Point learns your habits from things you say, one sentence at a time. What Point picks up sits in a list you can read, in plain sentences. There’s no training period and no preferences screen.
Here’s how a rule gets made. You read a draft and the wording is off. You say so, in plain words, while you’re looking at it. Point offers the rule back, and you keep it with a click.
That’s the whole mechanism. The rest of this page is what Point can pick up, what stays out of reach, and how to tell the two apart before you rely on it.
- A habit is the move you already make, before you’ve decided anything. That’s what makes it teachable. It’s also why you can’t list your own on request.
- The learning attaches to the thing that was wrong. A rule you make while looking at a bad draft beats one you make from memory on a settings screen. It’s also the kind people actually get around to making.
- One list serves everything. A sentence you say while reading a reply reaches your sign-offs, your register and your summaries alike. There’s one memory, and every part of Point reads from it.
- A good list stays short. You’re teaching the exceptions to what Point already reads out of your mail. A memory that keeps growing is a symptom, and a short one is the sign it’s working.
- Facts and judgment sit outside all of this. A perfectly learned habit still writes a confident sentence about the wrong Thursday.
The part of your work that repeats
Most of what you do at your desk runs on rails. The share is larger than it feels from the inside.
Wendy Wood, Jeffrey Quinn and Deborah Kashy ran two hourly diary studies, 70 students in the first and 209 in the second. Their paper, Habits in everyday life, came out in the Journal of Personality and Social Psychology in 2002. Each participant carried a wristwatch programmed to chime on the hour. Entries got written while things were happening, rather than reconstructed at the end of the day. A behavior counted as a habit when it had been “performed just about every day and usually in the same location”. Between a third and a half of everything people wrote down met that bar: 35 percent in the first study, 43 percent in the second.
The second finding is the one that matters here. While people were doing the habitual things, their minds were elsewhere. Habitual behavior, in the authors’ own summary, is behavior “performed almost daily in stable contexts”, during which people “were likely to think about issues unrelated to their behavior, presumably because they did not have to consciously guide their actions.”
Those were undergraduates keeping diaries, and the study is about their days rather than anybody’s inbox. What it establishes is the shape of the thing. A large share of what a person does repeats in a stable setting and runs on its own. Correspondence is about as stable a setting as working life provides. Your sign-off is the case in point. You wrote it four hundred times last year, thinking about the client’s question while you did.
So the material Point can learn from you divides into four kinds. Two of the four are worth a sentence.
Habits. The move you already make, most times, before you’ve thought about it. How you open, and how long you run. How firmly you commit to a date. What you sign off with, and what your practice calls the thing. These are consistent, so Point can learn them, and they recur, so they’re worth teaching. What that consistency is made of, broken down properly, is getting AI replies that sound like you wrote them.
Preferences. A choice you’d defend, rather than a reflex. You want the deadlines at the top of every summary. You want the amount kept in. These are teachable too, and they’re the easiest kind to state, because you’ve actually thought about them.
Facts. The fee is $4,200. The extension went in on the ninth. These are true about the world rather than about you. They change without warning, and a standing rule is the wrong container for them.
Judgment. Whether to hold the price. Whether this is the week to push. That stays with you.
One more distinction decides whether a rule holds, and it’s the one people get wrong most often when they start. Teach what you already do, not what you would like to do. “Be more concise” is an intention. It buys you a week of drafts you didn’t ask for, because Point has no way to know how concise you meant, or where. “I answer fee questions in two paragraphs” describes something you’ve already done a hundred times, and there’s one way to satisfy it.
Why a preferences screen asks the wrong question
A settings page puts an empty field in front of you. It asks you to produce an accurate description of your own behavior. There’s a long-standing reason to expect that request to go wrong.
Richard Nisbett and Timothy Wilson reviewed the evidence on how well people can report the workings of their own minds. Their paper, Telling more than we can know, appeared in Psychological Review in 1977. They concluded that “there may be little or no direct introspective access to higher order cognitive processes.” When people explain why they did something, the paper argues, they build a plausible account rather than read off an internal record. It comes from “a priori, implicit causal theories” about what would reasonably have caused a response like that.
