Search "AI email marketing" and almost everything you find is a tool trying to sell you a subscription. Very little of it tells a small business owner what AI is actually good at inside email, what it is bad at, and which of the two your money is currently buying.
This is that, written for someone with a customer list, no marketing department, and no appetite for another monthly fee that produces nothing.
What AI genuinely does well in email
Deciding who to send to, and when
This is the part that actually moves revenue, and it is the part nobody talks about, because it is not a feature you can screenshot. Most small business lists get one email to everybody at the same hour. AI segmentation looks at who opened what, who bought when, and who has gone quiet, and splits the list accordingly. The same message sent to the right third of a list routinely outperforms the same message sent to all of it — fewer sends, more replies, and less damage to your sender reputation.
If a tool offers you one AI feature, take this one over any writing feature.
Subject line variants and testing
Generating twenty subject lines is a genuinely good use of a language model, because the hard part of subject lines is not writing them, it is having enough of them to test. Where this goes wrong is letting the model pick the winner. It cannot. Send two, to two halves of a segment, and let the open rate decide.
Turning something you already wrote into an email
You wrote a long reply to a customer question. That reply, condensed, is a good email — and AI is reliable at that particular transformation, because the thinking is already done and the facts are already yours. This is the one writing task worth automating, and it is also the one that keeps the email sounding like you.
Cleaning the list
Duplicate records, dead domains, typo'd addresses, people who have not opened anything in fourteen months. This is dull, mechanical work that directly affects whether your email reaches an inbox at all, and it is exactly what a script or an AI-assisted tool should be doing instead of you.
What it does badly, and what that costs you
AI writing an entire campaign from a prompt is the feature every tool leads with and the one most likely to lose you money. Three reasons, in order of how expensive they are.
It writes email that reads like email. Generic, competent, forgettable — indistinguishable from the other nine promotional messages that arrived that morning. Small businesses do not win on polish. They win because the message sounds like a person the reader has met.
It invents specifics. Prices, timelines, guarantees, availability. A model asked to write a persuasive offer will produce a persuasive offer, and it does not know which parts of it you can honour. Every claim in an email you send is a claim you are legally standing behind.
It trips the filters it was supposed to beat. Deliverability is the whole game — an email in spam has a conversion rate of zero regardless of how good the copy is. Send volume, list hygiene and sender reputation determine that far more than word choice does, and they are the three things a generated campaign does nothing to improve.
The setup that works for a business without a marketing team
In order. Each step is worth doing even if you never get to the next one.
Authenticate your domain first. SPF, DKIM and DMARC records, correctly configured. Without these, a growing share of your email is filtered before anyone decides whether it was interesting, and both Google and Yahoo now enforce this for bulk senders. This is a one-time technical task with a larger effect on results than any amount of copy work, and it is the most common thing I find missing.
Use one list, segmented, not several lists. Duplicate records across separate lists are how people receive the same email twice and unsubscribe from all of it.
Write the first draft yourself, then let AI edit it. Not the reverse. Draft it badly and fast in your own words, then ask for a tighter version, three subject lines and a shorter opening. The voice survives, the polish arrives, and no facts get invented, because every fact came from you.
Send less than you think. A monthly email that people read beats a weekly one they have learned to delete. Frequency damages reputation faster than almost anything else.
Measure replies and sales, not opens. Open rates became unreliable when Apple began pre-loading images in 2021, and most dashboards still report them as though they mean something. Track clicks through to your site, and enquiries that arrive afterwards.
One thing worth knowing about AI on the other side
The same tools have made invoice fraud and impersonation email dramatically cheaper to produce. The badly-worded scam email is largely gone; what arrives now is clean, correctly formatted, and often references a real supplier of yours. Two consequences for anybody sending business email.
Domain authentication protects your customers from someone spoofing you, not only your own deliverability. And the habit of verifying a payment-detail change by phone, using a number you already had, is now the only reliable defence — because the email will look right.
Where to start
If you do one thing from this page, check whether your domain has SPF, DKIM and DMARC records. If you do two, segment your list once and see what a smaller send does to your reply rate.
Email is usually not the first thing costing a small business customers — it tends to be the website, or being invisible in search. Which is why it is worth knowing where you actually stand before spending money on any single channel.