Email Drafting That Still Sounds Like You

Email is the one place where the measurement is unambiguous

Most AI promises are soft. This one is not. When more than seven thousand people at 66 firms were randomly assigned an AI assistant, time spent on email fell 31 percent — while in the same study the time to complete documents did not change measurably. Email is the task where the gain is real and repeatable.

The second number says even more. In studies of style imitation, models perform very differently by genre: on a personal blog post, identifying the author drops to 17 to 21 percent, while on email it reaches 95 to 97 percent. There is a simple reason. The structure of your emails is more predictable than the structure of your thoughts. You greet the same way, you lead in for the same length, you close with the same phrase. The genre the machine imitates best happens to be the one you produce most of.

The three inputs

A machine-written email becomes recognisably yours because of what goes into it, not because of which model you use. Three things are needed, and most people leave out the third.

• The actual thread. Not your summary of the exchange but the text of the message as it stands. Summarising loses the register, the stakes, and the half-sentence the whole reply exists for.
• What only you know. One to three sentences about what the thread does not reveal: that you had a difficult project with this person last year, that the deadline can in fact slip, that the request is delicate. This is the one input the model structurally cannot supply.
• Three of your own past emails. From the same genre as the one you are writing. This is few-shot logic in practice, and measurements show that more than five samples adds nothing — while three adds a great deal.

The recipe

Five steps. The first is done once; the rest take half a minute per email.

1. Collect three to five sample emails, once. A reply, an awkward message, a reminder — one of each type. Save them in one place and copy from there. This is the step that makes everything after it fast.
2. Give it the thread, the background and the samples. In that order, separated by markers. That way the model knows which is which.
3. Ask for two versions, not one. A shorter one and a firmer one. Choosing between two is faster than repairing one, and the difference between them shows you where your request was vague.
4. Rewrite the first and the last sentence. Those two places give the most away, and they are the two that matter most. Thirty seconds.
5. Check the facts before you send. Date, amount, name, promise. A model invents precisely what you did not give it.

A prompt you can copy

Replace the square brackets. The sample emails are the important part — do not leave them empty.

You are drafting a reply on my behalf. You get three sections below:
the message received, the background only I know, and three of my own past emails.

=== THE MESSAGE RECEIVED ===
[the text of the message as it stands]

=== WHAT ONLY I KNOW ===
[1-3 sentences: relationship, stakes, anything not visible in the message]

=== THREE OF MY OWN EMAILS (for style) ===
[email 1]

[email 2]

[email 3]

TASK: write two versions of the reply.
A) shorter and matter-of-fact  B) firmer, with a concrete request

Rules:
- take the style from the three samples, not from generic email templates
- at most [150] words per version
- do not promise anything that is not in the background
- anything you cannot tell from what is supplied, mark as: [to confirm]
- do not open with "Thank you for reaching out" or "I hope you are well"

Four kinds of email, four different rules

Not every email is the same task. Where the stakes are higher, the machine's role is smaller.

• A reply to a question. This is where the machine is strongest, and it can safely write the whole thing. Check one thing only: that it has not promised something you did not promise.
• An awkward message. Ask for three versions — a direct one, a gentler one, and one that also offers a solution — but let the final wording be yours. The register of bad news is what people remember for years.
• A reminder. The machine does these well, precisely because there is no irritation in it. Supply one thing: which reminder this is. Everything else follows from that.
• Cold outreach. Here the benefit is smallest and the risk highest. If the message reads as a template, reply rates fall, and template-flatness is exactly the default of machine text. Write this type by hand and ask for help only with variants.

What gives it away anyway

Five signs stand out immediately on the receiving end, and all five can be fixed in thirty seconds.

• The empty opening sentence. "Thank you for getting in touch", "I hope this finds you well." Cut it; the email starts where the first real sentence is.
• The too-complete answer. A machine reply answers every point raised, with equal weight. A human reply picks two and disposes of the rest in one sentence.
• The missing small thing. A real email contains half a sentence that concerns only the two of you. Without it the message is polite and empty.
• The perfect length. Every paragraph is three sentences. A real email has a one-sentence paragraph in it somewhere.
• The machine sign-off. "Please do not hesitate to contact me should you have any questions." If that is not your usual close, swap it for whatever is.

Do you have to say a machine wrote it?

Legally, no. The EU labelling duty applies to published text aimed at informing the public on matters of public interest — a work email is not that. The details are a post of their own: when AI use has to be disclosed, and the answer is narrower than most people assume.

It is worth knowing what the research measures, though. Across thirteen experiments with more than five thousand participants, disclosing AI use consistently reduced trust — by roughly 20 percent among clients and 16 percent among students — and the effect held even among people who use AI frequently themselves. But the same research found the more important thing: trust falls further when someone else exposes you rather than when you say so yourself. Between those two, concealment is not the winning strategy.

So the usable answer is not about disclosure but about content. If every promise, figure and decision in the message is yours and the machine only assembled the sentences, there is nothing to declare — no more than you declare a spellchecker. If, on the other hand, the email asserts something you could not have written on your own, what is missing from it is not a label but a check.

What to take away

Email is the genre where the machine measurably wins — provided you give it three things: the actual thread, the background only you know, and three of your own past emails. Ask for two versions, rewrite the first and last sentence, and check the dates and the promises. Write cold outreach by hand. And do not agonise over disclosure — agonise over whether every claim in the message is yours.