The reply that writes itself

For the last couple of weeks a good proportion of my email has arrived with the reply already written. I set up a personal agent that reads my inbox, drafts an answer to anything that needs one and leaves it waiting, so that responding has become a matter of reading a draft and either sending it as it is or changing what needs changing first, and I wrote recently about how much time that saves me. It does save time. What I have been turning over since is what else it saves me from, because some of the effort it has removed was doing a job I had never thought to give it credit for.

The effort of replying used to sit between reading something and answering it, and that gap held more than typing. It held the time it took to calm down and to work out what I actually thought, and sometimes the realisation that the right answer was a phone call, or nothing at all. When the draft is ready before I have finished reading the email it answers, all of that gets squeezed out, and nobody asks me whether I wanted it to be.

Walking away from the machine

There is a piece of advice that has been around for as long as there has been email, and probably for as long as there have been letters: when somebody’s message makes you angry, do not reply until the anger has gone. Walk away from the machine, go and do something else, come back later, ideally tomorrow, and read it again. Most people who have worked in an office have seen what happens when it is ignored, the reply sent in the heat of the moment that reads very differently to the person receiving it, and that cannot be taken back once it has gone.

The advice works because of a fairly simple mechanism. The emotion arrives at once and fades on its own given time, whereas the processing, actually understanding what the other person meant and what the situation needs, takes longer and cannot be hurried. Writing a reply yourself used to take long enough that the first had usually started to fade before you had finished, and the act of composing it forced at least some of the second, because you had to decide what to say sentence by sentence, and somewhere in that deciding you often noticed that you were angrier than the email deserved.

A drafted reply removes the waiting and the composing in one go. It is there within seconds, competent and plausible, at precisely the moment I am least fit to judge it. To be fair to the draft, it is often calmer than I would have been, since the agent has not been insulted by anything; but a calm draft sent by an angry person is still a decision made by an angry person, and the edits I make in that state are usually the ones that put the anger back in. The machine I was always told to walk away from has become the thing that means I no longer have to.

A first response is not a final one

Some messages need thinking about before they need answering, and the thinking is not something I can do in advance or hand to somebody else. A lot of my working out of what I think happens in the writing itself; I start a reply believing one thing and by the third paragraph I have discovered that I believe something slightly different, or that the question I thought I was answering was not quite the one being asked. My first response to anything is rarely my last, and the distance between the two is where most of the useful thinking gets done.

A draft written for me cuts that distance out. Instead of forming a view and then putting it into words, I am reacting to a view that has already been formed, and it is surprisingly hard to read a finished, reasonable paragraph and notice what it has left out. The draft frames the situation, and once something has been framed it takes real effort to see it any other way. The risk is rarely that the draft is bad, because usually it is perfectly fine; it is that it is good enough, and good enough is exactly what stops me thinking any further.

I wrote a while ago about using AI to help me write these posts, and I want this piece to agree with that one rather than quietly contradict it. Help turning my own thinking into sentences a reader can follow is one thing. Having the thinking done for me and then putting my name to it is another, and an agent that drafts replies does both without asking which I wanted.

When every word might be read later

Where I feel this most is in a dispute over building works, the sort of situation where things could turn legal, where every email is potentially evidence and a careless sentence can be quoted back at you months later by somebody whose job is to make it look as bad as possible. Ours involves damp under our floor, a long run of investigations, and the question of whether work done around the house has played a part in it, and nobody has yet established who, if anyone, is responsible.

I use AI to help prepare that correspondence, and it has been useful for keeping apart what was observed, what is only a theory and what a professional has actually found, which is a distinction that matters enormously once a dispute starts and is the first thing to blur when you are frustrated. But the drafts have sometimes come back more abrasive than anything I would have written myself, stating things more firmly than the evidence allows, or taking on a legalistic tone that turns a request into an accusation, and I have had to rein it in and make the point that I am not asking anybody to accept blame without evidence.

The usual worry about AI in an argument is that it will be bland, and here the problem was the opposite. A model drafting a reply in a dispute does not know that what I actually want is the problem properly investigated and fixed rather than an argument won, and it has no reason to care how the people on the other end will feel about dealing with me once the letters stop. All of that is my responsibility once my name is at the bottom, whoever wrote the first draft.

The old advice still applies in those situations; it just has to be applied to the draft rather than to me. Let it be written, then walk away from it, and come back when the feeling has gone to read it the way the person receiving it will, and the way a stranger reading the file a year from now might.

The people who know how I write

The relationships I worry about are the ones outside the family that are carried largely in writing, with colleagues and former colleagues, people I do business with, and friends I mostly keep up with by message. Those relationships are built partly out of how we write to each other. People who have exchanged emails with me for years know how I put things, how long I take to get to the point and what I find funny, and when a message from me suddenly sounds smoother and more balanced than usual, some of them will notice, even if they could not say what has changed. Once somebody suspects a message was not really written by me, they start to wonder about the next one, and the one before, and the warmth in a message is worth very little if the reader is not sure whose warmth it is.

There is a version of this that is slightly absurd and probably not far off, in which my agent drafts a reply to an email their agent drafted, we both press send, and a whole exchange takes place without either of us properly reading or thinking about what the other said. Every message would be polite and reasonable, and the relationship would quietly stop being one.

Some messages only mean anything because somebody spent time on them, an apology, say, or a note to somebody who has lost someone, and with those a well drafted reply is worth less than an awkward one that is actually mine.

If you are using these tools too

I will keep using the agent I have, at least until the commercial ones reach the UK and I can hand the whole thing over to somebody else’s product, and I am not going to pretend the drafts are not useful, because most email does not need deep thought and a lot of it is better answered quickly than carefully. It has earned its place for research as well, where it can use free models rather than the paid plans I would otherwise reach for. What I am trying to do is notice which messages are which before the draft gets the chance to decide for me.

If you have started letting something draft your replies, it is worth asking before you press send whether you would have written it, and whether you have actually thought about it or only read it. If you are angry, the old advice still holds, and it applies to the draft just as much as it ever applied to you: leave it, go and do something else, and read it again once the feeling has passed. And with the people outside your own four walls who mostly know you through what you write to them, the effort is part of the message, and they can usually tell when it is missing, even if they never say so.