Skip to content

Writing e-mails with AI: friendly, clear, fast

By the KI können editorial team9 min read

Written with AI and checked against the sources by a second AI. As of November 16, 2026. More in the notes on AI.

The e-mails that take the most time aren’t the long ones but the tricky ones: a no, the third follow-up, the reply to an angry message. An AI helps you get a calm, clear draft for them in a few minutes. What the e-mail says is still up to you.

A woman types on a laptop at a standing desk, a mug and her phone beside her, a clock on the wall, a drawer unit with papers and a plant nearby.

Key points

  • For a good e-mail, the AI needs five pieces of information: occasion, recipient, goal, facts, and tone and length.
  • Write down your tone once as rules and include them with every e-mail.
  • With difficult e-mails, have the AI understand first, then write: what does the other side really want?
  • Replace names, addresses, contract and customer numbers with placeholders before anything goes into the chat.
  • Before sending, check commitments, dates and tone yourself, ideally by reading the e-mail aloud.

Which e-mails an AI can take off your hands

An AI phrases things quickly, kindly and without a bad mood. That helps with exactly the e-mails you keep putting off. But it knows neither the matter nor the person you’re writing to. What it needs to know goes into your instruction, the text you give it.

Tricky e-mails at a glance
E-mailWhat goes into the instructionWhat you check at the end
Refusal or noReason, alternative, what remains openWhether the no is clear and doesn’t sound like a maybe
Follow-up, for the second timeWhat is missing, since when, new deadlineWhether the tone is firm but not offended
Reply to an angry messageThe accusations without names, the factsCommitments, apology, question of fault
Admitting a mistakeWhat happened, what you are doing nowWhether nothing is glossed over or exaggerated
Replying to a long threadThe thread without names and signaturesWhether every question has an answer

In the trial lesson you put your first instruction together and get feedback. No account, in a few minutes. Start the trial lesson

The instruction for an e-mail: five pieces of information

“Write a friendly e-mail to the customer” leads to a smooth text that fits everyone and no one. It gets better with five pieces of information that you can jot down in a minute.

  1. Occasion: What is it about, in one sentence?
  2. Recipient: Who will read this, and how well do you know each other? First-name or formal terms?
  3. Goal: What should the person do, know or decide after reading?
  4. Facts: Dates, amounts, numbers, commitments, exactly as they stand.
  5. Tone and length: for example “matter-of-fact and warm, 120 words at most”.
Sample instruction
Write an e-mail to [recipient, e.g. “a supplier we have worked well with for years, we’re on formal terms”].
Occasion: [one sentence].
Goal: After reading, the person should [do what exactly].
Facts that must be right: [dates, amounts, numbers as placeholders].
Tone: matter-of-fact and friendly, no phrases like “I hope this e-mail finds you well”.
Length: 120 words at most, subject line with eight words at most.
Don’t add anything. If a piece of information is missing, write [missing: …] in the text.

The last sentence is the most important one. Without it, the AI fills gaps with details that sound plausible: a delivery date, a callback number, a discount.

Write down your tone once

Hana is executive assistant to the management of a furniture factory. At first, her drafts from the AI all sounded the same: polite, long and a little stiff. Sending her own e-mails along as models only helped halfway, because the AI picked up different features each time. So Hana turned her e-mails into fixed rules and adds them to every instruction.

Sample instruction
Below are three e-mails I wrote that were well received. Names and companies have been replaced with placeholders.
Describe my writing style in six rules at most that I can give an AI: sentence length, greeting, sign-off, typical phrases, what I avoid.
For each rule, give a short example from my e-mails.
E-mails:
[three of your own e-mails]

Hana then reads the rules herself and deletes whatever isn’t true. One of them now says: “The first sentence says what it’s about. No thanks for the e-mail unless it was something special.” Since then she has much less to rework.

Saying no without slamming doors

Marcus is a master electrician with four people in his business. A property management company asks whether he can still inspect three apartments next week. He can’t, but he wants to keep the client. His problem: until now, his refusals sounded either curt or so vague that they called again.

Sample instruction
Write a refusal to a property management company we work with regularly (we’re on formal terms).
They are asking about inspecting three apartments next week.
The no must come in the first or second sentence and be unambiguous.
Offer: from [date] we have two days free, or I can recommend a fellow business if they wish. Don’t name anyone.
No chains of apologies, no reasons beyond our workload.
90 words at most.

