A reply arrives with most of the setup already done. The recipient is decided. The subject line exists. The last twenty messages settled how formally to write, and the question you’re answering sits in front of you. Start a message yourself and you supply all of that. Composing from scratch is a different job, and it’s a job you can hand over once you know the four things to put in.
- A thread hands over five things at once. Who it goes to, what it’s called, how formally to write, what it’s about, and why you’re writing at all. On a blank window you supply all five, so the work moves from judging a draft to specifying one.
- The person reading it starts fresh too. That’s the same fact seen from the other side, and it’s why a message you started carries its own reason for existing in the first two lines.
- Whatever separates the words you typed from the message that came back is inference. A long email out of a short instruction is mostly a guess at what you meant, and the guess is the part to read.
- The field worth a second look is the one at the top. A reply goes back to whoever wrote to you. A new message goes wherever the name you typed resolved to.
- Most messages you start are asks. An ask travels worse over email than the person sending it believes.
What a reply gives you for free
Worth being literal about this. The list is the whole difference, and almost nobody has counted it.
The recipient. A reply goes back to the person who wrote, and the copy list arrives with the thread. The thread decides both, so both arrive right.
The subject. Already written, and written by them. You inherit it, good or bad.
The register. How close you stand to this person sits on the record in your own correspondence. Anybody who reads the exchange can see it, and so can anything that reads it. The register comes with the thread.
The content. Somebody asked something. The message has a job, and the job is stated.
The reason for writing. You’re replying because they wrote. That’s a complete justification, and the thread carries it for you.
A blank window leaves all five to you, and the last two matter more than they look. On a reply, the purpose of the message sits outside it: the thread explains why the message exists. On a message you start, the purpose rides inside the message itself, in words you put there. State the purpose and a drafting tool can do a great deal with it. Leave it out and the draft is guessing.
There’s a second difference underneath. Replies are mostly answers. Messages you start are mostly requests, announcements, or introductions, so the work has changed direction. You’re setting the agenda now. You’re asking a person whose mind was elsewhere to do something, and that’s a harder piece of writing before any software is involved.
The person reading it has no thread either
When a from-scratch email goes wrong, the writing is usually fine. What goes wrong is that the writer knows their own context and assumes it reached the page. That failure has been measured directly.
Four researchers put a number on it. Their paper is Egocentrism over e-mail: Can we communicate as well as we think?, published in the Journal of Personality and Social Psychology in 2005. Justin Kruger, Nicholas Epley, Jason Parker and Zhi-Wen Ng ran five studies on whether people can tell how their email will land. The third put 154 pairs of University of Illinois students in one of their own homes. Each person wrote five statements, each meant to carry a particular tone, then predicted whether the other would pick it up. Across all three media tested, senders expected 88.8 percent of their intended tones to be read correctly. The actual figure was 70.4 percent. The same gap appeared on the receiving side. Readers believed they had understood 89.4 percent of what they were sent and were right about 70.4 percent. In the email condition specifically, 89.3 percent confidence sat on top of 62.8 percent accuracy.
Two details make this more useful than a general warning about tone. First, the same study included voice-only and face-to-face conditions, and the overconfidence was distinctly worse over email. Second, and more to the point here, familiarity left the gap where it was. The researchers deliberately recruited half the pairs as friends and half as strangers, and found that “participants were just as overconfident when communicating with a stranger as when they were communicating with a friend.” The gap stays the same size with people you know well.
Those were students conveying sarcasm and sadness, rather than partners quoting a fee, and the study stays silent on business correspondence. What carries across is the mechanism. Your own context is impossible to un-know. You know why you’re writing, you know what happened last week that prompted it, and you know which of the four Karens you mean. All of that reaches the page when you put it there. A thread normally does that work for you. On a message you start, the sentence that would have explained everything is the one you’re most likely to skip, because to you it’s obvious.
So the practical shape of a good from-scratch email comes from the reader’s position rather than from style. Three things sit near the top. Which matter this is, why it’s arriving now, and what you want them to do. Hand a draft all three and it has what it needs. Those three come from you.
What the instruction has to contain
You’re specifying a message rather than reacting to one, so the specification is where the work is. It’s shorter than people expect, and it’s about facts rather than tone.
Four things belong in it: the person, the occasion, the outcome you want, and any limit that matters.
Write to Karen Alvarez at Brightline. The fieldwork moves to the week of the fourteenth. Apologize for the change, ask her to confirm that week works, and leave the delivery date open.
