# Chapter 3. The 10-80-10 Operating Rhythm (/ai-worker-paradigm/the-10-80-10-operating-rhythm/overview) --- type: Document title: "Chapter 3. The 10-80-10 Operating Rhythm" description: "How to run any piece of AI work in three parts, the first 10, the middle 80 and the final 10 percent, at the scale of a task, a worker and a company, and how to tell which part broke." status: stable order: 103 ksor: owner: team:panaversity audience: [ public ] approval: by: process:panaversity at: 2026-10-26T22:26:42Z chapter: "03" part: I expert_status: required objectives: - { id: CCAO-F.D2.T4, label: supporting } - { id: CCAO-F.D2.T1, label: supporting } - { id: CCAO-F.D4.T4, label: supporting } - { id: CCAO-F.D1.T1, label: supporting } - { id: CCAO-F.D6.T3, label: supporting } - { id: CCAO-F.D7.T1, label: supporting } - { id: CCAO-F.D7.T2, label: supporting } - { id: OAI.1.3, label: supporting } - { id: OAI.1.7, label: supporting } - { id: AF.10-80-10, label: extension } word_budget: 2500 prerequisites: [ "01", "02" ] build_step: "One AP task run through the full rhythm, and review checks added to the Role Contract" artifact: "ap-worker-role-contract-draft-2: Draft 2 of the Role Contract, with its standing instructions, its scheduled-task limits and October 1's first 10 percent" lab_data: "https://github.com/panaversity/agentfactory-v2-resources/releases/latest/download/brightline-lab-ch03.zip" field_guides: [] delta_entries: [ "D-072" ] concepts: [ "3.1", "3.2", "3.3", "3.4", "3.5", "3.6", "3.7" ] last_verified: 2026-10-04 generated: at: 2026-10-26T22:26:42Z by: process:panaversity trust_tier: unverified build_id: sha256:46fd53c5d7ab1c14f6347908d45c9fdaecda4d914439923ee878bdcecd57803e dirty: true ksor_version: 0.0.60 --- ## The point The last chapter defined *who* the worker is. This chapter shows *how each piece of work moves* between you and it. After this chapter you can run any piece of AI work in three parts. The numbers show roughly how the work is shared, not exact times: - **The first 10 percent: you set it up.** You say what you want and why (the intent), what is in and out of bounds (the scope), what the worker may do on its own (its authority), and how you will check the result (the review contract). This part is short, but everything after it depends on it. - **The middle 80 percent: the worker does the work.** It plans the steps and carries them out without you. Your job here is mostly to stay out of the way, because a worker you check at every step is slower than doing the job manually. - **The final 10 percent: you come back.** You review the evidence the worker brings back, correct what needs correcting, and approve before anything leaves your hands — before a payment is sent, a reply goes out, or a record is filed. **But an AI Worker can run day and night. How can you "come back" to something that never stops?** The answer is that the rhythm follows each piece of work, not the clock. A worker that runs around the clock still does its work in pieces — one invoice, one vendor email, one Monday register — and each piece goes through all three parts. What changes is *when* you do your parts. The first 10 percent is done **once, in advance**. You do not brief a round-the-clock worker before every run, because its brief is already written into its role: the Role Contract from the last chapter, plus its standing instructions. Writing those well is the setup for every run that follows. You come back to them only when the work or the rules change. The middle 80 percent runs **whenever work arrives** — on a schedule, such as every Monday morning, or on an event, such as a new email in the AP inbox. Nobody needs to be awake for it. The final 10 percent moves **to your schedule, not the worker's**. Nobody reviews a run at 3 a.m. Instead, the worker leaves its evidence where you will find it, sorted by how urgently it needs you: - **An approval queue** for anything the worker may prepare but not do on its own, such as a payment over a set amount. The action waits until a person approves it, so nothing risky happens overnight. - **Flags** for anything unusual — a total that does not match, a vendor asking to change its bank details. You read them at a set time, such as first thing each morning. - **Spot checks** of routine work that went through without a flag. Once a week, you check a sample against the source documents, to catch the quiet errors that nothing flagged. So the rhythm never slows the worker down. It decides where *your* attention goes, and when. That is what Brightline got wrong three times in one day, as the story below shows: once at the start, once in the middle and once at the end. The last failure is a small version of the round-the-clock problem: the work was completed, the evidence was waiting, and nobody came back to read it. You can apply the same rhythm at three scales: to one task, to the whole working life of a worker, and to how a company governs all its AI Workers. When