The reason most AI-drafted emails get sent back for “another pass” isn’t a grammar problem. It’s a voice problem. A generic ChatGPT reply sounds like a generic ChatGPT reply, and your CFO can spot one from across the room. The fix isn’t a longer prompt. It’s a tone library, built once, that you paste into every future request.
The Anthropic Economic Index shows that augmentation (working alongside AI) has overtaken automation (handing it off entirely) as the dominant pattern in real workplaces. The administrative assistants getting clean output aren’t typing one-off instructions. They’re feeding the model their own past work as a reference and asking it to match.
Build the voice card before you ask for anything
Pull six or seven emails you’ve sent on behalf of your executive that you were happy with. Mix of tones: a confirmation, a polite decline, a follow-up, an introduction, a delicate reschedule. Paste them into ChatGPT or Claude with this:
“Below are six emails I’ve written on behalf of my executive [NAME, ROLE, INDUSTRY]. Read them as a corpus and write a one-paragraph ‘voice card’ I can paste into future drafting prompts. Cover: average sentence length, formality level, signature opener and closer phrasing, words this writer favours, words this writer avoids, and how they handle bad news. Don’t editorialise. Reflect what’s actually on the page.
[PASTE 6–7 EMAILS]”
Save the output in a sticky note, a Notion page, or the top of a draft email you never send. That paragraph is now your tone reference. You’ll paste it into every drafting request from now on, and the difference is the difference between an email that gets sent and one that gets rewritten.
The drafting prompt that uses the voice card
Now compose a real reply. The prompt looks like this:
“Voice card: [PASTE THE PARAGRAPH FROM ABOVE].
Situation: [WHAT THE INCOMING EMAIL SAYS, OR PASTE THE THREAD].
Goal of my reply: [E.G. CONFIRM THE NEW DELIVERY DATE, ACKNOWLEDGE THE DELAY, KEEP THE RELATIONSHIP WARM].
Constraints: under 120 words, no apologies that imply fault, no bullet lists, sign off with [NAME] only.
Write three short variants. Variant 1: warm and reassuring. Variant 2: brisk and businesslike. Variant 3: deferential. Mark which you’d use if the recipient is a long-term client vs a new vendor.”
Three variants in one go costs you nothing and gives you the option to pick the tone that fits the moment. The “mark which you’d use” line forces the model to think about register, which is the part of writing it usually mishandles.
A worked example, before and after
Suppose a supplier has just emailed about a delayed shipment. A first-pass prompt gives you something like:
“Thank you for letting us know about the delay. We appreciate your communication and look forward to receiving the shipment on the new date. Please don’t hesitate to reach out if there are any further updates.”
Polite, but it sounds like every other AI reply on the planet. With a voice card that says “this writer avoids ‘don’t hesitate’, favours short sentences, and never thanks vendors for telling them about problems”, the same situation produces:
“Understood, the new date works on our end. Send the updated tracking through when it’s available and we’ll loop in receiving. Thanks, [NAME].”
That’s the version that goes out unedited. Same model, same email, completely different result. The voice card did the work.
The AI tells to scrub before sending
Even with a voice card, models drift. Run a 30-second pre-send pass on anything important. The phrases to delete on sight:
- “I hope this email finds you well.” Nobody under 50 has written this voluntarily in a decade.
- “I wanted to reach out to…” (just reach out)
- “Please don’t hesitate to…” (the hesitation phrase is a dead giveaway)
- “As we discussed…” when there was no prior discussion
- Three-clause opening sentences that warm up before they say anything
- Em dashes used as hedges (commas do the same job without the AI tell)
If you spot two or more of those in a draft, either rewrite manually or send it back to the model with: “Rewrite this without any of: [PASTE LIST]. Match the voice card.”
The failure mode nobody warns you about
The trap is feeding the model the wrong examples. If you build your voice card from the polished, formal, treaty-grade emails you write to the board, every casual internal note you draft from then on will sound stiff. Build two voice cards: one for external/executive-level correspondence, one for internal team communication. Label them. Paste the right one for the right job.
The other failure mode is treating the voice card as final. Update it every quarter or after any feedback round with your executive. Tone preferences shift, especially when leadership changes or the company does. A voice card built six months ago for a recently-departed CEO will quietly poison every email you send.
What this actually saves you
The original two and a half hours a day spent on email isn’t going to drop to twenty minutes overnight. What it does drop is the rewrite cycle: the version you draft, second-guess, rewrite, and second-guess again. With a voice card and a three-variant request, you make one call instead of four, and the version you pick is usually right the first time.
Writing is the part of administrative work AI is genuinely good at, but only if you give it your voice to imitate. Without that, you’re getting an averaged-internet version of professional, which isn’t what your executive needs to send.