Your writing will hold. When researchers put AI-written self-descriptions in front of 4,600 people, the people couldn’t tell. What a client feels is a reply that misses them, and in coaching the miss is almost never the phrasing. It’s what the reply does: it answers, reassures and closes, where you would have asked, reflected and left the thing open.
- Detection is the wrong worry. The cost comes from suspicion, which attaches to a belief rather than to evidence, and a wrong suspicion damages you as much as a right one.
- Your voice as a coach is a stance before it’s a style. The competency model your training was built on describes it in verbs: questioning, challenging, staying in not knowing, allowing silence. A general drafting model does something else by default.
- The expensive drift happens before you press send. Reading a finished answer changes the answer you give, and that has been measured. A coach is the one professional whose product is the withholding of one.
- Automatic replies and AI replies are two different things that share a search box. An out-of-office reads as an announcement, which is exactly what makes it safe and uninteresting.
What a client can actually hear
Start with the fear. It’s the reason most coaches leave drafting switched off.
The evidence settles it. People can’t pick AI writing out of a lineup. Six experiments with 4,600 participants, published in the Proceedings of the National Academy of Sciences in 2023, found readers unable to detect AI-written self-presentation. The cues readers trust are the wrong ones: contractions, first-person sentences. What people punish is the belief that a message was generated, and the belief does its work whether it’s true or false. A randomized study of smart replies in Scientific Reports, the same year, found conversations ran faster and warmer with them. It also found people judged more harshly once they were suspected of using algorithmic replies. Both studies, and what they mean for the general case, are set out in getting AI replies that sound like you wrote them.
Three things change once the reader is a coaching client rather than a stranger in an experiment.
The first is the baseline. Your client has one, and the study participants started cold. Years of your mail sit in their archive. That cuts in your favor far more often than against it, because one reply that reads a little smooth sits beside two hundred that read like you, and the average wins. It also means that when somebody does start wondering, they can name the exact message that started it.
The second is motive. A coaching client has a reason to ask the question that a supplier or a colleague lacks. If an accountant’s confirmation email was drafted by software, the accounting stays the same. If your reply was, the client is asking something else. They’re asking whether any of this is about them specifically, and that’s the thing they’re buying. That’s why an identical reply passes quietly in one trade and becomes a live question in yours.
The third is what actually provokes the question, and it’s behavior rather than prose. A well balanced sentence leaves a client exactly where they were. They start wondering because something missed: the reply came back in four seconds at 11:00 at night, or it greeted them like a stranger in week eleven, or they answered it and the thread went quiet. Those three belong to handing coaching admin to AI and staying yourself. The fourth sits inside the reply itself. It’s specific to coaching, and it’s the rest of this guide.
Voice is a stance before it is a style
The general account of what makes a reply sound like you is real and useful: sentence length, whether you greet, how hard you commit, how you sign off. Getting AI replies that sound like you wrote them works through it properly. Here it’s the floor rather than the subject.
Coaching has a layer underneath, and the layer is written down. The 2025 ICF Core Competencies came out in September 2025, built from data collected from “nearly 3,000 coaches across the world” (checked September 6, 2026). They define Competency 7, Evokes Awareness, as: “Facilitates client insight and learning by using tools and techniques such as powerful questioning, silence, metaphor or analogy.” Two of the sub-competencies under Maintains Presence are “Is comfortable working in a space of not knowing” and “Creates or allows space for silence, pause or reflection.” Under Demonstrates Ethical Practice, 1.06 asks you to maintain “the distinctions between coaching, consulting, psychotherapy and other support professions.”
Now look at what a general drafting model is built to do. It’s built to be helpful. Helpful means: identify the question, resolve the uncertainty, normalize the difficulty, offer options, propose a next step, close warmly. Each one is a competent, well-mannered act. Each one runs against the four lines above. Silence has no draft. Not knowing has no draft. A model that produced one would be reported as broken.
That’s the whole of the problem, and it hides from any check you’d think to run. Here’s the kind of message that arrives between sessions.
Did the conversation with my director. It went fine I think. Honestly not sure it changed anything. Maybe I’m overthinking this whole job question.
A good general draft comes back something like this, and every word of it is fine.
That’s really great to hear, and it’s completely normal to feel uncertain after a first attempt like that. Overthinking usually means you care about getting it right. A few things that might help before Thursday: notice what you actually said versus what you planned to say, jot down his reaction, and think about what you’d change. Let’s dig into it properly on our call.
