The thing AI is worst at in email is the thing everyone uses it for first.
Writing the email. Not because the prose is bad — it's usually fine — but because a competent generic email is precisely what your subscribers already ignore, and "fine" was never the bar. Email is a relationship medium. The whole value is that it sounds like a person they know.
Where AI genuinely earns its place is one step back from the writing: generating options you choose between, finding patterns in behaviour you couldn't see manually, and doing the assembly work around the copy.
Here's the honest division.
The short version
Where AI genuinely helps:
- Variation — twenty subject line options so you can pick two worth testing
- Segmentation — finding behavioural patterns across a list too big to eyeball
- Research — mining customer language from reviews, support tickets, replies
- Assembly — formatting, adapting a piece into a different length, drafting the mechanical parts
Where it reliably hurts:
- Writing the whole email unedited
- Personalisation that's technically impressive and obviously fake
- Producing volume because you can
And the deliverability question, answered directly: filters don't detect or punish AI-written text. There's no AI detector in Gmail deciding your fate. What they detect is sameness at scale — which is a different problem with a different fix.
Variation, not creation
The reframe that makes AI useful in email.
Creation: "Write me a subject line for this email." You get one option, chosen by a model, reflecting the statistical centre of subject lines.
Variation: "Here's my email and my audience. Give me twenty subject lines across five different frameworks." You get twenty options and you choose.
Why the second works far better: models are genuinely good at generating volume and genuinely bad at judgement. You have the judgement — you know your audience, you know what you tried last month, you know which promise you can actually keep. What you don't have is the patience to write twenty options at nine on a Tuesday.
Play to the split. Let it produce the field; you pick the winner.
The same applies beyond subject lines. Five different opening lines. Three ways of framing the same offer. Four CTA phrasings. In every case you're using it to widen the option set, not to make the decision.
Segmentation
The place AI adds something you genuinely can't do by hand.
What it's good at: finding behavioural patterns across thousands of subscribers. Who's likely to buy soon. Who's drifting toward dormancy before they've fully gone. Which purchase patterns predict a second order.
What it needs from you: enough behavioural data to learn from. Predictive segments built on a thin history produce confident nonsense, and the confidence is the problem.
And the caveat worth holding: for most small lists, a simple rule beats a model. "Opened in the last 30 days" is a segment you can build in two minutes, understand completely, and explain to anyone. A predictive segment you can't interrogate is harder to trust and harder to debug when it stops working.
What predictive segmentation actually requires.
Send-time optimisation: mostly theatre
Direct, because it's sold hard.
Most platforms now offer AI-driven send-time optimisation, promising to deliver each subscriber's email at their personal optimal moment.
The honest position from a deliverability standpoint: send time affects deliverability only marginally. Consistency of cadence matters far more than clock time.
And the practical reality: if your engagement is poor, send time isn't why. Subject lines, relevance, list quality, and placement all outrank it by a wide margin. Optimising send time on a list that isn't engaged is polishing a doorknob on a house with no roof.
When it's worth switching on: you have a large list, everything else is working, and you want a marginal improvement for zero effort. It's free on most platforms and it won't hurt.
When it's a distraction: anything before that point.
Personalisation that isn't creepy
AI makes deep personalisation technically easy, which is exactly the problem.
What works: using a first name naturally in a sentence. Referencing a genuine action they took — a purchase, a download, a page visit that makes obvious sense in context. Segmenting so the content is relevant rather than inserting tokens into generic copy.
What backfires: "I saw you were looking at…" Technically true, and it makes people uncomfortable in a way that costs you more than the sale is worth. Personalisation that reveals how much you're tracking crosses from helpful into unsettling, and the line is closer than marketers think.
And a specific one worth naming: personalisation tokens are mildly positive at best. "{first_name}, check this out" fools nobody in 2026 — everyone knows how it works. Relevance is what makes an email feel personal, not a merge field.
The test: would you say this out loud to the person? "Hi Sarah, I noticed you looked at our boots three times last week" is not a sentence a human says.
The three places AI reliably hurts
1 — Whole emails, unedited. The output is competent and generic, which in a relationship medium reads as a stranger. Your subscribers signed up because of something specific about you, and averaged prose removes exactly that.
2 — Volume because you can. Capacity becomes the plan. More sends to a list that isn't asking for them produces more complaints and lower engagement — and engagement is the single strongest input to whether you reach the inbox at all. Sending more is one of the fastest ways to reach fewer people.
3 — Sameness across sends. Same-shaped emails week after week decay engagement regardless of who wrote them. If every send has the same structure, the same length, and the same rhythm, subscribers stop registering them in the inbox list. Vary the shape deliberately.
Do filters detect AI-written email?
The question everyone has, so here's the direct answer.
No. Filters do not detect and punish AI-written text. There is no AI detector in Gmail deciding your fate.
What filters actually detect is sameness at scale — structural similarity across many messages. And that's where careless AI use becomes a genuine risk. If ten thousand senders prompt the same model for "a friendly newsletter about X," the structural similarity across all those emails is precisely the pattern-matching filters are built to notice.
The risk isn't the tool. It's volume plus sameness.
