Teaching note (for instructors)

For instructors. Students can read this page, but it works best if they discuss the case first.

What the case is about

In Case 1, Nisha managed two AI tools while she sat in front of them. In this case she owns an AI Worker that works at night. She writes a brief, builds a project with threads, schedules a nightly run, and decides what the worker can reach. On Thursday night, her client asks for the final numbers on time and for updates sent straight from the worker. One line of the bridge is not yet checked. The case ends before she decides.

The accounting is the setting, not the lesson. Students need no accounting knowledge. The lesson is about who does the work, who checks it, who sends it, and under whose name.

Important: the practical behind this case has not yet been run. Until it is, this is a fictional discussion case, not a tested demonstration of an AI Worker. Exhibits 3, 5 and 8 are illustrative. Tell students this at the start. Ask them to reason from the design of the setup, and to treat the illustrative numbers as "what might happen", not as evidence. When the practical is recorded, replace these exhibits and this note with the recorded results.

Learning objectives

After the discussion, students should be able to:

1. Explain what changes when work moves from "while I watch" to "while I sleep": the brief must stand alone, the record must live outside the desk, and the person checks results afterwards (Exhibits 2, 3 and 8).
2. Separate rules that are enforced by a setting or a missing permission from rules that are only written down (Exhibit 6).
3. Describe what several reviewers add, and what only the person with the client relationship can check (Exhibit 5).
4. Explain whose name a worker's actions appear under, and what a business would need to add (Exhibit 7 and the story).
5. Make and defend a decision under a deadline, with the risks named (core question 5).
6. Say what must be true before professional work is released, and who may release it (Exhibit 10).

Before class

If students will do the lab, set this order: "A paid job" and "Building the AI Worker", then the lab, then "Three nights of work", "Thursday night" and the exhibits. The later pages show what Nisha found, so reading them first spoils the count of who caught what. The answer key is not in the learner download. The lab links to it at the end of Part 7, so ask students to open it only then. Then ask students to read the four story pages and skim the exhibits. Ask each student to choose one of Nisha's three drafts before class, and to write one sentence on why.

Discussion plan (75 minutes)

For an 80-minute class, add series question 11 after core 3, for about 5 minutes.

Board plan

Use three columns.

Decision layout for core 5

Notes on the core questions

Core 1. Students often say "she saved time". Push them: did she? Her own time rose (Exhibit 8 is illustrative, but the design shows why: setup, morning checks, filtering, emails). What she gained was work done while she was away, and a record. What she gave up was being present when each decision was made.

Core 2. Most students will list the rule "Only I send things to the client" as her main safeguard. Ask what would happen if she had connected her email. The rule would still be only words. The missing email app is what enforces it. Also point out the "only asked" rows in Exhibit 6: nothing stops a thread from saving a changed copy of the client's file. Also ask about the thread that continued on Thursday: did it pick up Wednesday's two new points? Changes to instructions reach new threads, not threads already running, and a run in a project can continue a thread. Version 2 shows both points, so here it did, but only because Wednesday night's run had started a new thread. Optional question 7 picks this up.

Core 3. Exhibit 5 is illustrative, so focus on the design. Different reviewers saw different things (the Who saw what figure in Exhibit 5): the area threads and the reviewer thread saw the client's files in the Library, but each area thread saw only its own part of the bridge. ChatGPT saw the rules and the bridge, but not the files. Only Nisha saw the emails with Sutlej and Mark and could write to the client. Finding a gap is not the same as closing it: the reviewer thread listed the inventory gap and the rent figures, but it could not say which figures were right. Only the person with the relationship could ask Sutlej which rent figure was right.

Core 4. Anything a scheduled task does through connectors appears as the person, on every plan. Ask: would Mark be able to tell what Nisha had read? Link to Claude Tag (series question 13) and to what a business adds: its own identity for the worker, approvals recorded in a system, and a lasting record.

Core 5. There is no right answer. Strong answers name what each draft protects and what it risks, and say what Nisha would tell Mark in one sentence. Weak answers choose by feeling. Push students who choose draft 1 on the nightly emails. Push students who choose draft 2 on the client relationship.

Then push every group through Exhibit 10, the acceptance checklist. The board is deciding whether to invest. Ask: is this bridge a preliminary scenario analysis or a signed-off conversion? The lease contract is open, and it could change the balance sheet as well as the profit line. The inventory difference is explained only by an email. Who has the US GAAP expertise and the authority to release a signed-off conversion? Nisha is still learning US GAAP. A strong answer labels whatever leaves as "Preliminary scenario analysis, not a signed-off conversion" and names who would have to sign before it is called final. This is a discussion of professional judgment, not accounting advice.

Takeaways for the wrap-up

• A worker that runs while you sleep needs a brief a stranger could follow, and a record outside the desk.
• The safest rules are the ones a setting or a missing permission enforces. Written rules still matter, but they are not locks.
• More reviewers find more, but only the person who owns the client relationship can check some things.
• Through your connectors, the worker acts under your name. That is why the second signature matters most for anything that leaves the company.

Running the class with non-accountants

Tell students at the start that they do not need to understand the accounting. Explain a bridge as "a list of adjustments from one rulebook to another". Treat each line of Exhibit 4 as "one rule difference with an amount". The lease line is enough to understand as "one of two answers, and only a contract she has not seen can say which". Keep the discussion on the decisions, not the standards.

Notes on the series questions

• 11, where the work lived: the Library and Nisha's email held the lasting record. Threads and their desks did not. Deleting a project removes its threads, memory and files. Push further: a Library is right for drafts and learning, but it is not a governed business record. A business AI Worker keeps progress, approvals and accepted results in records with named owners and a history of every change. See From personal worker to business worker.
• 12, containment or authority: the threads were contained (they could not send). Mark was protected by Nisha's authority over what was sent.
• 13, a worker with its own badge: Claude Tag is in public beta on Team and Enterprise plans. An Owner or a Claude Tag admin sets up its connections for a workspace or a channel, and Anthropic's security page for Tag covers audit trails. In a one-to-one message, Tag runs on the member's own Claude account, and in a channel it can use a member's own connectors for that member's requests, after the member allows it.¹ Rebuilt projects are not yet offered on Team or Enterprise plans, so moving to a team plan for Claude Tag would cost Nisha her project setup. Even with Tag, nothing by itself gives her the client's business record or an enforced approval before anything is sent to Mark.

Grading the lab (10 points)

Use the same rubric that is on the lab page.

The answer key lists the eight items, the expected figures and how to mark the lock test and the checklist.

Give full points for honest records of what went wrong, such as a run that stopped or a lock that failed its test. Take points off for claims with no record behind them.

The final assessment

After both cases, set the final assessment. It gives learners a new job with no accounting and only made-up data, and asks them to plan an AI Worker on their own: what it may execute and what it may only draft, where state is saved, which permissions are locked and how they test the lock, and what evidence a person needs before approving. It tests whether the lessons transfer, not whether students can repeat Nisha's steps. Model answers and marking notes are in the model answers file. The final assessment links to it at its end, for learners who study alone.

Running the practical and using it in class

Use the practical run sheet to run the exercise before teaching, and replace the illustrative exhibits. Classes can also run it themselves with the lab, which uses the ready-made data pack in labs/nisha-lab-case-2/. The click-by-click steps and the fallbacks for learners without projects are kept in one place, the setup guide, with the date it was last verified. Check it before each class. The planted errors are listed in the answer key, and shown in the story pages that learners read after the lab.

Facts and fiction in the case

The accounting differences are real and sourced. The bridge is a teaching example and needs an accountant's check.