How to write marketing emails with AI, and what to check before you send

From brief to approved send#

1. 1

   Write a five-line brief

   State the goal, the audience, the offer with its exact terms, the one product or story to feature, and anything that must not appear. A brief this short takes two minutes and removes most rewrites.

2. 2

   Ask for a draft and variants

   Ask for one full email plus three subject lines and two opening lines in different directions. Variants show you options you would not have written, and you keep the one that fits.

3. 3

   Edit in plain words

   Reply with specific changes: “Cut the intro to one sentence”, “Use the fabric name, not ‘premium materials’”. Each pass should change one thing so you can see what improved.

4. 4

   Check facts and claims

   Compare every price, date, discount code, product detail and claim against your store. The model writes what sounds plausible, not what is in your stock system.

5. 5

   Preview and test

   Look at the email on a phone and on desktop, and send yourself a test. Read the subject, the preheader and the footer as a recipient would.

6. 6

   Approve the send

   Confirm the audience is people who agreed to hear from you, then send or schedule. With a connector like Tellnod, this is the confirmation card that only you can press.

What AI does well in email marketing#

Language models are fast at the parts of email marketing that are mostly writing. For a small shop without a copywriter, that covers most of the hours a campaign takes. The model is weaker wherever the answer lives in your own data: stock levels, real prices, who opted in, and what your customers bought.

A sensible split: let the assistant produce options and do the rewriting, and keep the decisions (what to offer, to whom, and whether it goes out) with a person.

Where a person has to check#

Anthropic’s own guidance says Claude can produce text that is factually incorrect, and that critical information should always be validated ^S2. ChatGPT works the same way. In a marketing email, an invented detail is a promise to customers, so five things always get a human review.

• Facts and prices. The model fills gaps with plausible details: a discount that sounds right, a delivery promise, a fabric weight. Paste the real numbers into the brief and check the draft against them.
• Claims. In the US, the FTC expects advertising to be truthful and non-deceptive, with evidence for a claim in hand before the ad runs ^S3. “Warmest sweater we have made” needs a basis; “100% merino” needs to be true.
• Subject lines. CAN-SPAM prohibits deceptive subject lines ^S4, and Google’s sender guidelines say the subject and display name should accurately represent the sender and content, with no “Re:” or “Fwd:” unless the message is a real reply or forward ^S5.
• Consent. A model can suggest an audience; it cannot know who agreed to hear from you. Send only to contacts with a consent record. See email marketing consent laws.
• Brand voice. Drafts drift toward a generic register. Read the email aloud; if it could be from any shop, it needs your details.

The FTC is clear that a business stays responsible for its marketing email even when another company sends it ^S4. The same logic applies to a draft from an AI: the sender owns the content.

Give the model your brand context#

Anthropic suggests treating Claude like a capable new employee who lacks context on your norms, and explaining the reason behind an instruction, because the model generalizes from the explanation ^S1. For email, that context is a brand brief you write once and paste at the start of each chat, or keep in the place your AI app stores standing instructions.

• Who you are in one sentence, and who buys from you.
• Three words for your voice, and three it should never sound like.
• Words you use and words you avoid (for example “knit” instead of “garment”).
• Your standing terms: shipping threshold, returns window, how you handle discounts.
• Two past emails that worked, pasted in full as examples.

An example brief

“We are Ninefold, a small apparel brand selling knitwear, outerwear and basics online. Customers are mostly 28–45, buy a few good pieces a year, and care about fabric and fit. Voice: plain, warm, specific. Never: hype, exclamation marks, ‘must-have’, ‘elevate’. We say ‘knit’, not ‘garment’. Free shipping over $120, 30-day returns. We run two sales a year and do not invent urgency. Below are two emails our customers liked.”

Examples carry more weight than adjectives. Anthropic recommends a few relevant, varied examples as one of the most reliable ways to steer tone and structure ^S1; OpenAI’s prompting guide also advises giving examples and the context the model needs ^S6.

How to avoid generic AI copy#

Generic copy comes from generic input. When the brief says “write an email about our new collection”, the model has nothing to work with except the average of every new-collection email it has seen. These habits fix most of it.

• Give one concrete detail per email. The yarn, the mill, why a coat has a two-way zip. The model builds around it.
• Name the phrases to avoid. List the openers and words you never want (“In a world where”, “elevate your wardrobe”). Explain why: your customers read past them.
• Ask for one idea, not five. An email that features one product and one reason to click reads like a person wrote it.
• Cut the opening line. Drafts often warm up for a sentence before saying anything. Ask the model to start with the news.
• Write the call to action yourself. “Shop the Harbor coat” is clearer than “Discover more”.
• Keep a swipe file. Save the drafts you sent and the edits you made; paste them next time as examples.

For subject lines specifically, see email subject lines for e-commerce, and ready-made requests in ChatGPT prompts for email marketing.

Review checklist before you send#

• Every price, discount code, date and end time matches the store.
• Every product named exists, in the sizes and colors mentioned.
• Links go to the right pages and carry your UTM tags if you use them.
• Claims are true and you could show why (materials, origin, “best-selling”).
• The subject line and preheader describe what is inside, with no fake “Re:” or false urgency.
• Merge tags have a fallback, so nobody gets “Hi ,”.
• The audience is people who opted in, with unsubscribed and bounced addresses excluded.
• The footer has your postal address and a working unsubscribe link ^S4. The email footer checker can confirm both.
• You have read it on a phone, and sent yourself a test.

An assistant that drafts versus one connected to your list#

Claude or ChatGPT on its own writes text. You then copy it into an email platform, rebuild the layout, pick the audience there and send. With an MCP connector, the assistant can work with the email tool directly: it reads your list, renders the real email and prepares the send, while the approval stays with you.

With Tellnod connected, the same brief becomes a working session. You might type: “Draft an email to customers who bought knitwear last winter about the new merino crewnecks, no discount, link to the knitwear page.” The assistant drafts it and shows the preview card; you ask for changes in words; you ask for a test; then it prepares the send and the approval card shows who receives it and who is excluded and why. The assistant cannot press Confirm, and a scheduled send can be cancelled until it goes out.

The checks above still apply: Tellnod adds the unsubscribe link, postal address and consent handling, but you still verify the facts in the copy. Setup steps are in the guides for Claude and ChatGPT, or start on the connect page. Tellnod is free up to 500 contacts and 2,000 emails a month; see pricing.

FAQ#