.
• Sequence steps. The model can structure a welcome series; only you know what belongs in step four.
• Replies to support-style emails. With your actual policy attached.

Keep human

Do not delegate these, and be suspicious of any tool that offers to:

• Decide what to say. The strategy, the offer, the reason you are emailing today. A model cannot know that your pricing changed or that a customer segment is unhappy.
• Make factual product claims. This is the one to guard hardest. A model asked to "sound more persuasive" will invent a capability. Every claim in every send must trace to something true.
• Handle a complaint or a churn conversation.
• Sign off on a legal or compliance statement.

If you only remember one line from this article, make it this: use AI to widen the option space, and humans to narrow it. Generate broadly, select carefully. The mistake is letting the model make the selection too.

How to brief AI so the output is specific enough to use

A generic brief produces generic output. The fix is not a magic prompt, it is supplying the four things the model cannot invent:

1. Who is receiving this, and what is true about them. Not "customers" but "trial users on day nine of fourteen who have imported contacts but never sent a campaign".
2. What happened that makes today the day to send. A release, an expiry, a milestone, a seasonal moment.
3. What you want them to do. One specific action.
4. What you are not allowed to say. Constraints are the single most under-used part of a brief. Naming the banned phrases is more effective than asking for a desired tone.

A generic brief versus a usable one

Generic:

Write a re-engagement email for inactive users.

Usable:

Write a re-engagement email for users who signed up 60-90 days ago,
imported contacts, sent at least one campaign, and have not logged in
for 30 days.
They stopped for one of three reasons: the campaign underperformed,
they got busy, or they only needed a one-off send.
Goal: get them to log in and view a campaign report showing deliverability.
Do not use "we miss you", "still there?", or any guilt framing.
Do not claim results we cannot show. Max 150 words. Plain, direct tone.

The second brief is longer, but it is the difference between an email you can send and an email you rewrite from scratch.

Constraints beat adjectives

People spend most of their prompt-writing effort on tone words: friendly, professional, warm, bold. Tone words are vague and the model interprets them loosely. Constraints are concrete and they work.

Useful constraints:

• Length. "Under 120 words." Models pad by default.
• Banned phrases. "Do not use: game-changer, seamless, unlock, elevate, revolutionary, exciting news."
• Structural requirements. "One idea per paragraph. No paragraph over three sentences."
• Evidence rules. "Only reference features from the attached list. Do not infer additional capabilities."
• Reading level. "Write for someone reading on a phone in ten seconds."

The banned-phrase list is worth maintaining as a standing document. Every time your team sees an AI-generated phrase that makes them wince, add it. Within a month the output stops sounding like the output.

Writing subject lines with AI, properly

Subject lines are the best AI use case in email, because you want volume and the cost of a bad one is low. But the standard approach wastes the model.

Asking for "ten good subject lines" gets you ten variations on one idea. Asking for diversity explicitly gets you range:

Generate 20 subject lines for this email. Deliberately use different
angles across them:
- 4 that state the benefit directly
- 4 that pose a question the reader has
- 3 that lead with a specific number
- 3 that reference the reader's situation
- 3 that are short and blunt (under 5 words)
- 3 that are playful

No line may reuse a noun phrase from another.
Flag any line over 45 characters.

Twenty lines with enforced variety beats fifty that are the same sentence rearranged. You are using the model to explore a space, not to converge on an average.

One caution: your subject line and preheader are read together in most inboxes. Generate them as a pair, or you will get a subject line and a preheader that repeat each other.

Personalisation that is not just a first name

There is a meaningful distinction between three levels, and most teams stop at the first:

Level two is where AI becomes genuinely useful, because the work is in generating variants, not inventing strategy. If you have a segment that "imported contacts but never sent", you know exactly what to say. The model just saves you writing it four ways.

Level three is where a real segmentation setup pays for itself. It is also where the risk sits: a model combining facts about a customer into new sentences is the point at which it might combine them wrongly. "You've been with us since 2023 and sent 12 campaigns" is fine.

"You've been with us since 2023 and your best campaign got 40% open" is a claim that needs to be true.

Testing: where AI adds real leverage

A/B testing has always been limited by one thing: the number of variants you can be bothered to write. That ceiling disappears with AI, and the practice should change accordingly.

• Test ideas, not adjectives. "Save time" versus "save money" is a test. "Save time" versus "Save time!" is not.
• Use AI to generate the hypothesis space, then pick two. Testing six variants splits your audience too thin to reach significance.
• Have the model predict which will win and say why. Then compare with the result. This is a surprisingly effective way to teach your team what your audience responds to.
• Keep a running log of what won. Feed past winners into future briefs. This is the single highest-value thing you can do with AI over time, and almost nobody does it.

Seven things to check before you send

1. Every claim is true. Read the copy and ask of each sentence: can I prove this? Delete anything you cannot.
2. No invented features. Particularly after asking the model to make copy stronger. Check it did not add a capability.
3. Links work and point where you think. Models will generate plausible placeholder URLs.
4. Merge fields have fallbacks. A blank first name is worse than no first name.
5. Reading it aloud sounds like a person. The fastest test for output that is technically clean and obviously generated.
6. One clear action. If there are three CTAs, there is none.
7. It is not the same email you sent last month. Open your last three sends. If they share a shape, that is the sameness problem and no prompt will fix it, only a different brief will.

Prompts worth keeping

Four that do most of the work:

1. Tighten this without adding claims:
   [paste copy] — cut at least 30% of the words. Do not introduce any
   fact, number or capability that is not already present.

2. Find the generic in this:
   [paste copy] — identify every sentence that could appear in an email
   to any audience. List them. Do not rewrite yet.

3. Diversify the angles:
   [paste brief] — give me 8 approaches to this email that argue from
   different premises. Label each with the premise in five words.

4. Adversarial review:
   [paste finished email] — you are a skeptical recipient who gets 40
   marketing emails a day. What makes you delete this? Be specific.

The fourth is the most valuable and the least used. Asking a model to critique the output it helped produce sounds circular, but it reliably surfaces the weak claim and the vague CTA that you have stopped seeing.

How this works inside Cresca

Cresca's campaign generation takes a description of what you want to send and produces a structured campaign: subject line, content blocks and layout. The intent is to remove the blank page, not to remove you.

Two things about how it is built that matter for this topic:

• You generate, then edit. The output is a draft you work on in the editor. Nothing sends without you reviewing it, and the review step is where all the judgement above happens.
• Segments come from your real contact data. When a campaign targets a segment, the copy is working with actual attributes rather than invented ones.

The honest limitation: no generator knows your product's specifics or your customers' current mood. That is the part you supply. A campaign generator is a production accelerator; it is not a substitute for knowing what you want to say.

The short version

AI is excellent at producing options and terrible at making choices. Use it to generate twenty subject lines, four variants of a paragraph, a restructured version of your draft. Keep the decision about what to say, and the verification of every claim, with a person.

Brief it with specifics it cannot invent. Constrain it more than you instruct it. And read your last three sends occasionally, because the failure mode of AI email is not bad writing.

It is every email being the same email.

Continue learning

Related Cresca resources

• Email marketing with Cresca
• Explore Cresca features

External references

• Gmail email sender guidelines
• RFC 8058: one-click unsubscribe