Applied here, that says something narrow and useful. Ask yourself in the abstract how you sign off to clients, and you’ll give an answer. It’ll be the version of you that seems right in the moment, rather than the one in your sent folder. Ask how formal you are, and you’ll describe an average that fits none of your correspondents. The blank field gets your theory about yourself.
Recognition is a different faculty, and yours is intact. Show the same person a draft that closes “Kind regards” over a full name. They know inside a second that it should be their first name and their cell number. On demand, that’s out of reach. On the page, it’s instant.
That’s why the learning here attaches to an artifact. Every rule Point ends up holding got made while you were looking at something that was off, which is the one moment you have reliable access to what you actually do. It’s also, as a practical matter, the one moment you’re willing to spend thirty seconds on it.
Three ways a habit gets in
There are exactly three, and they cost different amounts of your attention.
You write one. Open the memory list and type a sentence. “Sign client mail with my cell number and my first name.” It’s saved, marked as added by you, and applied from then on. Use this route for the thing you already know is wrong. It’s the one route that works ahead of a failure rather than after one.
You say it in passing and Point offers it back. You’re reading a thread. The draft under it is polite and formal, and the person it’s addressed to is a friend. You say so, in the thread, in the words you’d use to a colleague. Point turns that into a proposed rule in plain language, and you keep it with one click. This route works whether or not you know a memory exists, which is why it collects most of what a mailbox learns.
You correct something and the correction is kept. A summary comes back faithful, and three times longer than the decision needed. You say what you’d have wanted instead. That sentence becomes a standing rule, rather than something spent on the summary in front of you. Which corrections deserve that treatment, and how to phrase one so it lands the way you meant, is worked through in when a summary learns from the corrections you make. So is how to check afterward that it took.
Two disciplines apply to all three routes. Each gets a line here, because the pages either side of this one work them through properly. Name the audience inside the sentence itself. An instruction given over a note to a friend, with nobody named, is an instruction about every letter you write, your bank included. And add one rule per occasion, since two changes made together leave you nothing to test when the next draft is still off.
One memory, used everywhere Point works
People reasonably expect a drafting setting that lives inside the drafting feature. The arrangement here is simpler. There’s one memory, and every piece of work Point does on your behalf reads from it.
The consequence is worth being concrete about, because it’s what makes one sentence worth saying. Say once that your cell number belongs on client mail, and there it is under a reply, on a thread where the phone never came up. Say once that a particular person gets a warmer register, and the next reply to them arrives that way. Say what a useful summary looks like for you, and the next status report from that sender comes back in the new shape. You said one thing, in one place, and it reached three surfaces beyond the one you were on.
The same property is the risk, and it’s why scope matters more here than it would on a settings screen. A preference with nobody named on it is a preference about your entire mailbox, including the correspondence you’d have wanted left alone.
Three things sit next to the memory and hold still when you teach Point something.
How much Point does on its own. That’s a separate control, set one action type at a time: suggest-only at the bottom, review in the middle, fully handled at the top. Teaching Point a habit changes how a piece of work is prepared, and leaves that control where you put it. Setting how much your inbox does on its own is where that decision lives, and drafting starts on review.
Who matters. Marking somebody a VIP is its own act, with its own consequences for what surfaces first. It’s covered in teaching Point who your VIPs are.
What’s true. A rule changes how something is written or read. The underlying messages stay exactly what was said, and one click away.
What Point has to go on in the first week
On the first morning the memory is empty, and the drafts already sound roughly like you. Those two facts sit close together, and they answer different questions.
Your sent mail already exists. It’s years of offhand choices about length, distance, how you open and how you close. That’s a better record of your defaults than anything you could write down. Point has it as context from the first reply, which is why the first draft usually arrives in roughly your shape rather than in the shape of generic business email.
There’s a question underneath that, and it’s worth settling before anything else. Does a tool read that mail at the moment of drafting, or absorb it into a model that other people’s mail also shaped? The two answers commit you to very different things in front of a client. The page for it is keeping client data out of AI training.
The memory is the other half, and it starts empty. It holds what your sent mail leaves out, what’s wrong in it, and what you’ve changed your mind about. Your old sign-off is in every message you ever sent. The fact that you stopped liking it in March lives in the memory alone.
