AI Written Email Etiquette: What to Send and What to Rewrite

AI written email etiquette comes down to one test: would you sign this if a colleague read it over your shoulder? Using a model to draft, tighten or translate an email is normal now. Sending one where the personal parts have clearly been outsourced is what reads as careless, and recipients notice far more often than senders expect.

Where AI help is normal and where it reads as careless

Nobody objects to a model doing the scaffolding. Brainstorming a subject line, turning three bullets into a paragraph, shortening a long update, or drafting a first version of a routine status note are all fine. So is translation and tone softening when you are writing in a second language and want to avoid sounding blunt.

What reads badly is when the model writes the parts that only you can know. A client who has worked with you for two years can tell when the paragraph thanking them for their patience is generic. The tell is not vocabulary. It is the absence of specifics: no reference to the call on Tuesday, no acknowledgement of the thing they actually asked about.

Our own measurement, run on 2026-09-13 with the DeAIze detector on a 30% flag threshold, found that 12 unedited model passages scored a mean AI score of 33%, while 23 hand-written DeAIze guides scored a mean of 21%. Two things are worth taking from that. First, unedited output is measurably more machine-patterned than human writing. Second, the gap is not enormous, which is why you cannot rely on a tool to tell you whether an email is fine. You have to read it.

Situation	AI draft is fine	Rewrite yourself
Routine internal update	Yes, with your own first line	—
Scheduling and logistics	Yes	—
Client apology or delay	Structure only	The whole apology
Bad news or a mistake	Maybe a skeleton	Every sentence the reader will reread
Sensitive feedback to a colleague	No	All of it
Cold outreach	Opening line only	The reason you are writing to them

The three things to always rewrite yourself

If you remember nothing else from this piece, remember these three. They are the parts recipients quote back to you, forward to other people, and remember.

1. The first line. The greeting plus the opening sentence sets the temperature. A model defaults to something like “I hope this email finds you well” or “I wanted to reach out regarding”. Both are dead on arrival. Replace with the actual reason: “Quick update on the Henderson timeline — we slipped by a day.” One concrete noun beats three polite ones.

2. Anything that carries an emotion or a judgement. Apologies, thanks, disagreement, a gentle push, a refusal. Models smooth these into a neutral register that reads as insincere precisely because it avoids committing. If you are sorry, say what you are sorry about. If you are pushing back, say what you need changed.

3. The ask. The single sentence that tells the reader what happens next. Models bury it, hedge it, or split it into three sentences. Write it yourself, put it near the end, and make it a verb: “Can you confirm by Thursday?” or “Send me the file and I will handle the rest.”

If you are worried a draft has drifted into generic territory, the same sentence-level reading helps: DeAIze’s AI detector highlights which lines carry the machine-patterned signal, so you can see exactly which sentences to rewrite by hand.

How tone shifts between colleagues and clients

Internal email is short, blunt and full of shared context. External email is longer, more careful and assumes no context. A model does not know which one you are writing, so it defaults to the careful version, which is wrong for your team and often too thin for a client.

With colleagues, cut the throat-clearing. Drop “I just wanted to” and “just following up”. Say “Following up on the pricing doc.” Internal readers reward brevity and read terseness as competence, not rudeness.

With clients, the opposite applies in one specific way: you need to name the context and the consequence. Not more adjectives — more nouns. Which project, which date, which person, which number. A model will happily write a paragraph that sounds warm and says nothing. Read it back and ask whether a client who missed the last two calls would understand what you are telling them.

The gap widens when you are writing in a second language. A model can fix register quickly, but it also flattens the small idiosyncrasies that make a client trust that a specific person is on their account. Keep one or two of your own phrasings, even if they are slightly less polished.

Should you disclose that AI helped?

Usually no, and occasionally yes. Routine drafting help is the same category as spellcheck or a template. Nobody discloses spellcheck.

Disclose when the reader has a legitimate reason to care who wrote it. That means: when the email contains a commitment, a price, a legal position or a professional judgement attributed to you. If a client asks directly whether a message was written with AI, answer honestly and briefly. “I used a tool to tighten the draft, the content is mine” is a complete answer and ends the conversation.

If your workplace, your client contract or your professional body has a rule about AI-generated communication, follow it over any general advice here. Our guide to AI content disclosure walks through the same decision for published content, and the logic transfers to email.

One honest limit: detector output is a probability estimate, not proof of authorship. If someone runs your email through a tool and gets a high reading, that is not evidence you did not write it — and it is not evidence you did. Do not let a score settle an argument about intent.

An editing pass in five steps

Do this before you hit send on anything that matters. It takes about as long as rereading the email twice, which you were going to do anyway.

1. Read the first line out loud. If it sounds like a form letter, rewrite it with one specific detail from the actual situation.
2. Find the ask. Circle it mentally. If you cannot find it in one pass, the reader will not find it either. Move it near the end and make it a verb.
3. Delete every sentence that only softens. “I just wanted to”, “I hope you don’t mind”, “at your earliest convenience”. Keep one if the relationship genuinely needs it.
4. Check the facts you own. Dates, names, numbers, commitments. Models occasionally invent a plausible detail, and the reader will act on it.
5. Read the last line. If it is a generic sign-off stacked on a generic closing sentence, cut one of them.

A worked illustration. An AI-drafted client email opened with: “I hope this message finds you well. I wanted to follow up regarding the project timeline and see if there are any updates you might be able to share at your earliest convenience.” Rewritten by hand: “Checking in on the Henderson timeline — are we still on for the 14th? If the review is holding things up, tell me and I will chase it.” Same request. Half the words. The reader knows exactly what to do next.

When the draft is mostly fine

Most AI-assisted emails do not need a rewrite. They need a first line and an ask. That is roughly two sentences of your own work, and it is the difference between a message that reads like you and one that reads like a template with your name on it.

If you are editing a longer document rather than a short email, the same principle scales: fix the seams, leave the rest. Our guide to mixed human and AI writing covers how to find those seams in longer drafts without rewriting everything.

Write the parts only you can write. Let the model handle the rest.

---

Measurement note: figures in this article come from our own detector run on 2026-09-13 — n = 23 hand-written guides versus n = 12 unedited model outputs, median AI score 22% and 34% respectively. Method: /methodology/.