Everything Point does on your behalf is written down. Which message it touched, what it did to it, when, and on what reading, set out in plain words rather than in a code, with the thread itself a click away from the row. Most of those actions can also be put back.
That is the third of the six questions answered, and the answer is duller than the question sounds, which is usually a good sign. What deserves the rest of a chapter is why the record matters, because the obvious reason is not the real one. It is not for catching the software out. You will open it with any suspicion in mind perhaps three times in a year. Its real work is quieter, and the way to see it is to look at how you already come to trust work you did not watch being done.
Why a record beats watching
The reflex when you hand something over is to supervise. Keep an eye on it for a fortnight, satisfy yourself, then relax. That is the natural instinct, it is the wrong instrument here, and it was never really the instrument in your own firm either.
Think about how work you did not see gets accepted in a practice. Somebody does a piece of it while you are with a client. You do not stand behind them and you do not perform the work again to check it. What you do is look at what they left behind: the figures, where each came from, what was assumed, what was queried and by whom. The work becomes trustworthy because it can be reconstructed afterwards by a person who was not in the room when it was done. Every trade that has to answer for work carried out of sight has arrived at some version of that arrangement, and none of them arrived at watching.
Which is precisely the instrument the first chapter said this was going to need. A model’s prose gives your eye nothing to catch, so finding a summary satisfying to read is not evidence about the summary. A record has a property the prose does not. It is made of things that either happened or did not, at a time, to a named thread, in an order. A paragraph can only be judged. An action can be looked up.
A wrong outcome you cannot place
Here is where a record earns its keep, and the scene is more ordinary than the one people imagine.
Four weeks in, on a Tuesday, something is off. A thread you needed was not in front of you and turns out to have been filed. Or a client who answered you in March has just been chased for the thing they answered about. Nothing that will end a relationship. Enough to make you sit back.
Take that Tuesday with nothing to consult. You know the arrangement produced a wrong outcome and you cannot say what produced it. A one-off. A setting sitting a notch too high. Something it picked up in week two from a message you happened to deal with strangely. Your own hand on a Friday you barely remember. When a fault has no location, suspecting the whole arrangement is the only proportionate response available, because you have no way to isolate the part that failed. And that is what actually happens to people. One unexplained Tuesday ends most trials of most tools of this kind, and the owner walks away holding a verdict about an entire category assembled out of a single event they never got to the bottom of.
Take the same Tuesday with the record open and the event has coordinates. Filed at twenty to nine, on the reading that the thread was a receipt rather than a client. It has stopped being a mood about software and become a fault in one place, and a fault with a location can be repaired. One setting moves, or one learned habit gets corrected, and everything else stands, because nothing else was implicated.
That is the defect the first chapter found in vague caution, met again from the opposite end. A doubt with no parts in it cannot be settled by anything. Evidence with no parts in it cannot settle anything. What the record protects is your ability to revise your view of this software, and a view you can revise is the only sort worth forming while you are still early.
What a way back is worth
The other half of the answer is that most of what appears in the record can be put back, and the value of that sits somewhere other than where people put it.
It is not that mistakes get erased. It is price. When being wrong inside your own account costs a click, you can find things out by trying them, which is how you learned every other piece of software you rely on. Nobody read the manual for the filing portal. You pressed something, watched what it did, and pressed it back.
A decision that cannot be taken back has to be made on faith, in advance, out of a supplier’s description of itself. A careful person will not extend faith on those terms, and should not, which is exactly why careful people spend two years on a question a fortnight of use would have closed. Being able to reverse something is what turns it from a decision you have to reach into an experiment you can run.
What comes back is what Point did inside your own account, and that is a wide territory. A thread it filed returns to the feed. A task it lifted out of a message can go. A draft it prepared can be discarded with nobody any the wiser, and the ranking and the labels and the tidying it did while you were out are all sitting in a place you administer. Standing instructions work the same way: ask to be told when a particular client replies or when a deadline shifts, and that watch is listed where you can see it and stopped whenever you decide. Which closes something the outbound perimeter left open, since an instruction you gave in March and cannot see the results of in November is not one you are supervising. Where all of it stops is the same edge as before. A message sitting on somebody else’s server is past every mechanism described here.
The expectation you do not have yet
There is something you have about a colleague of four years that you have never had to put into words. You know what they will get right unasked, where they go thin, and which kind of thing they will come and query. It is worth more than any procedure you have ever written down, and you did not get it by reading their finished work. You got it by watching a few hundred of their small decisions go past, in order, over months.
On the first morning you have none of that about software, and there is no way to acquire it from the output, for the reason running underneath this entire series: a sound decision and a poor one arrive in the same untroubled prose. The information lives in the sequence or it lives nowhere. Twenty consecutive actions tell you what it reached for first, what it treated as routine, and where its reading of a thread was thinner than yours would have been. Twenty finished summaries tell you almost nothing.
So the early reading is not supervision, whatever it looks like. It is how the expectation gets built, and the reason to do it properly is that it has an end. Once you have the sense of the thing, the record stops being a study and becomes something you open on the rare morning when something surprises you. Reading it closely for a while is how you earn the right to stop. What a row has to contain before it is worth reading at all, and how to read one without turning it into a second inbox, is worked through in the activity log itself.
Being able to answer for it
One more use, and it is the one that counts on the day it counts.
A client rings about something that went out under your name. Why was I chased for a document I sent you in March. Two answers are available to you and only one of them is survivable. The first is that the software did it and you will look into it, which tells a client something worse than the original error, namely that things occur inside your practice for which you cannot account. The second is specific. This is what happened, this is when, this is the reason it happened, and this is what has changed since.
Every firm produces errors and clients have always known it. What they weigh is whether you knew, how quickly you knew, and what you did about it. Responsibility did not move anywhere, and no record moves it: the work went out under your name and you are the one holding the phone. What the record changes is whether that responsibility can be discharged or only apologised for. What you owe a client in writing about any of this, and whether the arrangement belongs in the terms you send them at all, is a separate question answered in what to put in your engagement letter about AI.
What the record is really for
Every row is written after the event. Nothing here prevents anything, and that is worth stating flatly rather than letting the comfort of a thorough record do work it cannot do.
It is less a weakness than a division of labour, and seeing the division is the point of this chapter. Prevention is a setting. The record is the instrument that tells you where the setting ought to sit. Without one, how much a piece of software does on its own is a single act of faith, made once, in advance, with no way of learning afterwards whether it was well placed. That is why the choice tends to collapse into all or nothing, and why it lands on nothing. With one, the move gets much smaller. Raise a single kind of work, read what it did for a fortnight, and either leave it there or put it back where it was.
The three answers so far have each been a different shape. Reading was broad and could be withdrawn. Sending was narrow, sorted by weight, and at the far edge permanent. The record is complete about what was done and silent about what nobody did, which is the absence that gives off no signal rather than anything a log could fix. None of the three is the decision itself. They are the conditions that make one possible.
The decision is the dial, held separately for each kind of work, sitting on review when Point arrives and pointing at everything a mailbox can be handed. It is the only thing in this series you move yourself, and it is the question the whole argument has been walking towards.
So: how far, on what, and what it costs when a setting is one notch higher than it should have been.