One thoughtful AI-assisted newsletter a week is invisible to fingerprinting. Fifty thousand AI-spun cold emails with swapped tokens is exactly the thing.
And the second-order point: AI-written mail that nobody opens is still mail nobody opens. The engagement penalty for boring, undifferentiated content is identical whether a human or a machine wrote it. The filter doesn't care about the author — it cares about the reaction.
Voice across a sequence
Automated sequences are where voice drift shows most, because the emails were written at different times, possibly by different tools, and nobody reads them in order after launch.
Two habits worth building:
Read the whole sequence in order, out loud, once a quarter. Drift is obvious when read consecutively and invisible when reviewed individually.
Write the openings yourself. The first two lines of each email carry most of the voice, and they're also the preview text shown in the inbox. That's perhaps 15% of the words doing most of the work.
The voice method. · Email copywriting.
What AI can't do
Three things, and they're what make an email worth opening.
Your results. The number from your own work, the client outcome, the thing that happened when you tried it. Your only genuine information advantage.
Your story. The mistake, the reversal, the specific situation. Situated experience no model has.
Your offer. What you're actually selling, why it's priced that way, who it's genuinely for and who it isn't. That's judgement about your own business.
The test for any email: what's in here that only I could have sent? If nothing, it's an email your subscribers could have got from anyone — and increasingly, one they could have got from a model without you.
A practical division of labour
| Step | Who |
|---|---|
| Research customer language | AI, from real source material you supply |
| Decide the angle and the offer | You |
| Draft the body | AI, from a proper brief |
| Write the opening and any opinion | You |
| Generate subject line variants | AI |
| Choose which to test | You |
| Fact-check every claim | You |
| Approve before sending | You |
Frequently asked questions
Can AI write good marketing emails?
It can write competent ones, which in email is a lower bar than it sounds. Email is a relationship medium — subscribers stay because of something specific about you, and averaged prose removes exactly that. Where AI genuinely helps is generating variants you choose between, drafting from a detailed brief with your own material in it, and handling the mechanical parts. What it can't supply is your results, your story, and your offer, and those are what make an email worth opening.
Do spam filters detect AI emails?
No. Filters don't detect or punish AI-written text — there's no AI detector at Gmail deciding your fate. What they do detect is sameness at scale: structural similarity across many messages. That's why careless AI use is risky, since ten thousand senders prompting the same model produce structurally similar emails, which is exactly the pattern filters are built to notice. One thoughtful AI-assisted newsletter a week is invisible to this; fifty thousand spun cold emails is not.
Is AI personalization worth it?
Segmenting so the content is genuinely relevant is worth a great deal. Inserting more tokens into generic copy isn't — personalisation tokens are mildly positive at best, and "{first_name}, check this out" fools nobody. The failure mode to avoid is personalisation that reveals how much you're tracking. "I saw you were looking at…" is technically true and reads as surveillance, which costs more than the sale is worth. The test is whether you'd say the sentence out loud to the person.
Will AI replace email copywriters?
It's already replaced the mechanical parts — first drafts, variants, reformatting, adapting length. What it hasn't replaced is knowing what's worth saying to this particular audience, having a position worth taking, and supplying the specific results and stories that make an email distinctive. The copywriters losing work are the ones producing competent generic email, because that's now free. The ones doing well are using it to spend more time on the parts that need a person.
What to do next
Take your last three sends and ask one question of each: what's in this that only I could have sent?
If the answer is a number from your own work, a client story, or a position you'd defend, you're using email properly and AI can help you do more of it.
If the answer is nothing — three competent emails that anyone in your category could have sent — that's the problem, and it isn't a tooling problem.
Free: The subject line swipe file.
Go deeper: The Email Copywriting Playbook — the framework library, sequence templates, and the pre-send audit.
Related guides
- Using AI to generate subject line variants — volume, then judgement
- AI-assisted email segmentation — what it needs from your data
- Do AI spam filters detect AI-written emails? — the four signal layers
- An AI-assisted email workflow — end to end
- Email copywriting — the craft underneath
- Why authenticated emails still land in spam — what actually decides placement
Join the Newsletter
Get practical marketing tactics delivered straight to your inbox.
Written by
Muhammad Basim
Related Articles
What Marketing Automation Actually Costs at Scale
Nobody's automation bill jumps because they added more automations. It jumps because they added steps to workflows that were already running — a filter here, an enrichment lookup there, a second notification — and the billing model charges for every one of them, on every run. That's the mechanic behind almost every "why did this […]
One Workflow, Three Tools: A Build Comparison
Feature tables don't tell you what a tool is like to use. The only thing that does is building the same thing twice and noticing where you got annoyed. So here's one realistic workflow — lead routing with enrichment and conditional assignment — specified once and mapped across Zapier, Make, and n8n. What changes between […]
GoHighLevel: 7 Automation Workflows to Build First
GoHighLevel gives you a hundred things you could automate, which is exactly why most accounts end up with forty half-built workflows and no measurable result. Seven are worth building first. They're the ones tied directly to revenue — catching leads before they cool, recovering appointments that would otherwise vanish, and asking for reviews at the […]