Use AI in email for subject line variants, segment descriptions, first drafts and summarising replies. Do not use it to write more emails, to personalise at a scale you cannot verify, or to answer anyone who is unhappy. The rule is that nothing reaches a subscriber without a person reading it.
Email is unforgiving because the cost of a bad send is not a bad post, it is an unsubscribe and a complaint.
Where it genuinely helps
These save real time and carry little risk, because they are verifiable.
- Generating twenty subject lines so you can pick two to test, rather than agonising over one
- First drafts of structured emails: confirmations, reminders, announcements
- Turning an existing long piece into an email, which is a format conversion
- Reading a batch of replies and telling you the themes
- Tightening your own draft, which is editing rather than writing
The volume trap, applied to email
The moment writing becomes cheap, sending more becomes tempting, and email punishes that faster than any other channel.
Frequency increases driven by capacity rather than by having something to say produce exactly the pattern that damages a sender: falling engagement, rising unsubscribes, and eventually filtering that affects the emails people did want.
The capacity gained is better spent on segmentation and on making each send genuinely worth opening.
Personalisation you cannot verify is a risk
Generated personalisation is convincing until it is wrong, and when it is wrong it is memorably wrong.
Referring to a purchase that did not happen, a company detail that is out of date, or a problem the recipient does not have reads worse than no personalisation at all, because it demonstrates that nobody checked.
Personalise on data you hold and can confirm. Anything inferred should be phrased as a question rather than an assertion.
Never automate the reply to an unhappy person
This is the one absolute. A complaint answered by a system that did not understand it converts a solvable problem into a public one.
Use AI to draft the reply if it helps, then have a person read it, change it, and send it. The drafting saves time; the reading prevents the incident.
Deliverability implications
Nothing about AI-generated text is inherently filtered. What gets filtered is the behaviour that often accompanies it.
Sudden volume increases. Sending to segments you have not mailed in a year because a tool suggested it. Templates that all look alike. Content that recipients do not engage with.
The safeguards are the same as ever: authenticate, warm up changes, watch complaints, and keep sending to people who read you.
A workable division
Let AI handle the parts with a right answer and a verifiable output. Keep for yourself the parts that require knowing something it cannot: what actually happened this month, what you are willing to promise, and what tone this specific list expects.
That division holds up in practice and it survives whatever the tools do next.
Common questions
- Where does AI help in email marketing?
- Subject line variants, first drafts of structured emails, converting a long piece into an email, summarising replies, and editing your own draft. All of those are verifiable, which is what makes them safe.
- Can AI-generated email hurt deliverability?
- Not the text itself. What hurts is the behaviour that often comes with it: sudden volume increases, mailing segments you have not touched in a year, and content recipients do not engage with.
- Should AI reply to customer emails automatically?
- Never to an unhappy one. A complaint answered by a system that did not understand it turns a solvable problem into a public one. Draft with AI if it helps, then have a person read and send it.
Keep reading