Most people's AI email workflow is one step: paste a topic into a chat window, take what comes back, hit send.
That produces a competent generic email, which in a relationship medium is worse than it sounds — your subscribers signed up because of something specific about you, and averaged prose removes exactly that.
Here's a seven-step version that takes about forty minutes and produces something worth sending. Two of the steps are ones AI should be nowhere near.
The short version
| Step | Who |
|---|---|
| 1. Mine customer language | AI, from real source material |
| 2. Decide the angle and the offer | You |
| 3. Write the brief | You |
| 4. Draft the body | AI |
| 5. Write the opening and the opinion | You |
| 6. Generate subject line variants | AI |
| 7. Fact-check and approve | You |
Steps 2 and 5 are the ones that make it yours. Everything else is assembly.
Step 1 — Mine your customers' language
The most underused AI capability in email, and it takes ten minutes.
What you're doing: feeding it real text your customers have written, and asking what patterns emerge.
Where the source material comes from: support tickets, sales call notes, product reviews, replies to past emails, forum threads in your niche, survey responses.
What to ask for: the exact phrases people use to describe the problem, the words they use for the outcome they want, and the objections that recur.
Why this matters more than a clever subject line: people search and respond to their own vocabulary, not yours. Your customers say "my emails are going to junk." You say "inbox placement optimisation." Writing in their words rather than your professional register is the single biggest improvement available to most email copy.
The constraint to state explicitly: use only the material provided. You want extraction, not invention.
Step 2 — Decide the angle and the offer
AI does none of this. It's the step that makes the email worth existing.
What you decide:
The one thing this email is about. One idea, one email. The more goals you cram in, the fewer clicks each one gets.
The angle. Not the topic — the specific take. Not "email deliverability" but "the thing everyone fixes first is the thing that's usually already fine."
What you're asking for. One primary CTA, and what happens after they click.
What you have that nobody else does. A number from your own work, a client situation, a mistake you made. This is the email's reason to exist.
If you can't answer that last one, stop. An email with nothing specific in it is an email your subscribers could have received from anyone — and that's the category AI has made free.
Step 3 — Write the brief
Where the output quality is actually determined.
What goes in:
- Who it's for, specifically — their situation, not their job title
- The one idea, stated as a sentence
- Your angle, as something you'd defend
- Source material — your data, the customer language from step 1, anything factual it should use, with instructions to use only what's provided and flag gaps rather than filling them
- Must include — your specific example, your number, the CTA
- Must not include — clichés, hedging, "in today's fast-paced world," a summary paragraph at the end
- Length and shape — roughly how long, and whether it's plain-text personal or a designed campaign
- Voice reference — two of your own past emails, pasted in
Ten minutes, and it's the difference between editing and rewriting.
Step 4 — Draft the body
Now let it write.
Ask for the body only. Not the subject line, not the opening. Those come separately and get different treatment.
Section by section if it's long. For a short email, one pass is fine. For a longer piece, drafting in parts gives you control at each step.
What to check immediately:
Does it front-load the point? If it opens with a rambling intro, the reader has to scroll to find out why you wrote. The first two sentences also usually become the preview text shown in the inbox — so they're doing double duty.
Is it one idea? Models drift toward comprehensiveness. If it's covering three things, cut two.
Single column, short paragraphs, real text headings. Walls of text get deleted, and deletes-without-reading are a negative engagement signal.
Step 5 — Write the opening and the opinion yourself
The second step AI shouldn't touch.
The opening. The first two lines carry most of the voice, and they're the preview text. That's perhaps thirty words doing a disproportionate share of the work.
Anything expressing a position. A model's version of your opinion is a summary of your opinion, and it reads differently — hedged, balanced, unmemorable.
Your specific example or number. Write it yourself so it's stated precisely and honestly. "In 50 audits I ran last year, 34 had no valid SPF record" is checkable and citable. "Most companies have authentication problems" is neither.
