Chapter 10. Governance and Responsible Use

The point

This chapter is about governance: the decisions that let you defend how you use an AI Worker, an AI that holds a role at work. Governance gives each decision a named owner and a written answer, before the work starts.

Chapters 5 to 9 made the worker useful. But useful work can still be the wrong work, use the wrong data or reach the wrong person. Governance is not a way of saying no. Its goal is "yes, on these terms," with terms someone can check later.

The chapter answers seven questions, with one or two concepts for each:

Question	Concept	After it, you can
Should a worker do this at all?	10.1	Decide if a worker may do it, may do it with a person's check, or may not, and name what decides it
Can this data go in, and by which route (the way it reaches the AI)?	10.2, 10.3	Decide what data a task needs, and check what each AI vendor's privacy controls do and do not decide
What rules already apply?	10.4	Follow your organization's AI policy, or propose an interim (temporary) one when there is none
Who could be harmed or left out?	10.5	Check who an output affects, including people who never see it, and decide whether to say that AI was involved
Who owns what the worker knows?	10.6	Give knowledge an owner, and mark it as replaced (deprecate it) or take it down
Who must sign before it leaves?	10.7	Name the outputs a person must sign, and build a gate, a check that the worker cannot skip
What happens when it goes wrong?	10.8	Handle the first hour after something goes wrong
Which settings carry out the policy?	10.9	Find the settings that carry out one AI policy on each AI vendor, and the three differences that change your setup

The lab, the hands-on part of the chapter, turns the answers into documents a manager can use. A Governance Record is one page with the answers for one workflow. A workflow is the series of steps a team follows to finish a piece of work. The others are a sign-off matrix, a table of who signs and who sends each output, a takedown record, an incident note and an interim AI policy. This chapter covers Domain 6 of the CCAO-F exam, Anthropic's Claude Certified Associate: Foundations. The chapter is a method for the people who do the work. It does not replace your organization's policy, or advice about law, privacy or compliance.

This chapter deepens Diligence, one of the 4Ds of AI fluency: Delegation, Description, Discernment and Diligence.

Why it matters

By December 2026, the AP Worker was part of how Brightline Wholesale Supply, the book's example company, worked. AP means accounts payable, the bills a company owes. The worker answered policy questions from five approved concepts (Chapter 8), short files that each hold the rules for one topic. It prepared the Thursday pre-run note, the summary the controller reads before he approves each Friday's payment run. It drafted replies to vendors, and Maria, the office manager, sent them from the AP inbox. She ran the worker day to day. Dave Kowalski, the controller, approved what it worked on.

In the first nine days of December, four ordinary decisions went wrong. At the time, nobody had seen any of them as a decision.

Tuesday, December 1. Dave asked the worker for "a list of vendors worth paying early for the 2 percent discount, strategic vendors first." It returned a list of 14 large suppliers. Dave approved the list. Nobody looked at the vendors it left out. Brightline promises to pay small suppliers within 15 days of receiving their invoice. Three of them were already past that date, and the list delayed their payment by another week.

Thursday, December 3. The auditors, the outside company that checks Brightline's accounts, asked for a schedule of open payables, a list of the bills Brightline still owed on November 30. They wanted it signed by the controller. The worker built it from the register, Brightline's list of invoices, and drafted the email. At the bottom it put Dave's signature block, his name and title, as if he had signed it. Maria sent it from the AP inbox, like the worker's other drafts. Dave never saw it. It listed the balance owed to Northern Maple Paper, a Canadian supplier, of CAD 18,400.00 (Canadian dollars) as USD 18,400.00 (US dollars).

Sunday, December 6. Maria needed to find duplicate vendors, the same vendor listed twice, before the year-end tax forms. The company's AI workspace felt slow. So at home, she uploaded the full vendor master to her personal AI account. The vendor master is the main file of all vendors and their details. It held bank account numbers. It also held the Social Security numbers, used as tax IDs, of 23 sole-proprietor contractors, people who each run a business on their own. Her personal account was set to allow its chats to be used for training, to improve the AI vendor's models.

Wednesday, December 9. Jordan, a clerk hired to help at year-end, asked the AI in the new-staff project, a shared workspace in an AI product, how to handle a vendor's new bank details. It answered from onboarding notes, written for new staff in 2025: "Confirm by replying to the vendor's email." Jordan did. The AP policy's callback rule stopped the payment: a change waits until someone phones the vendor on the number already in the vendor record. The notes had been removed from the AP Worker's project in November (Chapter 8). But the copy in the new-staff project had no owner, so nobody had withdrawn it.

By Friday, December 11, Dave knew about all four. His first answer was "Turn it all off." That would have stopped the useful work as well as the harm, and left the four decisions as they were. This chapter gives another way: ask the seven questions above before any work that matters. Each failure in December skipped at least one of them. The early-pay list skipped the question of who it would leave out (10.5). The upload skipped the data, route and rules questions (10.2 to 10.4). The schedule skipped the question of who must sign (10.7). The notes had no owner (10.6).