AI Email Drafts That Sound Like You, Not a Stranger
Email is the most obvious thing to hand to AI. It’s writing, there’s a lot of it, and most of it isn’t precious. So people try it, and within about a week most of them quietly stop — because the drafts come back polished, professional, and not theirs. Too many words. Too warm. “I hope this message finds you well.” A closing paragraph that restates the whole email back to the reader.
The productivity research says this should be working. In a well-known MIT experiment, professionals given ChatGPT finished mid-level writing tasks 40% faster and scored 18% higher on quality. That’s a real result from a real experiment. But read the fine print and you find the catch: the researchers deliberately chose tasks that “didn’t require precise factual accuracy or context about things like a company’s goals or a customer’s preferences.”
Which is to say — they studied writing that has no history, no relationship, and no stakes. Your inbox is nothing but history, relationships, and stakes. That’s the whole gap. Here’s how to close it.
Why the drafts sound wrong
The model isn’t trying to write your email. Absent other instructions, it writes the statistical average of every business email ever published — which skews longer, more formal, and considerably more eager than how any actual person writes to a colleague they’ve known for three years.
It’s also pulling toward The tendency of AI models to be agreeable, complimentary, and eager to please — thanking you unnecessarily, praising the question, softening anything that might read as blunt. It comes from how models are trained on human feedback: people rated pleasant responses highly, so the model learned to be pleasant by default, sometimes past the point of sounding sincere. , which is why the drafts open with “Thank you so much for reaching out!” when your actual reply would have started with “Yep — Thursday works.”
You can’t fix this by asking for “a more casual tone.” Casual is a direction, not a target. You have to show it the target.
Step 1: Make it name your pattern
Don’t describe your voice. You’ll get it wrong — almost nobody can accurately describe their own writing. Instead, go into your sent folder and copy three emails you actually wrote: ideally one quick logistical reply, one where you delivered mildly bad news, and one where you explained something to someone junior. Then:
Below are three emails I wrote. Describe my writing style specifically enough that someone could imitate it. Cover: typical length, how I open and close, sentence rhythm, formality, how direct I am, punctuation habits, and anything I conspicuously don’t do. Be blunt — I want the real pattern, not a flattering one.
This is Giving a model a handful of examples of what you want instead of describing it. Models are far better at pattern-matching from samples than at following adjectives — three real emails carry more usable signal than a paragraph of instructions about tone. pointed at yourself, and the output is usually a small shock. Mine came back with “opens without a greeting when replying in a thread” and “uses dashes where most people would start a new sentence.” Both true. Neither something I’d have thought to say.
Step 2: Keep the description, not the samples
Read what it produced and correct it. It’ll overstate something — the description of my style claimed I was “consistently informal,” which isn’t true when I’m writing to a client. Edit it directly, add the exceptions, and save that paragraph. That description is the reusable asset, not the three emails.
Where you save it depends on your tool: custom instructions, a saved project, a stored persona. If you want it available every time without re-pasting, this is exactly the setup covered in Build a Reusable AI Assistant for the Task You Do 20 Times a Week.
Step 3: Give it the two things it cannot guess
Now the part that closes the MIT gap. When you ask for a draft, the model already has your voice. What it doesn’t have is your situation. Supply two things, always:
The relationship and history. “This is a vendor we’re likely to stop using next quarter.” “She’s asked about this twice already and I haven’t replied.” “He’s new and doesn’t know the acronyms yet.”
What you want to happen next. Not the topic — the outcome. “I want to buy two more weeks without sounding like I’m stalling.” “I want to say no and keep the door open.” “I want a decision on the call, not another thread.”
So the whole request looks like:
Draft a reply to the email below. Use my saved writing style. Context: she’s asked twice, I dropped it, and she’s right to be annoyed. Goal: acknowledge the delay without over-apologizing, give her the real date, and make it easy for her to hold me to it.
That prompt produces something you’d actually send. “Write a professional reply to this email” never will — and the difference isn’t prompt cleverness, it’s that you gave it the two facts only you had.
When to skip it entirely
Some email isn’t worth the round trip. If your reply is under about three sentences, just write it — the time you’d spend on context is longer than the email. And for anything carrying real emotional weight (condolences, an apology that’s genuinely yours to make, praise you want someone to feel), write it yourself. Not for ethical reasons — because a person who knows you will hear the seam, and the message will land worse than three clumsy sentences you meant.
If you’re wondering whether to mention AI helped, that’s a separate question with a decent answer: Do You Have to Tell People You Used AI?.
The takeaway
The 40% time saving is real, but it was measured on writing with no context and no relationships — and your inbox is made of exactly those two things. Close the gap in two moves: show the model three emails you actually wrote and have it extract your pattern, then feed it the relationship and the outcome every time you ask. Ten minutes to set up, once. After that you’re editing your own voice instead of translating out of someone else’s.
Sources: The 40% time reduction and 18% quality improvement come from Shakked Noy and Whitney Zhang’s experiment with 453 college-educated professionals, published in Science in July 2023 and summarized by MIT News. The caveat about task design is the researchers’ own — they noted the assignments “didn’t require precise factual accuracy or context,” and that accuracy remains “a major problem” for real-world use. The argument that this caveat is precisely what makes AI email feel wrong is mine, not theirs.
Related: Build a Reusable AI Assistant for the Task You Do 20 Times a Week for saving your style description so you stop re-pasting it, and Giving AI Context: The Skill That Actually Makes It Useful on why supplying context — not clever wording — is what separates useful output from generic output.