Lesson 3.13.1 One rhythm in three parts

In everyday life. You hire a painter. You choose the colors and say which rooms, and you agree that you will walk through the rooms before paying. Then you go to work, and the painter paints the rooms. When you come home, you check the edges and the trim around the doors and windows, and only then pay.

At work. Dave (the controller who runs Brightline's accounting) gives the company's AP Worker one task: reconcile a vendor's statement, and change nothing. He should let it work, then read what comes back before he closes September's accounts.

The book's operating rule has three parts: Humans define intent. AI Workers execute. Humans verify outcomes. With the painter, you define the intent, the painter executes, and you verify the outcome. The 10-80-10 rhythm is that rule with numbers added. The human defines the first 10 percent, the worker executes the middle 80, and the human verifies the final 10.

The pattern is older than AI. The entrepreneur Dan Martell describes it as the 10-80-10 rule for delegating to people: you spend the first 10 percent of a project setting the vision (the outcome you want), a capable team does the middle 80, and you return for the final 10 to refine and approve.¹ It is also often credited to Steve Jobs, though no reliable source confirms that. The first edition of this book applied the pattern to AI, with AI Workers in the place of the team.² The shape stayed the same. What AI adds is in the details: the worker must bring back evidence you can check, flag what is unusual, and wait for a person before any risky action — and it may keep working long after you have gone home. The next pages are about those details.

The table below shows the three parts. The next three pages explain each part and the words in its row.

(The review contract is how you will know the work is right, agreed before it starts.)

Figure 3.1. The 10-80-10 rhythm. You own both ends. The worker owns the middle.

When the worker never stops

The table above reads like one job with a start and an end: you set it up, the worker does it, you check it. But a Standing AI Worker never finishes. The AP Worker builds a register every Monday and answers vendor emails all week, day and night. So where do your two ends go, when the worker is still working at 3 a.m. and you are asleep?

In everyday life. Think of a house cleaner who comes every Tuesday while you are at work. You do not stand at the door each week explaining the job. You wrote it down once: clean these rooms, water the plants, never throw away papers from the desk. If something is unusual — a broken glass, a leak under the sink — the cleaner leaves you a note. If something needs your decision — "should I throw out these old boxes?" — the cleaner asks and waits. And now and then, you look closely at one room to make sure the work is still good.

That is the whole answer. The rhythm still applies to every piece of work, but you no longer do your two parts at the same time as the work:

• Your first 10 percent is done once, in advance. For the cleaner, it is the written instructions. For the AP Worker, it is the Role Contract and its standing instructions. They are the setup for every run, until the work or the rules change.
• Your final 10 percent happens at times you choose, in three ways. An approval queue holds anything the worker may prepare but not do on its own, like the cleaner's question about the boxes. Flags mark anything unusual, like the note about the leak, and you read them at a set time. And spot checks look closely at a sample of ordinary work, like the one room you inspect.

Here is one week of the AP Worker, with Dave as its owner:

Figure 3.2. The rhythm at two speeds. The worker runs whenever work arrives; you review at times you choose.

Notice three things in the week. Nothing risky happened overnight, because the risky step — sending the reply — could not happen without Dave. Nothing unusual got lost, because the worker flagged it and Dave reads flags every morning. And even the work that raised no flags still got a human look, through Friday's spot check.

Each review also improves the next setup. Suppose Thursday's flag shows that the standing instructions never said what to do with statements that arrive in a different currency. Dave adds that case. From next week, every run starts from a better brief. This is the dashed arrow in Figures 3.1 and 3.2, from the final 10 percent back to the first: the end of one round of work becomes the start of the next.

So the worker runs day and night, and Dave works office hours, and the rhythm still holds. Dave made two moves. He moved his first 10 percent to before the week began, by writing the setup down once. And he moved his final 10 percent to fixed times in his own week: approvals and flags each morning, spot checks on Friday. The 4 p.m. run in the story failed because the second move had not been made: the worker flagged the error, but no one had a set time to read its flags.

Why the ends belong to a person

The two ends belong to a person, because judgment, values and accountability cannot be handed over. (Accountability means the work goes out under your name.) The middle belongs to the worker, because that is where execution scales (the work can grow without taking more of your time).

The rhythm is also why this book teaches you to manage an AI Worker before it teaches you to build one. The skills that matter at both ends are defining work well and judging it well, and every route through the book needs them.