AI drafts
AI email marketing
Using models to draft campaigns versus letting them send. Where they help, where they invent, and the review queue you still need.
Draft is a feature. Send is a decision.
AI email marketing, in practice, is a model proposing subject lines, body copy, or segments, and a person deciding whether that proposal may leave the building. ESPs added 'generate' buttons because blank pages scare people. The button does not know your refund policy, your unpublished pricing, or which competitor you will not name. If the same button can also schedule the campaign, you have a brand incident waiting for a tired afternoon. HubSpot and others ship this as a draft aid. Keep it as a draft aid.
Keep generation inside a draft state. The send path should still require the same approvals you use for a human writer: facts, legal claims, audience, and a test to a real inbox. Speed is the benefit. Skipping the test is how a model ships 'we are #1 in Europe' to 80,000 people.
Draft
A model proposes a subject and a body
then
Human
Facts, claims, audience
then
Send
After a real inbox test
What models are decent at
Restating a brief you already wrote. Offering three subject variants when you already know the offer. Turning a shipping delay into a calmer paragraph. Summarizing a blog post you trust into a shorter email, if you check the summary against the post. Those are typing assistants.
They are also decent at suggesting a structure: welcome, then proof, then ask. Structure is cheap. You still have to fill it with true specifics. If the brief is empty, the model will invent a customer named Sarah and a 40% lift. Delete Sarah.
Brand and legal risk
Models borrow tone from the internet, including other brands' slogans and leftover marketing cliches. They will claim certifications you do not have, quote reviews that do not exist, and offer discounts finance never approved. In regulated categories (health, finance, kids), that is not a quality problem. It is a compliance problem. Put prohibited claims in the instructions, and still read the output. Instructions are not a guarantee.
Personalization from CRM fields is riskier than it looks. A model that writes 'sorry your deal stalled in legal' to the wrong person has created a story about a deal that might be confidential. Limit merge data. Do not dump the whole record into a prompt. For one-to-one sales mail, a human should still own the send.
Automation plus generation is the sharp edge
A flow that generates a unique body for every abandoner sounds personal. It also means nobody read the message. Unique HTML is harder to QA. Unique claims are harder to recall. If you use generation in a flow, constrain it to a template with locked legal lines and a short free-text slot, or generate once per campaign version and reuse.
Measure the program the same way as before: complaints, unsubscribes, clicks, revenue. If generated mail gets more clicks and more complaints, you did not find a growth channel. You found a way to annoy people faster.
Common questions
Should AI pick the audience too?
Not without a human looking at the resulting list. Models will optimize for who might click, which can include people who opted out of that topic or who are already in a sales dispute. Audience is a permission decision.
Can I train a model on our best campaigns?
You can use past mail as examples if you have the rights and you strip personal data. That helps tone. It does not prove the next claim is true. Keep a fact sheet the model is not allowed to contradict.
Related
Prefer markdown? AI email marketing.md