A reply sounds like you when it gets four things right. How long it is. How close it stands to the reader. What it commits you to. How it ends. All four are habits rather than personality, which is why software can hold them, and why the drafts that miss tend to miss the same way.
- The part of your voice that transfers is smaller and duller than it feels. Sentence length, greeting, contractions, sign-off, how hard you commit to a date. Knowing what to say stays with you.
- A general draft misses in a predictable direction. With nothing to go on, a draft writes to the middle of business email. That middle is longer, more formal and more eager than almost anyone actually is.
- You have four or five voices, keyed to who you’re writing to. An instruction with an audience named on it is the one that lands.
- The test is how much of a draft survives to sent, and whether you make the same change twice.
- A draft in your own register is the one you skim. That’s the real hazard here, and it’s about accuracy rather than tone.
What your voice is actually made of
The best known work on spotting an author from their writing counted small words. In 1963 the Journal of the American Statistical Association published Inference in an Authorship Problem by Frederick Mosteller and David Wallace, who settled the contested Federalist Papers by counting words. The dull ones. Terms like war, executive and legislature moved with whatever the essay was about, so they tracked the subject rather than the man. The filler of the language, an, of, upon, held rates steady enough per author to decide the question.
That’s a deflating result if you think your voice is your judgment. It’s a useful one if you want software to reproduce it. What marks a message as yours is the several thousand unconsidered choices under the sentence you were pleased with. Unconsidered choices stay stable, and that’s the property that makes them reproducible at all.
For email, they come down to about seven.
Length. How many words you spend on a routine answer, and whether you write in paragraphs or in lines. People notice this first in somebody else’s draft, and it’s the hardest one to state about their own.
Distance. Formal, plain or familiar. First name, full name, or straight into the answer. Contractions or the full words. Whether you thank people for writing or take that as read.
The opening move. Some people answer the question in the first line. Some acknowledge the situation for a sentence and then answer. Some ask about the family. Whichever you do, you do it almost every time.
How you commit. The gap between “Thursday” and “we should be able to get that across to you by the end of the week” is rope you leave yourself. It’s one of the most individual things about a working writer.
How you deliver friction. Saying no, quoting a higher number, asking for the same document a third time. Everybody has a fixed way of doing this, and almost everybody keeps it in their head.
The sign-off and what travels with it. Best, Thanks, Kind regards, or straight to your name. Full name, initials, first name only. Cell number on client mail and your name alone on the rest.
What you call things. Every practice has its own nouns. The return, the file, the engagement, the close, the deck. A draft using the client’s word instead of yours reads as someone else’s work, even when every fact in it is right.
Then the thing that makes the rest of this work, and that most people skip. You have a handful of voices, and they’re keyed to the reader. You use one register with a client of nine years, another with a prospect, a third with your bank. When you ask for a reply “in my voice”, the honest version of the request is “in the voice I use with this person about this kind of thing.” Every instruction you give a system lands or misses on that axis before it lands or misses on any other.
Why a general draft misses in the same direction
Two things make a stranger’s draft feel synthetic. They’re worth separating, because telling the system about yourself fixes one of them.
The first is that a system starting cold writes to the middle. The middle of professional email is long, formal and eager, because that’s the safest place to stand when the relationship is a blank. So the misses arrive in a predictable order: too long first, too formal second, too eager to commit or too keen to hedge third. Underneath sit the tells everybody can now recite. Hoping the message finds you well. Wanting to reach out. Not hesitating to reach out. Thanking you for your patience on something that came back the same day. A closing paragraph that restates the message that just ended.
The second is structural, and it holds however much a system knows about you. Suggestion systems offer what’s likely, and what’s likely is ordinary by definition. The clearest measurement of what that does to writing is Predictive text encourages predictable writing, presented at the 2020 conference on Intelligent User Interfaces by Kenneth Arnold, Krysta Chauncey and Krzysztof Gajos. They had 109 people write 1,308 image captions on a phone keyboard. The one thing that varied was whether predictive suggestions were visible. “Our key findings were that captions written with suggestions were shorter and that they included fewer words that the system did not predict,” they report. The fraction of predictable words ran about six percent higher when suggestions were always visible than when they were hidden. Their illustration is the whole problem in one word. A photograph that would have drawn “hitter” or “baseball player” drew “man” instead.
The part of that study worth sitting with is that the writers liked it. Asked which of the three keyboards was most helpful, they picked the always visible one 206 times and the suggestion-free one 20 times. They reported lower effort with suggestions on. People prefer the tool that makes their writing more ordinary, and the drift stays quiet while it happens. Being a good writer leaves you just as exposed.
