Chapter 1 · A Sense of Rupture: What Is Really Going On

Extended content

Liquid part: this text keeps growing after the printed edition.

Monday, twenty minutes and a sheet of paper

Take a sheet of paper and set one condition: only tasks that repeat get in. A one-day feat is no use for this exercise; the matrix measures your ordinary week. Write down ten to fifteen. Draw the chapter's two axes and place the tasks one by one, without thinking about the tool that might do them. Think only about the task.

There is a trick for the vertical axis. Don't ask yourself how serious the error would be in the abstract, because any error looks serious when you picture it in detail. Ask who finds out, and when. If you catch it yourself that same afternoon, the cost is low. If a client catches it three months later, with the contract already signed, that is another matter.

A worked example

Picture a project lead at an installations consultancy. Her list for the week has fourteen tasks. The minutes of the coordination meeting fall without argument into the delegation quadrant: they are well defined, the error shows when you reread them, and nobody loses money over a comma. The comparison of three supplier bids lands in the middle. The mechanical part speeds up, but someone has to know which line item tends to blow up on site, and that is in no bid. It is a hybrid: the model organizes, she decides what gets compared. The call with the owner when a line item does blow up falls in the opposite corner, ambiguous and expensive. The AI can prepare three ways of explaining the problem. Choosing one and standing behind it in front of whoever is paying is her job.

The finding comes with task thirteen, the weekly follow-up email to the client. She had filed it as routine. Placing it, she hesitates: that email fixes in writing what was promised, and a badly placed nuance gets quoted six months later. It was in the wrong quadrant, and delegating it whole "because it's paperwork" would have been the most expensive decision of the week. That is what the matrix teaches: where you were blind about your own work. What the model can do matters least.

Two placement errors

The first is placing by effort. What is hardest to do is not what is costliest to get wrong: a forty-page report can have a closed specification and cheap errors, and a four-line message can be a minefield.

The second is measuring the cost of error by how often you get it right. A task that comes out well almost every time still belongs in the high quadrant if, on the day it comes out badly, someone else pays. When in doubt between two quadrants, put it in the more demanding one and review it in a month, with what you have seen.

Questions to work with your matrix

• Which task on my list has changed quadrant in the past year without my deciding it?
• In which quadrant do I spend most of my time today, and in which did I spend it two years ago?
• How much of what I call "my judgment" is really repetition of procedure?
• If the output goes wrong, who answers for it, and do they know it?
• What claim about AI have I repeated this week without examining it?

A log of claims

For seven days, every time you read or hear that a model "can already" do something, write down four facts: under what conditions, at what error rate, validated by whom, and deployed where. Almost always three of the four will be missing. That absence is information: it tells you how much of what you know about AI is demonstration and how much is operation. At the end of the week you will have a page worth more than most of the articles that produced it.

Moving on

The matrix tells you where judgment decides. What remains is to find out what shape yours is in when it gets to those zones, and Chapter 2 takes that up.