Do less

Nearly every AI product I use works the same way. I go to it, it asks what I want, and I come away with more to do than when I arrived. It is a very capable assistant that needs managing.

For some people that is a gift. Give them a tool that multiplies what they can do and they will do all of it, at three in the morning, planning their sleep around usage limits. A lot of the people building these tools are like that, and so are most of the people around them, or they would like to be. From inside that room it is easy to assume everyone wants more to do.

Most people want less. The appeal of being rich was never really the objects. It is the time back: the queue you skip, the form someone else fills in, the hassle you never hear about. AI could give that to everyone, and mostly it does not, because it is built as a conversation, and a conversation needs you in it.

Take an ordinary job. You want to sell a cabinet.

Every question was reasonable. Together they were the job. I answered them all, and at the end I had a paragraph to paste into a website, after which the real work (the enquiries, the haggling, the collection) was still mine. That is pull: I went to the AI and pulled the help out of it, one answer at a time.

Push runs the other way. The work comes to me already done, and the only part that reaches me is the part that only I can do.

The pricing, the listings, the replies and the arranging all still happened. They just did not happen to me.

Not having to ask

The better version does not wait to be told. Say I am flying somewhere. Booking it is the easy part. The tedious part is the day itself, and especially the day the flight is late. Drag through it, or let it play.

Nothing was asked of me. Three messages told me what had changed, and the claim I would have forgotten to make was filed while I was still in the air. The pull version of that day is an hour on hold to the airline, followed by rearranging everything else myself.

The draft and the button

The current tools do occasionally get this right. Copilot, which I would not otherwise recommend, is at its best when it notices an email that needs a reply, writes the reply, finds the attachment and leaves it there with a Send button. You did not open a chat or explain anything. You read it, change a word if you like, and send.

One action, and it was the one that mattered: deciding that this is what I want to say.

Skilled work

None of this is only for chores. Most professional jobs are a small core of judgement wrapped in a large amount of administration, and the administration is what fills the day. Policy work is a clear case. A consultation closes and four hundred responses come in. The job, the reason someone is paid to do it, is to weigh what was said and decide what changes. Before that can happen, somebody has to read all four hundred.

The empty box is where most AI for professionals stops: a capable reader, waiting for you to work out what to ask. The other version had read everything by the time the consultation closed. The campaign letters were counted once and attributed, the points the paper already answers had replies drafted against the paragraph that answers them, and what was left was the one question on which respondents disagree. That is the job. The rest was reading.

I have been building a drafting and consultation tool for policy teams on this principle for some time: every step that can be prepared is prepared before anyone opens it, and the person is left holding the judgement and the sign-off. It is why I think this is where the field goes.

The count

The test for a tool like this is not what it can do. It is how many things it asked of you, and whether the ones it did ask were worth your attention. Most products would do badly on that test today, and not because the models are not good enough. They were designed by people who like being asked.