The reason an AI-drafted email gets bounced for âtone is offâ is that âprofessionalâ isnât one tone. Itâs at least four. The note you write to a board chair confirming a date is not the note you write to a vendor who blew a deadline, which is not the note you write to your execâs direct report whoâs having a hard week. A model that doesnât know which register you want produces a generic-warm middle that fits none of them.
The fix is to tell it. Once you set up a register matrix, every drafting request takes the same shape, and the output stops needing an editorial second pass.
The four registers most EA correspondence falls into
Write these once, in your own words, with two or three example sentences for each.
- Board / executive peer: short, no apology language, no extraneous warmth, signs off with name only. Stakes are credibility.
- Client (active): warm but efficient, leads with whatâs resolved, ends with a clear next action. Stakes are the relationship.
- Internal team: friendlier, contractions fine, can use a small piece of context or humour. Stakes are clarity and morale.
- Difficult / corrective: precise, factual, no euphemism, no excessive softening. Stakes are that the recipient must understand the situation accurately.
The third bullet for each is your example sentences. Not generic, drawn from your actual sent folder. Thatâs what gives the model something to imitate.
The drafting prompt
Open ChatGPT, Claude, or Microsoft Copilot and use this exact shape:
âRegister matrix: [PASTE FOUR-REGISTER LIST WITH EXAMPLES]
Situation: [WHATâS HAPPENED, OR PASTE THE INCOMING EMAIL] Recipient and context: [WHO THEY ARE, RELATIONSHIP HISTORY, ANY POLITICAL NOTE] Goal of this email: [E.G. RESCHEDULE WITHOUT LOSING TRUST, ACKNOWLEDGE A COMPLAINT WITHOUT ADMITTING FAULT, INTRODUCE TWO PARTIES] Register to use: [BOARD / CLIENT / INTERNAL / DIFFICULT]
Constraints: under 130 words, no âI hope this finds you wellâ, no âdonât hesitate to reach outâ, no apologies that imply fault we havenât admitted to.
Draft three versions. Mark each: âVersion A: warmerâ, âVersion B: neutralâ, âVersion C: firmerâ. Donât pick for me; just produce the three.â
The three-versions ask is the key. Itâs faster than re-prompting for tweaks, and seeing the spread tells you which version actually matches the moment. Often the one youâd never have asked for (âfirmerâ) turns out to be the right one.
A worked rewrite, before and after
Suppose your exec needs to push a board call from Tuesday to Thursday because of a travel delay. Default-prompt output:
âDear Board, I hope this message finds you well. I wanted to reach out regarding our scheduled call on Tuesday. Unfortunately, due to unforeseen travel circumstances, Iâd like to kindly request that we move the meeting to Thursday at the same time. I sincerely apologise for any inconvenience this may cause and look forward to your understanding. Please donât hesitate to reach out if Thursday doesnât work for your schedules.â
Every AI tell is in there. Three apology phrases, one âreach outâ, one âI hope this findsâ, and 70 words spent on cushioning. Now run the matrix prompt with register set to âboardâ and constraints set as above. You get:
âTravel delay on my end. Moving Tuesdayâs call to Thursday, same time. Holding the agenda as-is. Reply if Thursday is a problem; otherwise see you then. [NAME]â
Three sentences, 27 words. Reads like an executive wrote it, because thatâs the register the matrix asked for. The board does not need a paragraph of softening; they need to know the new time and that nothing else is changing.
The pre-send pass for high-stakes notes
Before any email going to a client, board member, or anyone you cannot afford to misread, run one more prompt:
âHereâs the draft. Read it as the recipient. Answer:
- Whatâs the first impression in the first sentence?
- Is there anything that could be misread as defensive, dismissive, or evasive?
- Are there any soft-pedal phrases that hide a hard fact (e.g. âa slight delayâ for a two-week slip)?
- Is there a commitment in here that I shouldnât make without checking with [EXEC]?
Be specific. Quote the exact phrase youâre flagging.â
That last point is the safety net. Models sometimes draft confident-sounding commitments (âweâll have the deck to you Fridayâ) on your execâs behalf that havenât been confirmed. Catching them at draft stage is free. Catching them after send is a phone call you donât want to make.
The failure mode thatâs hard to spot
Over-cushioning. Models trained on the polite end of internet prose default to apologetic. A client email that thanks the recipient four times in 100 words doesnât read as warm, it reads as anxious. Your execâs voice probably isnât anxious. The pre-send pass catches some of it; an explicit constraint in the drafting prompt (âat most one thank-you in the whole emailâ) catches the rest.
The other one to watch: hedging modifiers. âJustâ, âperhapsâ, âif you have a momentâ, âwhen you get a chanceâ. A board email with three of these reads as someone who isnât sure theyâre supposed to be talking to the board. Strip them.
A counterintuitive observation
The shortest version is almost always the strongest, and almost never the first version a model produces. AI defaults to longer because longer feels safer (more polite, more thorough). Executives write short because their time is the scarce resource and they trust the recipient to fill in the rest. If you want the email to sound like your exec, ask the model for a version under 60 words after it gives you the first draft, and pick from there.
What the model is actually doing for you
Cutting drafting time from 12 minutes to 90 seconds for routine notes is real. The bigger gain is consistency. The fifth board email of the week sounds like the first, even at 5pm on a Friday, because the register matrix did the same job both times. Your execâs voice is now reproducible across volume, which means you can handle correspondence at twice the rate you used to without anyone feeling your bandwidth has stretched.
Polished doesnât mean longer or warmer. It means precisely calibrated to the recipient and the moment. AI is a tool for hitting that calibration on the first try, if youâve told it what calibration to hit.