The same pull shows up in the engineering. Google described Smart Reply in its 2016 paper at the Knowledge Discovery and Data Mining conference, and response diversity was one of the named problems they had to solve. Left alone, the model gravitated to a small pool of safe answers. So intent labels were attached to responses, to force variety into the three shown to you. Genericness is the default state of a system optimizing for likelihood. Matching your voice is the counterweight.
Which brings up the question everybody asks third and should ask first. Will the client be able to tell?
Probably not, and the worry belongs somewhere else. Human heuristics for AI-generated language are flawed was published in the Proceedings of the National Academy of Sciences in 2023. Maurice Jakesch, Jeffrey Hancock and Mor Naaman ran six experiments with 4,600 participants. They found people “unable to detect self-presentations generated by state-of-the-art AI language models in professional, hospitality, and dating contexts.” The cues people trust point the wrong way. Readers treat first-person pronouns, contractions and mentions of family as marks of a human writer, and a model produces all three on request.
So the risk is suspicion. Artificial intelligence in communication impacts language and social relationships was published in Scientific Reports in April 2023. Jess Hohenstein, René Kizilcec, Dominic DiFranzo, Zhila Aghajari, Hannah Mieczkowski, Karen Levy, Mor Naaman, Jeffrey Hancock and Malte Jung randomly assigned 219 pairs of people to have smart replies available or withheld. Conversations with them ran faster, warmer and were rated as closer. And people were judged more negatively when their partner suspected algorithmic help, whether or not they had used any.
Read those two together and the practical conclusion is narrow. You care about voice because a message that reads like you keeps the suspicion from starting. The penalty lands on the suspicion rather than on the fact.
What a system needs to sound like you
Three kinds of evidence are available to any drafting tool. They’re worth telling apart, because they differ in what they cost you and what you can check.
Your own sent mail. Years of it, most of it written on autopilot, which is exactly why it’s a good record of your defaults. The question to ask is how a tool reads it. Read at the moment of drafting, as context for this reply, is one arrangement. Absorbed into a model that other people’s mail also shaped is a very different one, with different consequences for client confidentiality. Settle that one before anything else. Keeping your data out of AI training is the page for that question, and this one assumes you’ve answered it.
The relationship in front of you. How this thread has gone, and how you’ve written to this person before. This is where the register actually gets decided. Your voice with a nine-year client is a fact about that correspondence.
What you’ve said out loud. One sentence, said once, kept as a standing rule. This is the one of the three you can read back, so you can tell whether the system heard what you meant. How a preference said in passing becomes a rule that holds everywhere is how Point learns your habits. Opening the list and seeing what Point believes about you is seeing and editing what it learned.
Four questions are worth asking any vendor before you judge the drafts. A bad answer to any of them makes the drafts uninformative.
- Where does the voice come from, and is my mail context at drafting time or training material?
- Can I read back, in plain sentences, what the system thinks my preferences are?
- Can a rule be scoped to one person or one kind of mail, or is every instruction global?
- Can I remove a rule as easily as I made one?
The last two do most of the work. A system where every instruction applies everywhere gives you one draft you like, and quietly reshapes your mail to the bank. Connecting those two events is the hard part.
Three sentences that do most of the work
A style guide is more than this needs. On the evidence of what people actually correct, three standing instructions cover most of the gap. Each is one sentence with its scope on the front.
The sign-off, with an audience. “Sign client mail with my cell number and my first name only.” Sign-offs are the most consistent thing you do and the easiest thing to get wrong. They show on every message, so fixing this once pays out constantly.
The register, with a name or a category on it. “Keep it plain and short with Marcus, he reads everything on a phone.” Name the person. “Be less formal” is a rule about your entire mailbox, including the letters you want to stay formal.
A ceiling, with an occasion on it. “Two paragraphs maximum when I’m answering a question about fees.” Length is where a general draft misses hardest, and a limit tied to a situation holds better than a limit tied to your mood.
A fourth is worth adding once you’ve seen a week of drafts. It’s the phrase you never want to appear under your name. Most people have one. Write it down as a flat prohibition rather than as a preference, because a prohibition is unambiguous.
Two disciplines about how you say them. Put the scope inside the sentence, because a rule you state while looking at a note to a friend reaches your letter to the bank too. And make one at a time, so the next draft tells you which half worked. Which corrections are worth turning into rules at all is worked through in when a summary learns from the corrections you make, and it holds the same way for writing as for reading.