So the first week has a predictable shape. Three or four sentences get kept. Most are about sign-offs and about one or two specific people, because those are the corrections that come up on every message. Then it goes quiet. The quiet is the finish line. You’ve covered the cases that recur, which was the entire job.
And if you say nothing at all, you’re fine. Point keeps writing to its reading of your mail, which is usually close. Once in a while Point is wrong in exactly the same way every time. That last category is what a sentence is for. A running account of the first two weeks, with the first drafted reply and what to watch on it, is your first reply written in your voice.
Why a short list is the sign it is working
After two weeks of this, the instinct is that a longer memory means a better-taught system. It works the other way around.
What you’re storing is the gap between what Point already infers and what you actually do. Most of your habits come through on their own, because they’re legible in your mail. The list holds the exceptions. On a mailbox that’s being taught well, that comes to somewhere around six to a dozen sentences and then stops growing. There’s a limited number of things about you that are both consistent and invisible in what you’ve already written.
A list that keeps growing is telling you something, and it’s usually one of three things.
You’re storing repairs. A wrong figure, a name in the wrong place, a date misread from message six of a thread. Fix those on the spot and let them go, because each one is a one-off.
A rule missed its mark. If you’ve written two sentences that say nearly the same thing three weeks apart, the first one either missed its scope or slid off. A third lands the same way.
You’re asking for things that were already true. The draft looks the same either way, so you write it again in different words. This one is quiet, and it inflates a list faster than either of the others.
There’s a fourth reason to read the list rather than let it accumulate, and it applies even when every line was right the day you wrote it. Rules stay until you retire them. An instruction that suited your February correspondence can be wrong by September. A year of that, with everything kept, leaves a mailbox running on preferences nobody would write today. Reading the list takes a minute a month. What to cut when you get there, and how to tell an over-reaching rule from a system that is simply mistaken, is seeing and editing what it learned.
The things a learned habit never reaches
This is the part that gets skipped, and it’s where the practical risk lives. Work that sounds like you slides past the eye faster.
A fact. Checking a date stays your job. A proposal buried in the sixth message of a thread can read as settled while it’s still open. A figure carried across from last year’s engagement is accurate and stale at once, which is the version of wrong that survives a careful read.
A commitment. Any sentence agreeing to a deadline, a scope, a discount or a meeting. Learning your habits makes a draft more agreeable, and agreeable is where commitments come from.
Whatever happened on the phone. Half of most decisions get settled on a call and stay there. Anything reading your mail works from the thread alone, however much it has learned.
Somebody else’s habits. What Point learns is about how you work. If you run more than one business, those stay separated from each other. It’s a boundary between the businesses you operate, rather than a wall around each client.
The words already gone. Undo reaches what Point did to things you hold. A message sitting on somebody else’s server sits outside that. The activity log is where you check what was done and turn back what’s still yours to turn back.
The useful conclusion is a small one. Register and accuracy are separate things, and teaching the first leaves the second exactly where it was. A reply that sounds exactly like you is the one you’re most likely to skim. So the reading gets a fixed list rather than an amount of care. Check the dates and the figures. Check any sentence that says yes to something. Check the list of people it’s going to. And check anything decided in a conversation the thread never saw.
How Point holds what it has learned
Point is a full email client, sitting on the Gmail or Microsoft 365 account you already use. Your mail stays where it is and your address stays yours. The reading and the preparing are finished before you sit down.
Everything Point has picked up about how you like things sits in a single list. It’s written as ordinary sentences rather than as settings, newest at the top, with the ones you wrote yourself marked as added by you. Each line is editable and each line is removable. Everything Point believes about how you work is there in a form you can read back, which is what makes the rest of this checkable rather than a matter of trust.
Rules get in the three ways described above. You open the list and write one directly. Or you say something in passing on a thread, and Point puts the rule into words and offers it back. Keeping it is one click, and the next reply on that thread comes back in the register you named. Or you tell Point what was off about a summary. That sentence is kept as a standing preference rather than spent on the one summary, so the following week’s report from the same sender comes back in the shape you asked for.
