Most writing about AI in email describes features. This describes a week: what you actually do, in what order, and which parts a model handles.
The version below takes roughly forty minutes a week for a weekly broadcast plus a maintained set of automated flows. The point is not speed for its own sake — it is that the time gets spent on the parts that need a person.
Monday: decide (10 minutes, you)
Decide what this week's email is about and who gets it. This is the part that cannot be delegated, because it depends on what you know about your business and what you are willing to promise.
Write four lines: the audience, the single action, the three facts that make the case, and anything you must not claim. That brief is the input for everything downstream and it is the difference between a usable draft and a generic one.
What AI can help with here is the audience question, if your data is deep enough to make it non-obvious — "which segments have not received anything in three weeks?" is a question worth asking an assistant connected to your account rather than working out by hand.
Monday: segment (5 minutes, assistant)
If you use an MCP server or an API, this is a sentence rather than a dashboard session:
"How many subscribers bought once in the last 90 days and have opened at least one email since? What's their average order value compared to the whole list?"
Answered from real numbers. If the segment turns out smaller or larger than you assumed, the brief changes — which is the point of asking before writing rather than after.
Monday: draft (10 minutes, AI, then you)
Feed the brief plus 200 words of your own past writing as a voice sample. Ask for three versions.
Then edit. The edit pass is where the value is: delete the rhetorical opening question, cut the closing summary paragraph, remove the tricolons, check every factual claim, and confirm there is exactly one call to action.
Expect to keep the structure and rewrite a third of the sentences. If you are keeping the draft unchanged, the brief was not specific enough and the email is generic.
Monday: subject lines (5 minutes, AI, then you)
Generate twelve with the email and your best-performing past lines as context. Delete the eight that match the known patterns — colon constructions, curiosity with no content, manufactured urgency. Pick the two most different survivors.
Write the preview text yourself, as a continuation of the subject rather than a repeat.
Monday: check (5 minutes, you)
The pre-send checklist, compressed: click the main link, preview with an empty first name, check the segment name and the recipient count, confirm the subject line is the one you meant.
The recipient count is the one that catches the expensive mistakes. A number ten times larger than expected means the segment is wrong.
Monday: send (1 minute)
Send with per-subscriber send time prediction on, unless the content has a deadline within hours.
Friday: review (5 minutes, assistant)
"How did Monday's campaign do compared to the last four? Break it down by segment, and tell me whether anything unusual happened with bounces or complaints."
Log the subject line result against your running record. One campaign tells you nothing; ten start showing patterns. See email A/B testing.
Monthly: the things that actually matter
The weekly loop above is maintenance. The monthly work is where the results come from, and none of it is AI-driven:
- Suppress the dormant. Anyone with no engagement in twelve months. This protects inbox placement for everyone else and matters more than any copy decision.
- Check complaint rate in Postmaster Tools.
- Review one automated flow. Flows produce most of the revenue and get looked at least, because nothing forces you to.
- Look at engaged list size, not total list size.
The three things to keep for yourself
Frequency. Cheap drafting makes sending more feel free, and send frequency is the strongest driver of complaints. Decide the cadence separately from the tooling and hold it.
Claims. Models produce confident, wrong specifics. Every factual statement in a sent email is yours.
Judgement about what to send at all. An assistant will happily help you write an email that should not exist. Whether this week's email is worth someone's attention is not a question a model can answer, because it does not know what you sent last week or what you are trying to build.
Frequently asked questions
How long should an AI-assisted email workflow take?
Around forty minutes a week for a weekly broadcast plus flow maintenance — roughly ten minutes deciding, twenty on drafting and editing, five checking and five reviewing. The saving comes from drafting and analysis, not from judgement.
What part of email marketing should not be automated?
Deciding what to send and how often, and verifying factual claims. Both depend on knowledge of your business and your commitments that a model does not have, and both are where the expensive mistakes happen.
Do I need an MCP server for this workflow?
No — the copy-paste version works. A connected assistant removes the export-and-analyse step and lets you ask questions of your real data, which is where most of the time saving in the analysis half comes from.
How do I stop AI making my emails generic?
Supply the actual facts in the brief, include a sample of your own writing, and edit with a specific deletion list. If a draft comes back usable without changes, the brief was too vague — a good brief produces something you have to shape.