Set that against the version most people type, which is write to Karen about the audit. That one produces a competent, plausible email about the audit. Every specific in it came from somewhere you left to the draft.
Look at what the good version leaves out. Warmth, formality, length and sign-off all stay unsaid. Those come from your own sent mail and from the standing rules you’ve already given, which is the subject of how Point learns your email habits and, for the register itself, getting AI replies that sound like you wrote them. Your instruction spends itself on facts and on the ask. Those are the parts only you can supply.
Two things are worth saying out loud, even when they feel obvious.
The limit you’re setting. A drafted message is written to be agreeable. Say in the instruction which date, number or scope stays open, and the draft leaves it open. Catching it afterward depends on you spotting it.
The one action. A message with two asks in it gets one answer, usually to whichever is easier. If you genuinely need two things, that’s two messages or a phone call. Decide which before the draft exists.
Where the context comes from without a thread
The blank window has more behind it than it looks. A tool sitting inside your mailbox has four other places to look, and knowing which they are tells you what you still supply yourself.
Your contacts. A name resolves to an address, so you write to a person rather than to a string. This is also the single point where a from-scratch message fails hardest, so it’s the one to read rather than trust. How that list gets into a usable state in the first place is cleaning up duplicate contacts.
Your past correspondence with this person, on other threads. This message starts fresh, and your history with the person usually runs deep. That history is where the register comes from. It’s why a first draft to a nine-year client sounds different from a first draft to your bank, and that happens on its own.
Your calendar. The meeting you’re writing about, when it is, and whether the time you’re about to propose is one you actually have.
Your standing rules. The sign-off, the phone number on client mail, the person you agreed to keep things short with. Those apply to a message you started exactly as they apply to a reply, which is the point of keeping them in one place. Reading back what those rules currently say, and taking out the ones that have gone stale, is seeing what an AI learned about you.
Three things sit outside all four, and that’s where the confident errors live. Anything settled on a call and still unwritten. Anything you decided and still hold in your head. Any figure that has changed since the last time it appeared in your mail, which is the hardest kind of wrong to catch, since the sentence is accurate about last year.
Which gives the one reading habit that matters most here. Compare the size of what you typed with the size of what came back. Every specific in the message beyond your instruction, your history with that person and the file you attached is an inference. Inference is most of why the draft is useful, and it’s also why some sentences deserve a second read. On a from-scratch message those are the long ones that arrived on their own.
The subject line is yours to decide
A reply inherits its subject. A new message invents one, and it’s the one part of the message you can be sure the recipient reads.
A subject line does two jobs, and a draft can infer one of them. Saying what the message is about is a summary of the body, which is straightforward. Saying how much of their attention it deserves is a judgment about your relationship and their week, and that judgment lives outside the body. That second job is the reason a drafted subject line is worth a second of your time even when it’s accurate.
Three things make one work, and all three are plain.
Use the noun the recipient uses. They say the audit where you say the engagement. They say their taxes where you say the file. A subject line written in your practice’s vocabulary reads as internal mail from somebody else’s firm.
Put the date in it when there is one. “Fieldwork moving to the week of the 14th” is findable in eighteen months. “Quick question” disappears into the archive for both of you.
Keep the ask in the body. A subject line gets skimmed, so an ask that lives only there leaves you following up on a message that was technically complete.
That last point has a cost attached to it. An unanswered message stays open in your head, and where that debt lives is unanswered emails and the threads that come back.
Writing to somebody for the first time
Cold contact is the hardest version of this, and it’s the one where a draft has the least to work from. So it earns a section rather than a caveat.
Start with the honest part. Everybody is worse at this than they think, and again it has been measured. The paper is Ask in person: You’re less persuasive than you think over email, published in the Journal of Experimental Social Psychology in 2017. Mahdi Roghanizad and Vanessa Bohns had 45 participants approach ten strangers each with a small request. By Cornell’s own account of the study, face-to-face requests were 34 times more effective than emailed ones. Requesters consistently overestimated how many people would say yes over email, and underestimated it in person.
Strangers on a campus asked to fill in a survey sit a long way from a referred prospect reading an introduction from a CPA, and the multiple belongs to that setting rather than to your desk. What survives is the direction of the error, and its cause, which is the same egocentrism as before. The sender can see how reasonable the request is. The recipient sees an email from a stranger, and reads it with a suspicion the sender would find surprising.