Warm, correctly punctuated, plainly written, in your register if your rules are set up well. It’s also consulting with a coaching accent. It takes the client’s own question off the table. It tells them what their uncertainty means. It gives them homework they didn’t ask for.
What you’d have written is shorter and does almost nothing.
“It went fine I think” and “not sure it changed anything” are pulling in different directions in that sentence. Which one is louder when you sit with it? Bring whichever it is on Thursday.
Two lines. Nothing resolved, on purpose. That withholding is the thing the client pays for. It lives below tone, so tone matching reaches straight past it.
Four ways a draft misses that a read-through will not catch
Each of these survives editing, which is why they’re worth naming. You’ll read past all four without flinching, because each one looks correct.
It answers. A question mark in a client’s email is sometimes a request and often something else. Coaching mail is full of questions that are the client thinking out loud in your direction, and the coaching move is to hand the question back. A drafting model reads every question as a job.
It normalizes. “That’s completely normal” and “a lot of my clients feel exactly this way” are the two most common sentences in a drafted coaching reply. Both close a thing down. Both are occasionally true, and that occasional truth is exactly why they get left in.
It runs warm. The smart-reply study found more positive emotional language, and warmth is free for a model. A reply a degree warmer than you’ve ever been reads as generous the first time. By the fourth it reads as unearned, and it spends something you’ll want later. A coach who is delighted about everything has no way left to be delighted about the thing that mattered.
It commits you. A time, an extra session, a scope, a cheerful offer to jump on a quick call. This one is a promise rather than a voice problem, and it’s the most expensive of the four. Give it the first pass of your eye rather than the last. The reply written in your voice on day five separates that pile from the writing pile, and it’s worth reading once before you rely on drafts for anything.
Decide the move before you read the draft
A version of this problem happens earlier than editing. It’s the one that quietly changes what a coach writes over a year.
Co-Writing with Opinionated Language Models Affects Users’ Views was presented at the 2023 CHI Conference on Human Factors in Computing Systems. In it, Maurice Jakesch, Advait Bhat, Daniel Buschek, Lior Zalmanson and Mor Naaman asked 1,506 participants to write a post on whether social media is good for society. Some had a writing assistant configured to argue one side. “Using the opinionated language model affected the opinions expressed in participants’ writing and shifted their opinions in the subsequent attitude survey.” The shift showed up in what people typed. It showed up again in what they reported believing afterward, on a survey taken once the writing was over.
That was an opinion task with strangers, so the finding transfers as a mechanism rather than a number. The mechanism is ordinary and hard to argue with. A finished answer in front of you is an anchor, and you edit from where you’re standing rather than from where you’d have started.
For most trades that’s a curiosity. For a coach it’s the trade itself, because what you sell is the capacity to withhold an answer, and the anchor always supplies one.
The protection is eight seconds long. For any message where your reply is coaching rather than logistics, decide the move before you open the draft. The move, in one word. Everyone tries to decide the wording instead, and the draft will do the wording better anyway.
- Ask. Give the question back, usually narrower than they asked it.
- Reflect. Say what you heard, and stop there.
- Hold. Acknowledge that it landed and leave it for the session.
- Answer. Rare between sessions, and almost always about logistics, money or a boundary.
Say the word to yourself, then read the draft. If the draft is doing the same move, you have a clean editing job and fifteen minutes back. If the draft is doing a different move, you’ve learned something worth more than the fifteen minutes, and you write that one yourself.
This also gives you a number to watch, and it beats any impression of tone. Over a month, how often did the draft’s move match the one you’d already picked? Say the answer is most of the time on scheduling and rarely on between-session mail. Read that as information rather than as a complaint about the software. That’s the software telling you which half of your mailbox should ever arrive with a reply already in it.
The mail that should never start from a draft
Which messages in a coaching practice are arrangement and which are the work gets sorted in handing coaching admin to AI and staying yourself. The line this guide needs is narrower, and it sits inside the second pile. Among the messages that are genuinely yours to write, some can start from a draft and some want writing cold.
Starting from a draft is fine for the ordinary run of client mail. The answer you’ve written two hundred times about what happens if they need to cancel. The explanation of what the discovery call involves. The package description going out to somebody who asked what you charge. Mail in a language you read comfortably and write badly, which has its own page. All of that is arrangement, and a draft you edit beats the version you’d have written at 9:40 at night.
