AI email writing: what it gets wrong and how to review it

We build an AI writer into Mailnaptic, so you might expect us to tell you to trust it. We would rather tell you how to check it. A draft that reads well is not the same as a draft that is true, legal and ready to send, and the gap between those is where the damage happens: a customer who tries to claim a discount you never offered, a subject line that promises something the body does not deliver, a layout that falls apart in Outlook.

This post covers the six ways AI-written email tends to go wrong, a review routine in order of risk, what to put in the brief, and the parts we think should stay human. If you are still setting up your wider programme, the 30-day plan for a small business gives the context this fits into.

Six ways AI-written email goes wrong

None of these are exotic. They come from the same cause: a language model produces plausible text, and plausible is not the same as accurate. If you give it no facts, it supplies some.

1. Invented claims and made-up offers

Ask for a spring sale email without saying what the sale is and you will often get a percentage, a deadline and a flattering adjective about your products. All three are guesses. The risk is practical (customers expect the offer) and legal. The FTC's deceptive pricing guide says a former-price comparison is only valid when it reflects the price at which the item was genuinely offered "on a regular basis for a reasonably substantial period of time", and our reading is that a discount the model made up cannot meet that standard. This is not legal advice; check the rules where you sell.

2. A subject line that does not match the body

The US CAN-SPAM guide for business is direct: "The subject line must accurately reflect the content of the message." A model asked for "an attention-grabbing subject line" will optimise for attention. The fix is to write the body first, then judge the subject against it. We go deeper on this pairing in subject lines and preheaders as a pair, and on claims specifically in using AI for subject lines without making claims you cannot back up.

3. False urgency

Models love "last chance", "ends tonight" and "only a few left". If the deadline is not real, you are teaching your list to ignore your deadlines, and you may be making a deceptive claim. Tone settings matter here: in the Mailnaptic writer, "Urgent" is one of six tones (Professional, Friendly, Casual, Urgent, Formal, Playful). Choose it only when the urgency is true.

4. Tone drift

The model can sound upbeat in a message about a failed payment, or stiff in a welcome note. Tone also drifts across a longer email: the first paragraph sounds like your brand and the fourth sounds like a press release. Reading the opening line aloud usually exposes it.

5. Repeated phrasing

AI copy loves a frame and then reuses it: "designed to help you relax, designed to help you sleep". Across a month of sends the same openers appear, and subscribers notice. Mailnaptic has an "Avoid recent topics" option (last 5 and up) so the writer does not repeat what you recently covered, but it cannot hear your cadence for you. Read the first three words of every sentence in a draft; if they match, rewrite.

6. Broken or fragile HTML

If you ask a general chatbot for raw email HTML, you may get something that looks right in a preview and breaks in the inbox. The next section covers what two independent tests found.

What testing of AI-generated HTML found

MailerLite published a test on 25 August 2026. They asked Gemini, ChatGPT and Claude for the same email using identical prompts, previewed the results in Gmail and Outlook on desktop and mobile, and had developers read the code. Their summary: all the emails had issues when displayed in the inbox. The specifics are useful because they are so ordinary. One benefits block was misaligned on mobile. One output used http:// links instead of https://. One dark background was ignored by Outlook and Gmail mobile and reverted to white. Unicode characters stood in for icons, and the preheaders were not handled cleanly. With shorter prompts, all three tools left out the business address that CAN-SPAM requires. Their full write-up is worth reading, including their reasoning for why a polished builder helps.

Knak ran a separate test with nine coding tasks across three models and reported that none of the nine outputs was production-ready, with most failures in Outlook for Windows. The two tests used different methods and neither is a measure of how often this happens in general. What they agree on is the pattern: the HTML looks fine in a preview and fails in specific clients.

Rehearsal is the cheap insurance here. Our post on rehearsing a campaign before you send it covers test sends and previews, and the email accessibility checklist covers the things AI tends to skip, such as alt text and sensible heading structure.

A review routine in order of risk

Review the things that can hurt you first, then the things that make the email better. The sequence below is ours, built from the failure modes above rather than from any standard, so adapt it. It should take about ten minutes for a normal newsletter.

What to feed the model

Most bad AI copy starts with a thin prompt. When the model does not know the offer, it writes one. A good brief is mostly facts: what you are selling, the exact price or discount and its dates, who reads it, the single action you want, and the words you never use. The more of that you supply, the less there is to invent.

In Mailnaptic you can set this once per sending domain. Under the domain's AI settings there are fields for Brand description, Target audience, Default writing tone and Writing rules (optional), and the writer picks up the default tone. In the campaign builder, "Write email" takes your brief, a tone and a length (short is 3 to 4 blocks, medium 4 to 6, long 6 to 8). There is also "Start from URL": paste a landing or product page and the writer extracts key points into a brief. Always read the extracted points. They are only as accurate as the page, and a stale page gives you a stale email.

Once you have a draft, the Brand Voice check compares it with your domain's brand settings and returns a score with detected and expected tone. In the editor, the thresholds read 75 and above as a strong match, 50 to 74 as partial and below 50 as a mismatch. Treat the score as a prompt to reread, not a verdict. It checks tone against your settings; it does not know whether your discount exists. For the prompting side, see prompts for email copy that keeps your brand voice.

What stays human

Three things we would not delegate. First, promises: anything that costs you money or creates an expectation, like a discount, a guarantee or a delivery date. Second, anything touching an individual in a hard moment, such as a refund dispute, an outage apology or a complaint. Third, the decision to send. Name one person who approves each campaign, even on a team of two.

Judgement about your own voice matters too. Models average the internet, and your email wins by being specific to your business. Use AI to get past the blank page and to tidy, then put in the detail only you have: the story about why the batch was late, the opinion on which colour to buy. That is what readers remember.

The full picture of where this fits in a small programme, from list building to cadence, is in our 30-day email marketing plan.

A workflow that holds up

Put the review into the process so it does not depend on a good day. Write the brief with facts, generate once, check facts and claims against your records, edit for voice, send a test, and let a named person approve. Keep a short note of the mistakes you catch. After a month you will see which ones your model repeats and can add them to the writing rules.