Lesson 3.53.5 Three scales: a task, a worker, a company

In everyday life. The rhythm has three parts: someone sets the work, someone does it, and someone checks it. A family does this at three sizes.

• The chore, like one task. You ask a child to wash the dishes (the first 10 percent). The child does it (the middle 80). Then you check the kitchen (the final 10).
• The babysitter, like a worker's whole working life. You agree the rules on the first day (the first 10 percent). The babysitter looks after the children for months of evenings (the middle 80). You check in with them regularly, and change the rules when something keeps going wrong (the final 10).
• The house rules, like the rules a company sets for all its AI Workers. The family sets them once (the first 10 percent). Everyone follows them (the middle 80). The family looks at them again each year (the final 10). For a company, this is called governance.

The same rhythm repeats at three sizes. Only the time span changes, from minutes to years. And the sizes fit inside each other: every evening with the babysitter is a small rhythm of its own, inside the months, inside the house rules.

Figure 3.11. The rhythm at three scales. Each final 10 percent feeds the next first 10 percent.

Notice who holds the two ends at each scale. For a task, it is the person who set the task. For a worker, it is the one person responsible for the worker. For the company, it is its leaders, and the auditors who check for them. The rhythm never runs without a person at both ends. Only the person changes.

A worker's working life

The AP Worker's whole working life follows the same rhythm.

The first 10 percent is its specification. It starts with the Role Contract from the last chapter, which names Dave as the owner and sets the most the worker may ever do. Around it sit the standing instructions for each kind of recurring work, the knowledge it draws on, its inputs, and the tests it must pass. Before the worker does any real work, Dave questions its plan. He also gives it a dry run on last month's invoices (see Lesson 3.2).

The middle 80 percent is the months of work after that. Every Monday the worker builds the invoice register, and it drafts replies to vendor questions as they arrive. Each of those runs is a task with its own small rhythm inside: its setup comes from the specification, it runs, and Dave reviews it at 9 a.m. (see Lesson 3.1). The task rhythm repeats many times inside the worker's middle 80.

The final 10 percent is Dave's monthly review. It steps back from single runs and asks four questions:

• Is the worker meeting its KPIs? For example, are duplicate payments still at zero?
• Do the tests still pass? Rerun them, because a change to the specification, the data or the model can break something that used to work.
• What did this month's corrections say? A correction made twice is a gap in the specification. Fix it in the part where it belongs, and keep the case as a test (see Lesson 3.4).
• Does the specification still describe the job? Vendors change, policies change, and the real job slowly moves away from what the specification says.

Then Dave decides: keep the worker as it is, change it, or retire it. A worker whose specification has not been reviewed in a year is running on an old first 10 percent. Nothing breaks on the day the job changes. The worker simply goes on doing last year's job, carefully and on time.

A company's governance

The whole company follows the same rhythm too.

The first 10 percent is policy. The company's leaders set it, not each user. It says which work may go to AI Workers at all, who owns each worker, which actions always need a person, such as releasing a payment, and which knowledge is official.

At Brightline, the company's policy is short. It fits on one page and has three rules:

• AI Workers may prepare payments, but never release them. Only a person can do that.
• Every AI Worker has one named person who is responsible for it. For the AP Worker, that is Dave.
• The AP policy in the company's knowledge base is the only official copy.

The third rule fixes a problem from the last chapter. There, the assistant used an old copy of the policy, two versions out of date. One user could not fix that alone. Only the company can decide which copy is the official one.

The middle 80 percent is all the company's AI Workers doing their work, each with its own specification and its own rhythm.

The final 10 percent is audit: reading the evidence of what the workers did, reviewing incidents, and deciding which workers to expand, restrict or retire. Suppose Brightline's outside accountant asks at year end: "Show me every payment over $10,000, who approved it, and what the AI Worker recommended." Audit only works if every worker left evidence behind. A company cannot audit a worker whose runs left no record.

That is why the rhythm needs three governed layers under it, and the next chapter introduces them. KSoR, the Knowledge System of Record, governs what an AI Worker may know. DSoR, the Data System of Record, governs what it may do and keeps the evidence of each action. SSoR, the State System of Record, keeps one case file for each piece of work: what was tried, what failed, who approved what, and why. Together they make "check it" something an auditor can actually do, not just a hope.