From there the rules ride along with the work, whichever surface it appears on. You say each one once, and it holds on the threads it applies to. A reply to somebody who wrote in another language raises its own question, and that’s drafting emails in another language. Writing to somebody cold, with no thread underneath to work from, changes what there is to go on, and that’s composing an email from scratch.
How much Point does unprompted is set per type of action rather than once for everything. The scale runs from suggest-only, through review, to fully handled. Drafting begins at review. The message is written, and then it waits for you. Raising that is something you do for one kind of work at a time, and why a draft waits for you by default is the case for leaving it where it is for a while first. Everything Point has done sits in a plain-language log. The entry that recorded an action is where you go to reverse it, as far as the thing is still yours to reverse.
The benefits page lists the rest of what’s already finished by the time you open the inbox. And the case for the remembering living inside the thing you read mail in, rather than in a separate window alongside it, is what an AI email client is.
Common questions
Does Point watch what I click and learn from that?
What Point keeps comes from things you said: a rule you typed, a sentence you offered in passing on a thread, a correction you made to something it produced. There’s a good reason to draw the boundary there. Actions are ambiguous where a sentence is plain, because archiving a message unread looks identical whether it was junk, already settled on a call, or read on a train. A sentence says which.
How long before it knows how I work?
Whatever comes out of your existing mail is in the very first draft, since your sent folder was written years ago. Rules are the slower half, and they arrive one at a time. Most mailboxes are through the bulk of them inside a week, because the corrections that repeat are few and they announce themselves quickly. The longer part is you, watching enough drafts go by to tell a habit from a one-off.
Where does what it learned live, and can I read it?
In one list, in plain sentences, newest first. Each line is editable and removable, and the ones you added are marked as yours. This matters more than it sounds. Reading the list is how you separate a rule that reached further than you meant from a plain mistake, and the two call for opposite responses.
Is what I tell it used to train the AI?
That’s a different question from whether it’s remembered, and the two are worth holding apart. A rule stored against your account and text used to fit a model are separate arrangements, and one of them raises a question you’d have to answer to a client. What happens to your mail on that front is where your data goes, and where it does not. Read that page before you connect anything.
Can a rule apply to one person and not to everyone?
Yes, and stating who or when is the single most important thing about the sentence you write. A preference with no audience on it is a preference about your whole mailbox. “Keep it short and plain with Marcus” changes one correspondence; “be less formal” changes your letters to the bank as well. The two events will look unrelated to you when it happens.
I said something and nothing changed. What now?
This is common, and it has three ordinary causes. Check them in order, and each takes a moment. First, whether the rule is in the list at all, in the words you meant. Second, whether the next message you’re judging it on falls inside the scope you named, since a per-person rule proves nothing on a stranger’s thread. Third, whether you asked for something Point was already doing, which is common and produces exactly this feeling. If all three are clear and the drafts still read the same, edit the sentence rather than add another.
Does teaching it my habits mean it starts acting on its own?
They’re separate settings. What Point has learned changes how a thing is written or summarized. How much Point does on its own is a level you set per type of action, and it starts on review, where work is prepared and waits for you. Raising it is a deliberate decision about one kind of work at a time, made in setting how much your inbox does on its own.
The short version
A habit is the move you already make before you’ve decided anything. That consistency is what makes it something software can hold. It’s also why a settings page gets a theory when it asks you to list your own. What you can do reliably is spot a violation in about a second. So the learning attaches to the moment of noticing. You see a draft that reads wrong, you say so in one sentence, and that sentence is kept and reused rather than spent on the message in front of you.
Teach what you already do rather than what you’d like to do. Put the audience inside the sentence. Expect the list to stop growing after the first week. It stays short because you’re storing the exceptions to what your own mail already shows. Read it once a month and delete what’s stopped being true, since a rule stays until you retire it. And keep the reading of drafts on the facts, because a message that sounds exactly like you is the one that slides past. Checking the date is still yours.
Which rules you end up writing depends on what lands in your inbox. In an accounting practice they tend to be about brevity and about what has to sit above the fold, a figure and a date. In a consultancy they tend to be about how hard a sentence is allowed to commit you. If your desk is a third thing, there are eight versions of the morning.