A regular customer messages you with a question. You ask AI to draft the reply, to save time. It hands back an email that's spelled right, polite enough, ready to send. But reading it over, something grates: it addresses someone you've worked with for three years like a stranger who just filled in a contact form. "Dear valued customer," "Please do not hesitate to reach out"... correct, and cold. Send this, and they'll notice you're different today.
01Polite is its safe default — not what you needed
This isn't a writing problem. That email is well written — it's just written for the wrong person.
When you say "reply to this email," AI has no way of knowing how close you and the recipient are. It can't see the three years you've worked together, doesn't know you two trade jokes, doesn't know you've always used their first name and never "valued customer." Without that, it does the safest thing a machine knows: maximally formal, maximally polite, maximally distant. Because with a stranger, too polite beats too familiar — it's the move that's never wrong, so it always picks it.
The catch is that your relationship isn't a stranger relationship. The machine's "safe," dropped into something already warm, comes out as distance. It didn't write badly. It wrote for a version of the relationship that doesn't exist.
02Four signs an email sounds like a machine
The hard part is that emails like this look fine — easy to send without noticing. These signs help you catch it in the act:
Notice none of the four are writing errors — they're relationship errors. The email gets everything right except the one thing that matters most: it doesn't sound like you talking to that person.
03Let it hear how you actually talk to this person
The worst fix is typing "make it warmer and friendlier." Those adjectives are vague — "friendly" to a machine is an averaged notion, not your voice. Voice is hard to describe in words and easy to show by example.
"Here are a few emails I've sent to this kind of customer. Match this voice and level of formality." A sample says more than any adjective.
"This is a regular of three years — relaxed, first-name, no 'valued customer.'" Give it the thing it can't guess.
No need to re-paste later — save those samples in one place and just point to them. Your voice becomes reusable.
Two or three old emails in "the right voice" pull its target from "polite for everyone" to "sounds like you talking to this person" — and the second is much closer to you.
Notice this isn't teaching AI to write better. It already writes smoothly. You're teaching it the one thing it could never know on its own: what this relationship sounds like. That's your job, because it lives in your head and in the history between you two, not in the machine.
So next time an AI email sounds like a machine trying to be polite, don't jump to "AI is tone-deaf." Ask instead: did I ever let it hear how I actually talk? Almost always, the answer is no. This is also the doorway to a bigger skill in the craft: giving a structured instruction to steer the voice, and choosing what to put in for the agent to anchor on. That cold email was never the AI being tone-deaf — it just never heard you talk.