Three consequences for the message itself.
A first message reaches for your general default, since that’s the only register on file. The default is usually fine. It’s also the one case where the drafting tool works from the least information, which is worth knowing before you judge the result. The related question of whether a first message is one to write personally is handled from the voice side in getting AI replies that sound like you wrote them.
The first line has to earn the second. Who referred you, where you met, what they published, which of their problems you already know about. That’s the part you supply. Your mailbox holds every other part somewhere, and it holds this one only in your head.
Make the ask small and singular. A cold message asking for a decision stays unanswered. A cold message asking for twenty minutes gets an answer either way. If the meeting is what you actually want, the machinery for arranging one without the chain of messages is a meeting scheduler without the email chain.
And when the thing genuinely matters, the email works as the follow-up to a call rather than the substitute for one. That’s the practical reading of the research, and it’s free to act on.
Reading a message you started yourself
A message you began has its own checklist, because it fails in its own ways. A reply can be wrong about the thread. A message you started can be wrong about the world, and it can go to the wrong person entirely.
Who it is going to. This is the field that fails. Verizon’s 2026 Data Breach Investigations Report (checked September 6, 2026) puts misdelivery at the top of the error varieties behind the 1,719 breaches it classes as miscellaneous errors, ahead of misconfigured servers and lost devices. It defines misdelivery as data delivered “to the wrong recipient”. Its own comment on the category is the useful part: “It is very easy to make this type of mistake.” That’s a corporate breach dataset rather than a study of small-firm email, and most misdirected messages cost embarrassment rather than a notifiable event. The transferable part is narrow and worth having. The To field fails through convenience, when a name matches two people and the faster one wins. Read the resolved address, not the name you typed.
Who else is on it. A reply inherits its copy list. A new message invents one, so the person you meant to add and the person who slipped on are both your doing.
Anything specific you left to the draft. Dates, figures, document names, promises about timing. Anything beyond your instruction and your history with this person came from inference, and inference is confident.
The ask, in one sentence. Read the message as though you were busy, which the recipient is. Find what you want them to do in a single scan, and so will they.
What they are missing. The context that’s obvious to you is the context most likely to be absent, and you’re the worst-placed person to spot the hole. If a colleague is nearby, thirty seconds of somebody else reading it beats any amount of your own rereading.
The attachment. A new message is where “please find attached” goes out with the file still on your desk.
One thing this list leaves standing. Once the message is gone it sits on somebody else’s server, and a reversal reaches what is still yours to reverse rather than what has already been delivered. Where exactly that edge sits is set out in undoing what Point did. So on a message you started, the check happens before the send.
The ones still worth typing out yourself
Say it once and move on. Some messages are worth writing by hand, and the share is higher here than it is for replies. The reason is worth naming. You start a message when something has changed. Replies are mostly the ordinary traffic of work. New messages skew toward news, and a fair share of news is the kind where the fact that you sat down is part of the content.
The clear cases are bad news you’re responsible for, an apology, a fee increase, and a condolence. Add anything heading toward a dispute, where every sentence gets reread later by somebody hunting for an advantage. The same holds where your practice has house wording for something, a statutory notice or engagement terms, and the standard language beats a personal one.
There’s also a quieter case, which is the message you can’t specify. If you’ve started three instructions and abandoned all three, the tool is usually working fine. What’s happened is that you’re still deciding what to say, and a draft written from an undecided instruction reads exactly like an undecided message. Write two sentences yourself, then ask for help with the rest.
Everything outside that is ordinary correspondence, and ordinary correspondence is most of what leaves your desk.
How Point writes a message from nothing
Point is the email client, and the Gmail or Microsoft 365 account underneath it is the one you already have. Your address stays the same, and your mail stays where it is. A message you start here leaves as a message from your own account.
One click on write opens a new message in a corner window, and your feed stays where it was. You address it by name rather than by address, because Point already knows the people you write to. The chip shows which address the name resolved to, before you send anything. A copy list works the way it does in any mail client. Subject, message, all of it familiar. Attach a file and a chip confirms which file went on.
You can also just say it. In the prompt bar, one sentence names the person and the point of the message, and the draft comes back addressed and written. With a thread open, Point already knows who and what, which is why a reply takes seven words. From a blank start those are the two things your sentence carries, and that’s the whole difference this page is about. If something is off, say so in the same breath and the draft follows, and you can ask by voice instead of typing.