How to tell whether it worked
The instinct is to judge each draft as it arrives. A draft you liked and a draft you’d have written are different objects, so one draft tells you very little. Judge the sequence instead. Two weeks of ordinary mail is enough.
- Measure what survives. For each drafted reply, mark it as sent untouched, sent after a change or two, or discarded and written yourself. Three tallies on one piece of paper. A proportion is the honest evidence here, and it beats an impression formed on a Tuesday afternoon after two good drafts.
- Treat a repeated change as an unwritten rule. The same correction twice means you’ve found a habit the system has yet to learn. That’s the moment the correction is worth thirty seconds, and it’s the only moment it is.
- Keep register separate from fact. A draft that’s a shade too formal costs you ten seconds. A draft that’s perfectly in your register and wrong about a date costs whatever the date was worth. Sorting the two is the practical skill, and your first drafted reply is where it gets learned.
- Check the edge as well as the middle. After you set a rule for client mail, open one message that fell outside it and confirm it reads the way it always did. A rule that improves what you aimed at and quietly reshapes everything else shows up later, somewhere you’re looking away from.
- Edit rather than rewrite while you’re measuring. A draft you deleted at the first line looks the same afterwards as a draft that was bad, so the week ends with sent replies and an empty tally.
One caution about the reading. Your standard rises. In week one you grade against “would I send this”. By week three you’ve drifted to “would I have written this”, and that’s a bar you’d fail yourself, because you write the same email two ways on two days. Sendable is the target. Identical is beyond anybody.
The draft that sounds most like you
Here’s the part that gets underweighted, and it’s the reason to care about any of this beyond vanity.
A draft in a stranger’s voice gets read. Every sentence is faintly wrong, so your attention stays on. A draft in your own voice gets skimmed, because it reads like something you already decided to send. Voice fidelity is the feature, and it takes away the friction that was doing your proofreading for you.
So the reading gets a fixed shape rather than an amount of care. Care comes and goes on a Thursday in March; a list holds.
- Dates and times. The commonest confident error. A proposal buried in message six of a thread reads as settled when it’s still open.
- Money. Figures carried across from the last engagement are correct and out of date, which is the hardest kind of wrong to catch.
- Anything that agrees. Any sentence accepting a deadline, a scope, a discount or a meeting. A draft is written to be agreeable, and agreeable is where commitments come from.
- Who’s on it. Reply versus reply all, and whether the person you’d have added is there.
- What happened off the thread. Half of most decisions happen on a call. What you settled by phone on Wednesday lives outside the messages, and voice matching leaves it there.
All of that sits outside tone, and it gets harder as the drafts get better. That’s why a fixed list beats vigilance.
The other half of this is when a draft is allowed to leave at all. That’s a separate setting from how well it’s written, and it deserves its own decision: why a draft waits for you by default. Worth knowing before you make it. A reply in your voice goes out from your own address and threads under the original. On the recipient’s side it’s your message in every sense that matters, and on your side it’s your message in every sense your engagement letter cares about.
When your voice is the wrong tool
Say this once and move on. Some messages are worth writing yourself, because the fact that you sat down is the content.
Bad news you’re responsible for. A fee increase. An apology after a mistake on your side. A condolence. Anything heading toward a dispute, where the wording gets read twice by somebody looking for a lever. And the first message to a person who’s never heard from you, where you’re setting a voice rather than matching one.
There’s a quieter case too. Your practice has standard wording for some things: engagement terms, a disclaimer, a statutory notice. The house language is better than your personal language there, and reproducing your voice actively works against you.
Everything else is ordinary correspondence, and ordinary correspondence is most of the mail.
How Point drafts a reply
Point runs as a full email client on top of the Gmail or Microsoft 365 mailbox you have now. You keep the mailbox you have and you keep the address. A thread that needs a reply has one waiting before you get to it.
You open the conversation and the reply sits under it. One reply, whole: addressed, in paragraphs, with a sign-off. A gray auto-drafted mark is the one sign Point wrote it, and changing a single line removes the mark, because at that point the words are yours. Point writes it out of your own mail rather than from a template, so the shape of it is familiar before you’ve read a sentence.
Standing preferences ride along with every draft. Tell Point once that client mail carries your cell number, and it appears in the sign-off on a thread where you never mentioned it. Say in passing on a thread that somebody is a friend and should hear from you informally, and Point offers the rule back in plain words. One click keeps it, and the next reply on that thread arrives in the new register. Every preference lives in one list in plain words, newest first, marked as added by you, editable and removable. What Point believes about how you write is a list you can open and read.