Together that's maybe 15% of the words and it does most of the work on whether the email sounds like you.
Step 6 — Subject line and preview text
Now the subject line, once the email exists.
Why the order matters: a subject line written before the email promises something the email might not deliver. Overpromising buys one open and spends future ones — a disappointed open is worth less than no open at all.
Ask for twenty variants across four named frameworks. Then cut to three, then test two from different frameworks.
Ask for preview text in the same pass, with the instruction that it should complement rather than repeat the subject line.
Step 7 — Fact-check and approve
Never skip, and do it as a dedicated pass.
Check every: number, name, date, claim about how something works, and every link.
Models produce fabricated statistics and citations with exactly the same confidence as accurate ones. There's no tell in the text — you catch it by checking categories, not by reading carefully.
Then the send checklist:
- Every factual claim verified against a source I can see
- One clear CTA, repeated once near the end
- Preview text set deliberately
- From address is a real mailbox that receives replies
- Sending to an engaged segment, not the full list
- Read the whole thing out loud once
- Test send to Gmail and Outlook, checked in both inbox and spam
What this actually costs
Roughly 40 minutes for a considered email, against maybe 90 doing it entirely manually and 5 doing it badly.
Where the time goes: ten minutes on research, ten on the brief, five reviewing the draft, ten writing the parts that are yours, five on subject lines and checking.
The counterintuitive part: the brief takes almost as long as writing a mediocre draft would. That's the trade — you're moving effort from writing to specifying. The specification is reusable and improves over time; a draft isn't and doesn't.
Can AI write a whole sequence?
It can draft one. Whether you should ship it is a different question, and there's a specific risk.
The problem with sequences: they're written at once, launched, and never read again in order. Voice drift and repetition are invisible when you review emails individually and obvious when you read them consecutively.
Two habits:
Read the whole sequence out loud, in order, before launching. You'll find the same phrase three times and two emails making the same point.
Re-read it quarterly. Sequences run for years, and the offer, the pricing, and the tools you reference all change.
And vary the shape between emails. Same-length, same-structure emails across a five-part sequence decay engagement regardless of who wrote them.
Frequently asked questions
Can AI write a whole email sequence?
It can draft one from a good brief, and the risk is specific: sequences get written at once, launched, and never read in order again. Voice drift, repetition, and two emails making the same point are invisible when reviewing individually and obvious when read consecutively. Read the whole sequence out loud before launching, vary the length and shape between emails, and re-read it quarterly — sequences run for years while your offer and the tools you reference change.
How long does an AI-assisted email take?
Around 40 minutes for a considered send: ten minutes mining customer language, ten writing the brief, five reviewing the draft, ten writing the opening and your own specifics, and five on subject lines and fact-checking. The brief takes almost as long as writing a mediocre draft would, which is the actual trade — you're moving effort from writing to specifying, and the specification is reusable while a draft isn't.
What should I never let AI write?
Two things. The angle and the offer — deciding what this email is about, what position you're taking, and what you're asking for is the judgement that makes it worth sending. And the opening plus anything expressing an opinion, because the first two lines carry most of your voice and double as preview text, while a model's version of your opinion reads as a hedged summary of it. That's roughly 15% of the words doing most of the work.
What to do next
Try step 1 on its own before changing anything else.
Take fifty real customer messages — support tickets, replies, reviews — paste them in, and ask what phrases keep recurring when people describe the problem you solve.
Most people find their customers use completely different words from the ones they've been writing. That single insight improves every email you send afterwards, with or without the rest of this workflow.
Free: The subject line swipe file.
Go deeper: The Email Copywriting Playbook — the framework library, sequence templates, and the 13-point pre-send audit.
Related guides
- AI for email marketing — the wider division of labour
- The content brief — step 3 in detail
- Using AI to generate subject line variants — step 6
- Fact-checking AI output — step 7
- Email copywriting — the craft underneath all of it
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 […]