What you’ve taught Point rides along, and you say it once. The sign-off, the cell number that belongs on client mail, the person you asked to be written to plainly. Those live as plain sentences in a single list you can read, edit and remove from, so they stay checkable rather than a matter of trust. Writing to somebody who does business in another language is a case of its own, and it’s worked through in drafting emails in another language.
How much Point does on its own is held per kind of work, on a scale of three positions: suggest-only, then review, then fully handled. Drafting starts at review. By default the words are prepared and the message waits in the composer for you, and raising that is a decision you make one kind of work at a time in setting how much your inbox does on its own. The case for leaving drafting where it starts for a while is why a draft waits for you by default.
When you send, Point hands the message to your own mailbox and it goes out from your address. The message lands in your real sent folder like anything else you wrote. Everything Point did is written down in plain language as it happens. Composing next to your contacts, your calendar and ten years of your own correspondence is the whole argument for the writing living in the mail client rather than in a window beside it, which is set out in what an AI email client is. The benefits page covers everything else that’s finished before you open the inbox.
Common questions
Can Point write an email that is not a reply to anything?
Yes, and starting from nothing is a different interaction rather than the same one with an empty thread. You click write, or you say what you want in one sentence. With a blank screen in front of you, that sentence names the person and the point of the message. Everything a thread would have supplied for free, the address, the subject, the reason, comes from what you typed or from your contacts and your history with that person.
How much do I have to type to get something usable?
Less than a paragraph and more than four words. The useful test is coverage rather than length. The person, the occasion, the outcome you want, and any limit that matters. Leave tone out, because the register comes from your own sent mail and from the rules you’ve already given. If you find yourself writing instructions about warmth and formality, the rule you actually want is a standing one rather than a note on this message.
Will Point get the right person if I only give a name?
Point resolves the name against the people you write to, and shows you the address it landed on. That address is the part to read. The check exists because the To field is where a message you started fails, and it fails quietly. A name that matches two people picks one, and the message is otherwise perfect. A reply inherits its recipient. A new message decides one, so treat it as a decision.
Should I use this for cold outreach?
You can, and the honest advice is to keep the ask small and expect less than you’d predict. Research on requests to strangers found emailed asks far less effective than in-person ones, with senders overestimating how many people would agree. The one line a draft leaves to you is the one that earns you the second, the referral or the shared context. If the outcome matters, the email works better as the follow-up to a conversation than in place of one.
Does Point know what we agreed on the phone?
No, and this is where a message you started is more exposed than a reply. A reply at least sits on a thread that records part of the decision. A new message is often the first written trace of something settled by voice, so anything from that call comes from you. A date or a number you never typed arrived by inference, and inferred details about a phone conversation are the ones most likely to be confidently wrong.
What is the difference between this and asking a chatbot to write me an email?
A chatbot in a separate window writes good English about a subject you describe to it. Point knows who Karen is and which address that resolves to. Point knows what you’ve said to her before, whether the Thursday you’re proposing is free, and how you sign off to clients. The finished message goes into your sent folder. The prose is the same either way. The difference is everything on either side of the prose, which is why the composing sits inside the mailbox rather than beside it.
Can I keep this for replies and not for messages I start?
Yes. How much happens on its own is held per kind of work rather than as one switch. Drafting can sit at review, where the words are prepared and wait, while something else sits higher or lower. The autonomy dial is where you set that. What Point has learned about how you write is a separate list, and the distinction is worth keeping straight. Teaching Point a habit changes how a message is written, and the dial decides whether it goes.
The short version
A reply hands you the recipient, the subject, the register, the content and the reason for writing. Composing from scratch hands all five to you, so your job moves from judging a draft to specifying one. The specification is short: the person, the occasion, the outcome, and any limit that matters. Leave tone out of it. That part comes from your own mail and from the rules you’ve already set.
Then remember that the reader starts fresh too. People consistently overestimate how well an email conveys what they meant, and friendship leaves the gap where it is. So the context that seems too obvious to state is exactly the sentence to put in. Check the gap between what you typed and what came back, because that gap is inference. Check the address the name resolved to, since that’s the field that actually fails. Check the ask, the copy list and the attachment. And when the message is the kind where sitting down is part of the point, write it yourself.
What you start from scratch depends on the desk. In an accounting practice it’s mostly notices with a date in them and requests for a document. In a consultancy it’s mostly introductions and messages that commit you to something, and the commitment language is what to read hardest. For a different morning, there are other versions of it.