Five kinds are worth writing cold, before you look at anything.
The between-session note. The ICF model puts this inside the competencies: 8.06 has you partner with the client “to summarize learning and insight within or between sessions.” Between sessions. It’s delivered coaching that arrives as email, and it’s the thing you’re paid for.
The first message after a hard session. Whatever you write is going to be read four times.
Anything about ending, renewing, or the money attached to either. The client will read that message more carefully than any other they get from you. You should be the first person who read it carefully too.
Any reply to a client who is drifting. The draft answers the message. The problem sits underneath the message.
Anything where your first reaction was that you didn’t know what to say. This is the strongest of the five and the easiest to lose. That reaction is usually accurate, and it’s information: something in the message is still unplaced. A competent draft on the screen leaves the difficulty exactly where it was and removes your awareness of having one. The awareness was the part you needed.
All of this is a habit rather than a setting. You need to know, before you open the thread, which pile it’s in.
Six sentences worth saying out loud
The mechanics of turning a correction into a standing rule work the same in coaching as anywhere, and they’re worked through in how Point learns your email habits. The failure modes are in when a summary learns from the corrections you make. What’s specific to a coaching practice is the content of the rules, and it has a shape worth noticing.
Most of your useful instructions will be prohibitions. A model’s default is to add. Your voice as a coach is substantially what you leave out, so the rules that do the work are the ones that subtract.
- Never tell a client their feeling is normal.
- Don’t offer advice in client mail. Ask a question instead.
- Never propose an extra session or a call in a draft to a client.
- Keep replies to clients under 120 words unless the question is about scheduling or money.
- Sign coaching mail with my first name only.
- Don’t use the word journey.
One line of caution, because this is the mistake people make with rules of this kind and the fix is cheap. A rule carries the edges its sentence carries. Say “be less formal” while you’re looking at a note to a friend, and it turns up in your next message to a corporate sponsor. Put the who or the when on the front. Then read the list back every couple of months and take out the lines that stopped being true. That’s what seeing what an AI learned about you is for.
Automatic replies are not AI replies
Two different things get called automatic replies, and the search box treats them as one. The first is an out-of-office autoresponder: a fixed message you write in advance, sent to whoever writes to you while you’re away. The second is a reply drafted for one specific message, which you read and send. They fail in opposite directions, and it’s worth knowing which one you’re actually shopping for.
The autoresponder is clean on authenticity, and that’s the point of it. Nobody has ever believed an out-of-office was written for them personally, so it has nothing to betray. Its failures are practical. Microsoft’s own documentation names three worth knowing before you plan a week away around the feature (checked September 6, 2026).
It goes out once. “This feature automatically sends a customized message once to each person who emails you during your absence.” A client who writes on Monday and again on Thursday hears from the machine on Monday. On Thursday, as far as they can tell, they’re ignored.
Outside is the normal case for a coach. Microsoft warns that “Sending replies outside your organization will send your automatic reply to every email you receive, including newsletters, advertisements, and potentially, junk email.” In a solo practice every client you have is outside your organization. That warning covers your entire client list.
The feature is also tied to the account type: “Automatic replies are not supported for Gmail, Yahoo, or other POP or IMAP accounts.” Worth checking in advance rather than on the morning you leave.
The one authenticity rule that does apply is a rule against trying too hard. The temptation is to make the out-of-office warm and personal, and half of that is right. Warm is fine. Personal costs you, because a message that appears to engage with what somebody just wrote, followed by nine days of silence, is precisely the experience people mean when they say they were answered by a machine. Announcements should read as announcements. What actually belongs inside one for a coaching practice, including the part most coaches leave out, is set out in the inbox for a solo coaching practice.
Drafting a coaching reply in Point
Point works on the mailbox you already keep, Gmail or Microsoft 365. Your address, your history and your contacts stay exactly where they are, so there’s nothing to migrate.
Replies come back written the way you write them, learned from your own sent mail rather than from a template with a name dropped into the top. A good answer is usually one edit from sent. Your standing instructions live as plain sentences you can read down, change and delete, which is where the six above would go. Those rules hold everywhere Point writes for you, so an instruction about how you sign client mail lands on a thread where you never mentioned signing. A client who writes to you in another language gets a draft in that language, ready for you to check.