Marcus only gives a name once the property management company asks and he has reached his colleague. That’s why the e-mail only offers the recommendation; without the last sentence in the instruction, the AI might have named an invented business.

Following up for the second time

Diego is a dispatcher at a freight forwarding company. A subcontractor isn’t sending the signed delivery notes, even though he has already been reminded once. Diego doesn’t want an argument, but he wants the papers by Friday.

Sample instruction
Write a second reminder to a subcontractor (we’re on first-name terms).
Missing: signed delivery notes for [number] runs from [period]. First request on [date], no answer.
New deadline: Friday, 12 noon. After that I have to put the billing for these runs on hold; this has been agreed with our accounting department.
Tone: firm and collegial, no accusation, no threat beyond this consequence.
Put the deadline and the consequence in a sentence each.

The consequence in the instruction is true because Diego has agreed on it. Otherwise an AI likes to invent one that sounds good, such as “legal action”. You only write sentences like that if you mean them and are allowed to follow through.

Replying to an angry e-mail: understand first, then write

Leah works as an insurance clerk. A customer writes a long, angry e-mail because his claim has been sitting for weeks. Whoever answers such an e-mail right away often writes against the tone instead of addressing the questions. So Leah has the AI work in two steps. She uses the AI her insurer has approved, because even without names, the account of a claim often contains personal details. She still removes names, address, contract number and claim number.

Sample instruction
Step 1: Read the e-mail below. List:
a) What questions does the person ask, including indirectly?
b) What accusations does the person make?
c) What does the person want to achieve in the end?
Don’t write a reply yet.
E-mail:
[e-mail without names and numbers]
Sample instruction
Step 2: Now write a reply.
Facts: The documents have been complete since [date]. The loss adjuster will get in touch by [date]. After that we decide within [period].
Acknowledge the anger in one sentence without repeating it.
Answer every question from step 1 or say when it will be answered.
Make no commitment about the amount of the benefit and no admission of fault.
150 words at most.

Whether an apology is appropriate and how far it goes is something Leah decides with her team lead, not the AI. Especially when money or liability is involved: better one sentence less than a commitment nobody can keep. Fixed reply templates for frequent complaints are described in the guide AI in sales and customer service.

Long threads: don’t miss anything

After twelve e-mails between five people, often nobody knows which question is still open. Here an AI helps with sorting. Delete signatures, names and phone numbers first, or replace them with roles such as “purchasing” or “trade fair booth construction”.

Sample instruction
Below is an e-mail thread. Turn it into a table with the columns: question or request, from whom (role), answered yes/no, open with whom.
Include only what is in the thread. If it’s unclear whether something was answered, write “unclear”.
Then write a draft that answers all points open with me in one e-mail; where I’m missing the answer, put [answer: …].

You check the table against the thread, not against your memory. With “unclear” in particular, it pays to look at the original.

What doesn’t belong in the chat

Almost every e-mail contains details about people: names, addresses, numbers, sometimes health or money. A contract or customer number counts too, because it leads to a person. Such details don’t belong in a private AI account. Replace them with placeholders such as [customer] or [no.] and only fill in the real details in your e-mail program.

The e-mail program itself often has a writing aid built in. It usually sees the whole e-mail and sometimes the thread. Whether you may use it at work, and with which account, is up to your employer; in your own business you decide yourself, ideally once, as a general rule and in writing.

Six questions before you hit send

  • Do all dates, amounts, names and numbers match the matter?
  • Does it contain anything you can’t or may not commit to?
  • Is the first sentence the most important one?
  • Would you say it like that on the phone too?
  • Have all placeholders and every [missing: …] been replaced?
  • Is the greeting right: first name or last name, formal or familiar?

Read tricky e-mails half aloud once. Whatever stumbles or sounds unfriendly when spoken, you cut. With very tricky e-mails, sleeping on it helps more than any wording.

Keep learning with your own e-mails

With every e-mail, your instruction gets a little better: one more rule for the tone, one less remark you have to delete afterward. If you want to practice this on purpose, the learning app “KI können” has bite-sized lessons for exactly that: write an instruction, get the result from the AI you already use, and check it point by point. You build your first instruction right away in the trial lesson.

Sources

Information as of November 16, 2026. Laws and offers can change; this article is not legal advice.