Yes, an AI can write emails in your voice, not because it knows you, but because you hand it material about you. The difference isn't the model, it's what you show it: type "casual but professional" into the prompt and you get the average of millions of other people's emails back. Analyze thirty of your own sent emails and turn them into concrete rules, and you get drafts that sound like you instead of like the average. Your tone is made of six layers, all of which can be written down: greeting and sign-off, sentence length and paragraph shape, where your core message sits, three to five phrases that keep showing up in your writing, the list of things you would never write, and how you handle humor and emojis. The no-go list matters most, because generic text gives itself away through what it has too much of, not through what's missing. At the end there is a blind test: it shows you which rule is still missing before you trust the first real email. And approval stays with you: an email in your voice is a draft, not a decision.
In my business, Sonja handles this. She's my AI assistant for business email, and her sense of my tone doesn't come from a questionnaire, it comes from 50 of my own real sent emails. The style guide that fell out of it is an unspectacular table: which greeting belongs to which contact group, how I sign off, what I explicitly don't want. How I build my roles in general lives on the AI employees page, and the difference between an assistant with a voice and a chatbot is in Build an AI Assistant: Three Paths.
What your own tone is actually made of
Most people treat their writing style as a feeling. It's really a set of decisions you make the same way every time, and that is why it can be written down. Six layers are enough, and the examples on the right come from my own guide:
The fifth layer is the one most often overlooked, and it makes the biggest difference. A model writes politely, evenly and completely all by itself. Your tone only appears where you refuse to write the obvious thing.
On top of that: your tone isn't one tone, it's two or three. I write to prospects differently than to my tax advisor, and each is consistent in itself.
Why "casual but professional" gets you nowhere
An adjective is a shortcut. You know what you mean by it because you have your own emails in mind. The model doesn't, so it fills the gap with the average of the internet. That's why each of these adjectives comes back as something other than expected:
Examples solve the problem because they carry all six layers at once, without you having to name a single one. Anthropic's own prompting guide says examples are "one of the most reliable ways to steer Claude's output format, tone, and structure", recommends three to five of them, and names three criteria: close to your real use case, diverse enough that no unintended pattern sticks, and clearly separated from the rest of the instruction (platform.claude.com).
For email that means three to five real, complete messages from different situations. A yes, a no, a follow-up question. Not the same appointment confirmation five times over, or the model will mistake your confirmation voice for your whole style.
Six steps to make your tone describable
- Pull 30 to 50 of your own sent emails from the past few weeks. Sent, not received: this is about you, not about the people writing to you. Before any of them goes into an AI tool, replace names, addresses, amounts and transactions with placeholders: you do not need them for the tone, and anything specially protected stays out entirely.
- Sort out what isn't you: automatic calendar replies, one-word answers, forwards with no text of your own. In my own analysis, about 30 of 50 messages were genuinely written by me.
- Split the rest into two or three contact groups. Mine: new prospects, ongoing project clients, and more formal contacts like my tax advisor or insurer.
- Walk the six layers per group and write one sentence per layer. No adjectives, evidence: the greeting you actually used, not the one you consider appropriate. The gap between the two is often the real find.
- Write the no-go list. Fastest route is other people's emails that annoy you: every "I would never write that" is a line.
- Run a blind test before you send anything. Generate three drafts, mix in two of your real old emails, and let someone guess which ones are yours. Whoever spots the drafts can usually say why, and that reason is the rule still missing from your guide. If you have nobody in the business for that, bring the round into my community, where people have already run the same test.
One thing does not belong in the guide: your typos. They aren't style, they're haste, and the one point where an email from Sonja beats one of mine.
Before and after: turning a generic email into yours
The setup: a prospect got a demo a week ago and hasn't replied. The draft when the prompt says nothing more than "write a friendly follow-up email":
Subject: Follow-up on our proposal
To whom it may concern,
I hope this message finds you well. I wanted to follow up once more
regarding our proposal from last week and kindly ask whether you have
already had the opportunity to take a look at it.
Of course I remain available for any questions and would very much
appreciate a positive response.
Kind regards
Kevin Welter
And here's the same email with the style guide in the prompt:
Subject: Quick nudge about the demo
Hi Jana,
quick nudge: the demo of your new homepage has been sitting in your
inbox since Monday.
I have to be honest: I don't know whether you haven't looked yet or
whether things are simply too busy right now. Either is completely
fine.
If you do want a look, it takes two minutes:
- open the demo on your phone
- scroll down to the contact form once
- ask yourself whether you'd send an enquiry there
Then just tell me yes or no. And if it's not for you: the pointers in
the demo are yours to keep anyway.