You can also just say what you want. With a thread open, seven words is enough. Point already knows which conversation you’re looking at and who’s on it, so the draft comes back addressed to the right person on the right thread. Starting a message with a blank thread behind it is a different job with different inputs, and it’s composing an email from scratch. A reply to somebody who wrote to you in another language raises its own question, covered in drafting emails in another language.
How much Point does without asking is a level held separately for each type of work, running from suggest-only through review to fully handled. Drafting starts at review. The words get prepared, and the sending waits for you. Moving that level is a decision you make per kind of work rather than all at once: setting how much your inbox does on its own is where that lives. Everything Point has done sits in a plain-language log. Undo reaches what Point did to things you hold, and a message already on somebody else’s server sits outside that. The benefits page lists the rest of the work that’s already done by the time you sit down, and the case for the writing happening inside the client rather than in a tab beside it is what an AI email client is.
Common questions
Can software really write in my voice, or is that a sales line?
It holds the reproducible part, which is smaller than the phrase suggests and covers most of what makes a message recognizable: length, register, opening move, how firmly you commit, sign-off, the nouns your practice uses. Judgment stays with you, and the gap between those two is exactly where the reading has to happen. The way to test the claim is to count, over two weeks, how many drafts you send with nothing changed.
How long does it take before the drafts sound like me?
The part that comes from your existing mail is there on the first draft, because your sent folder is already written. The part that comes from instructions arrives one sentence at a time, and three sentences cover most of the gap: sign-off, register with a name attached, and a length ceiling for one kind of message. The two weeks are yours rather than the system’s. You’re accumulating enough drafts to know which corrections are habits rather than one-offs.
Do I have to tell clients an AI drafted the reply?
That’s a firm policy question rather than a voice question, and it’s worth deciding deliberately rather than by default, because the research says people penalize suspected AI use whether or not it happened. The practical answer for most practices is that a message you read, edited and sent is your message. The disclosure conversation belongs with your engagement terms rather than with each email, and engagement letter language for AI works through what to say and where.
What if I write completely differently to different people?
Then you’re normal, and it’s the single most important thing to tell the system. Voice is per relationship, so the instructions that work carry a who or a when on the front of them. “Be less formal” is a change to your whole mailbox. “Keep it short and plain with Marcus” is a change to one correspondence, and that’s the shape almost every real preference actually has.
Will leaning on drafts make my own writing more generic over time?
It’s a real effect and worth naming rather than dismissing. Writers given predictive suggestions produced shorter, more predictable text and preferred the experience, so the drift stays quiet while it happens. Two things blunt it here. A draft written from your own past mail pulls toward your defaults rather than toward the population average. And editing a finished draft is a different act from accepting a suggestion mid-sentence: you react to a whole message rather than picking the next word from a list.
What if the draft is perfectly in my voice and says the wrong thing?
That’s the failure this page is most worried about, because a message in your register is the one you skim. Register and accuracy are separate axes, and improving the first leaves the second where it was. The defense is a fixed reading list rather than extra care. Read the dates and times. Read the money. Read any sentence that agrees to something. Check who’s on the message. Then ask what was settled on a call and never made it into the thread.
Can I undo what it learned about how I write?
You should be able to, in the time it took to say it. An instruction that suited your March correspondence can be wrong by the fall, and it stays until you retire it. So what matters is that the list reads as plain sentences and that deleting a line takes one action. Seeing what a system believes about your writing is what lets you tell an over-reaching rule from a system that’s simply wrong, and those two have opposite fixes.
The short version
Your voice is a set of habits rather than a personality, and the habits are the reproducible part: how long, how close, how firmly you commit, how you sign off, what you call things. You have several of them, keyed to who you’re writing to, so every useful instruction carries a who or a when inside it. A draft starting cold will be long, formal and eager, in that order. That’s the middle of business email, and the middle is where a system stands until it knows you.
Judge the sequence rather than the draft. Count what you send untouched. Treat any correction you make twice as a rule you have yet to write. Check that a new rule stayed inside the audience you named. Hold the target at sendable rather than identical. Then read the good ones hardest. A reply in your own register is the one that slides past, and the date is right or wrong whatever the tone.
What you end up teaching depends on the mail. An accounting practice mostly needs a short answer with the figure in it and a sign-off that carries a phone number. In a consultancy the commitment language wants watching more closely than the register. For any other desk, the same morning is set out for other work.