How much Point does on its own is held separately for each type of action, with three positions: suggesting only, preparing the work and waiting, or carrying it out. Every type starts at the middle one on a new account. For drafting, that means Point writes the words and leaves the sending to you. The dial’s own page sets out what each position reaches. Everything Point did is listed with a timestamp and the reading behind it, and most of it can be turned back from that entry. A message that has already left is sitting on somebody else’s server, here or anywhere, and it stays there. That’s the limit on undo, and it’s the reason drafting is the slowest row to move.
Two things matter in this trade specifically. A conversation can be moved onto a channel locked end to end, which puts it beyond Point’s reach entirely, and the cost of that is real: you write that thread yourself, and it sits outside everything Point sorts or summarizes. The other is that a thread stays in place after something is done to it, so a message that needed more than the reply it got is still there when you look.
Say plainly what none of that reaches. Point can hold “ask, don’t advise” as a rule, and that’s a genuine improvement, but it’s a rule about form. Whether this client, in week six, needs a question rather than an answer lives outside the mail, and it stays there. The full inventory is the flat list of what is finished before you open the inbox, and the version written for coaching practices is the same ground with a working week in it.
Common questions
Will my clients be able to tell that AI drafted my reply?
Almost certainly not from the writing. Experiments on several thousand readers found people poor at judging it. The exposure is suspicion rather than detection, and a wrong suspicion costs you the same as a right one. What provokes it in a coaching practice is a reply that misses the person. It comes back faster than you ever have, or it misses how far into the work they are, or it answers a question they were only thinking out loud.
Is it acceptable for a coach to use AI to write client emails?
For arrangement mail, yes, and most coaches are already doing it with a calendar tool. For mail that’s part of the coaching itself, the question worth asking is what the reply does rather than who typed it. Whether to tell clients, and what to say when one asks you outright, is a decision of its own. It belongs to whether AI is okay for a coaching business.
What should I never let AI draft?
The between-session note, the first message after a hard session, anything about ending or renewing an engagement, any reply to a client who has gone quiet, and anything you didn’t immediately know how to answer. The last one is the easiest to lose, because a competent draft removes the sense of difficulty before you’ve used it. The principled version of that line, beside this practical one, is what AI should never touch in a coaching practice.
The drafts are warmer than I am. How do I fix that?
Say so as a standing rule rather than editing it out message by message. Make the rule subtractive, because that’s the direction the problem runs in. Bans on normalizing a client’s feeling, on exclamation points, on offering extra time, and a word limit on replies to clients will do most of it. Then check a thread outside the scope you named, because a warmth rule written while looking at client mail can quietly reach your accountant.
Are automatic replies in Outlook the same thing as AI-written replies?
They’re two different things, and they’re worth keeping apart. An automatic reply is one fixed message you wrote in advance, sent while you’re away, and Microsoft’s documentation says it goes out once to each person who writes. An AI-drafted reply is written for a specific message and read by you before it goes. The first is an announcement, the second is correspondence, and the voice problem belongs to the second.
Will leaning on drafts make my own writing worse over time?
The evidence that exists points at something narrower and more interesting than skill decay. Writing alongside an assistant that favors one view shifted what people wrote and what they later reported believing. Your position erodes before your prose does. For a coach, the habit worth protecting is deciding your move before you read the draft, rather than judging whether the draft is any good.
Can I have different rules for different clients?
Yes, and per-client rules are the ones most worth writing, because coaching registers vary more between two clients than between two trades. Hold any system to one requirement: you can read back what it believes about you in plain sentences, and you can remove a line in the time it took to add one. A preference you can’t see looks exactly like a mistake.
The short version
The worry that keeps coaches off drafted replies is that clients will hear the machine in the writing. Your clients won’t hear it, and the research is clear enough to stop spending attention on the question. What clients respond to is suspicion, which runs on belief alone, and what provokes it is a reply that misses the person in front of it.
In coaching, the misfit has a specific shape. A drafting model answers, normalizes, warms and closes, because that’s what helpful means everywhere else. Your competency model asks for questioning, silence and comfort in not knowing, and none of those has a draft. So the discipline is choosing the move before you read the draft, keeping five kinds of message off drafts entirely, and writing your standing rules as prohibitions. Tone editing comes second. Your voice as a coach is mostly what you decline to add.
Here’s the test three months in. The client wrote four ambiguous lines on a Sunday, and a question came back, the way it would have before. The replies will sound like you, and that part takes care of itself.