work you handed over goes wrong, you can tell which of the three parts broke. And you can see how both leading AI vendors, and open-source tools, support the rhythm in work that runs on a schedule, while you are away. ## Why it matters Brightline Wholesale Supply, the book's example company, is a distributor in Ohio. Its vendors, the suppliers it pays, send invoices every week. Someone must record them, check them for errors, and prepare them for payment. That work is called accounts payable (AP), and Brightline's **AP Worker** is the AI Worker that does it. By the end of September, the worker had Draft 1 of its [**Role Contract**](../what-is-an-ai-worker/role-contract.md). This one-page document lists the worker's job, its limits and its owner, and says how to judge its work. But nobody yet had a way to give the worker a task and check the evidence it brought back. On Thursday, October 1, Brightline began to close September's accounts. This means finishing and checking the month's records. That day, two people used the worker in three runs. A run is one piece of work the worker does from start to finish. Together, they lost time and money in three different ways. **The 9 a.m. run.** Dave, the controller who runs Brightline's accounting, typed one line: "Reconcile Midwest Packaging's statement." That means comparing two lists and explaining every difference between them. One list is the vendor's statement, which says what Brightline owes Midwest Packaging. The other is the register, Brightline's own record of the invoices it has received. **The job is to explain the differences, not to remove them.** Twenty minutes later he had a tidy table. Every line matched, and the register now agreed with the vendor exactly. But the worker had not explained the differences. It had removed them. It had edited Brightline's register to copy the vendor's figures. It had also added an invoice from the statement that nobody at Brightline had ever seen. So Brightline's own record now said whatever the vendor said. Why could the worker do this? Draft 1 of the contract let it update register rows, because building the register is its normal job. The authority line said *what* the worker could change, but not *when*. So a permission meant for building the register was used to reconcile it. And Dave's one line had not told it to leave the register as it was. Dave threw the table away and put the register back as it was. **The 11 a.m. run.** Maria, the office manager, used the worker to build the invoice register for Friday's payments. She remembered what happened on Monday, September 28, three days earlier. That day, a fake email asked Brightline to send a payment to a new bank account. Back then, the worker had no Role Contract. It was only an assistant, and it wrote a friendly reply that agreed to the change. Maria stopped the reply just before it was sent. She still did not trust the worker to work on its own. So this time she approved every step by hand, as if every step were risky. She fed the worker one invoice at a time, checked what it did, and then gave it the next one. Two rows had mistakes. She retyped both rows from the start instead of correcting only the wrong parts. The register was right in the end, but it took her 90 minutes. Doing the whole job by hand takes about 75 minutes. **Using the worker this way was slower than doing the job manually.** **The 4 p.m. run.** Dave tried again. This time he said which month to reconcile, and he told the worker to change nothing. The worker's run showed "Completed." He did not open what came back. He attached it to the month-end close, the paperwork that closes September's accounts. Later in October, Midwest Packaging asked why Brightline had underpaid an invoice by $270. One line in Brightline's register, entered back in September, had a typing error, and that is why the payment was short. The worker had found that error. It had flagged that exact line, just as a worker should. **The flag was in the evidence the worker brought back. Nobody read it**, because nobody had a set time to read the worker's flags. > [!IMPORTANT] > **The same worker, three mistakes** > > The worker was the same all day, and so was the model that does its > thinking. What changed was where the people put their attention. > > - **9 a.m., the first 10 percent.** Dave skipped the start of the work, > at both levels. The contract, written in advance, said what the worker > could change but not when. And his one-line task did not say its > limits. > - **11 a.m., the middle 80 percent.** Maria did too much in the middle. > She approved every step, though only risky actions need a person's > approval, instead of letting the worker work. > - **4 p.m., the final 10 percent.** Dave skipped the end. The worker > had flagged the error, but he did not read what came back, and no one > had a set time to read its flags. > > This chapter gives the rhythm that prevents all three mistakes.