Best
Kevin
Both emails want the same thing. The difference isn't a matter of taste, it traces back to the guide line by line:
The second to last line makes the biggest difference for me. An email that lets the other person say no and keep something anyway sounds like a human. An email hoping for "a positive response" sounds like a process.
How to spot an AI email right away
Eight items sit on my no-go list. Take them over and the most obvious traces disappear in one go:
- The opener. "I hope this message finds you well." No email of mine starts like that.
- Evenness. Every sentence the same length, every paragraph the same size. Humans write unevenly, with a four-word sentence in the middle.
- The rule of three. "faster, clearer and more reliable." Three adjectives where two would have done.
- The not-only-but-also. "It's not just about looks, it's about trust." Once, that's a good line. Three times in one email, it's a fingerprint.
- The dash as favorite punctuation mark. It does not appear in my business correspondence; in AI drafts it is everywhere.
- Warmth with no cause. Enthusiasm attached to nothing concrete: "That sounds like an exciting project!"
- No concrete next step. "Feel free to reach out anytime" instead of a date, a deadline, or a yes-no question.
- The closing summary that only repeats what stood two paragraphs above.
Items 1 to 6 are easy to delete. Item 7 is the expensive one, because it doesn't read as AI, it reads as indifference: an email without a next step gets no reply from a human either.
Where the style guide belongs so it actually works
A guide in your notes app changes nothing, because you'd have to paste it in every time. It has to sit where it gets read automatically:
- A dedicated working folder for the role. The guide and example emails live there as normal files alongside the personnel file. That lets the same role work with Claude Code, Codex or an open model. Anyone who wants to start entirely inside the Claude app can reproduce the role principle in a Project: How to Use Claude Projects Right.
- Account-wide instructions for whatever should apply everywhere. Anthropic describes these as "account-wide settings" that Claude should consider in its responses (support.claude.com). Two or three sentences belong there, your sign-off and your most important no-go.
- Memory for what emerges as you work together. On Free, Pro and Max, Claude's memory is on by default according to Anthropic (support.claude.com). How to check and correct what's stored there is in Claude Memory: Check It and Use It.
If several people are supposed to write in the same voice, the guide belongs in one central place rather than in every individual account: what that costs is in Claude Team: Cost and When It Pays Off. The building blocks for such a role are in Claude Prompts: Templates, including the approval rule that the next section is really about.
What an AI in your voice still isn't allowed to do
Tone is one half, content is the other, and the second half is where it gets expensive. Three limits are written into Sonja's file word for word:
- She never invents the message. She improves language, tone and form, not the content and not the price.
- She never guesses. Name, address, link and price come from the source file. If something is missing, she asks.
- She doesn't send. Every text sits in my inbox as a draft. "Looks good" is not approval; it takes a clear sentence from me, logged word for word with a timestamp.
What that looks like in practice: in a campaign of 15 new-client emails in July 2026, I had already sent eight myself. The remaining seven drafts were in my inbox within two minutes, every demo link in them checked for reachability the same day, and one issue went back to my website colleague: one template still ended with the placeholder for the sender. I did the sending myself. The whole chain is on Sonja, who writes the email in my voice.
What else belongs to the anonymizing from step 1 is in AI and Privacy: What the AI Gets to See.
Common questions
Can an AI write emails that don't sound like AI?
Yes, if you give it examples and a no-go list instead of adjectives. Ban the patterns from the list above explicitly and most of it disappears before you even start on tone.
How many example emails does an AI need for my style?
Anthropic recommends three to five examples in the prompt, diverse enough that no unintended pattern sticks. For the analysis itself, take more: I went through 50 sent emails, of which roughly 30 were genuinely written by me. The guide comes out of those; the three to five examples sit next to it.
Can I have an AI write emails without reading them first?
Technically yes, in my business never. Tone can be hit reliably, content can't: commitments, dates, prices and deadlines are decisions, and those are yours.
Does this work in Outlook?
For me, yes. Sonja connects to my Microsoft 365 mailbox via device-code login, with no password in a file, and drops drafts where I read them anyway. That is my own setup, not a ready-made product feature. Alongside it, Anthropic offers a dedicated connector for Microsoft 365 (support.claude.com). The route via a project and copy-paste works just as well; what off-the-shelf AI email tools cover of this and what they don't is in its own post.
How to take it further
Take your last 30 sent emails and put down six lines, one per layer. Then the blind test: three drafts, two real old emails, somebody else guesses. Whatever gives them away becomes line seven.
Sonja exists with exactly this structure as a ready-made package in my community Claude Practitioners, alongside the other AI employees and the course that goes with her.
