Notes: Delegate more without creating another job to manage

Victor's week shows the problem. He is an operations manager in a prepared example, and he set up several AI prompts to summarize project threads, draft weekly status reports and flag schedule risks across his department.

The drafts arrived in seconds. By Friday afternoon, though, he was comparing 3 conflicting drafts and hunting down outdated numbers, and that took longer than writing the update himself used to.

The work he still carried came in three parts:

1. Explaining the context the drafts were missing.
2. Reconciling the different versions against each other.
3. Fixing claims that did not match his project records.

More agents do not prove a better workflow. If the output leaves you supervising and cleaning up after it, the speed of generation means very little.

For Victor, useful delegation would look like this:

• One clear summary to decide from.
• A plain reason beside any uncertain item that comes back to him.
• Final authority over consequential actions staying with him.

So before you add another layer of automation to a recurring job, I'd test one thing first: does the change lower your combined explanation, review and correction effort? The next 4 lectures build the controls that keep that number down.

Maya wants her phased Meridian proposal pressure-tested before she submits the bid. She has the Meridian scope document, her team's capacity record, her 4 evaluation criteria and an open question about specialist hours.

She can see two weak ways to do it:

• One chat builds the case and attacks it. The model can anchor on her preferred answer and offer only mild criticism.
• Two chats she compares herself. She ends up reading 2 long essays line by line, which is Victor's reconciliation burden again.

So she asks her usual agent in Work to act as the coordinator. It prepares 2 reviewer briefs for her to confirm, and each brief gives a separate agent one bounded review job.

Both subagents receive the same records and criteria. Only one receives Maya's preferred answer:

1. The Case-For Subagent also gets the phased proposal, and builds the strongest defensible case for it.
2. The Independent Evidence Subagent never sees the preferred answer. It evaluates the evidence from scratch and looks for facts in the records that could overturn the proposal.

Withholding your preferred answer reduces one source of confirmation bias. It never guarantees unbiased judgment. When you set this up yourself, check what each subagent can see, because the second review only stays independent if it cannot read your main conversation.

Maya tells the coordinator not to average the two reviews, and to return 4 sections instead:

1. Where the reviewers agree.
2. Where they materially disagree.
3. What evidence is missing.
4. Which uncertainty needs her judgment.

In the prepared case, both reviewers agree that the full 6-week bid is too risky. They disagree on week 5: one treats 20 hours of specialist time as available, and the other points to a calendar footnote showing an internal platform upgrade that same week. Averaging those two would hide that risk.

Because the conflict was kept, Maya sees in 30 seconds why week 5 is uncertain and which engineering lead to call. In ChatGPT Work, subagents are available on eligible accounts, and several agents running at once use more of your plan than one.

Start from receipts, never from trust. In Section 1, Alex's Monday preparation reminder stayed as Proposed only. Before you grant any advance authority, review what the agent actually did on earlier runs. The Northstar reminders are a prepared example, so no real receipts sit behind them. In your own work, you set the rule only after checking receipts from your own workspace.

Alex has now checked several weeks of reminder proposals and wants the 15-minute preparation block booked without a click every Friday. Alex describes the routine, and the agent proposes a rule for Alex to confirm.

The rule has 4 parts:

1. Action and destination. One 15-minute private preparation reminder on Alex's personal calendar only, 60 minutes before any confirmed Northstar meeting.
2. Required evidence. The review date and time are explicitly confirmed in the supplied transcript or project record.
3. Reporting. The agent reads back the saved calendar entry, where the calendar tool allows that, and logs it in Alex's weekly summary report.
4. Immediate return. An ambiguous date, a clash with an existing commitment, an unavailable calendar tool, or an event that would invite or notify another person: the agent books nothing and returns the item to Alex at once.

Compare that with "You have permission to manage my meeting follow-ups from now on." A broad grant like that blurs a private note to yourself with an unreviewed commitment email to a client.

Alex tests the rule on two cases from the prepared illustration:

• Monday. The review is confirmed for 11 in the morning and the 10 o'clock slot is open, so the agent books the reminder, checks it on the calendar and lists the receipt in the weekly report.
• Tuesday. A department budget call already sits in the 10 o'clock slot, so the conflict rule fires: nothing is booked, and the conflict is flagged at the top of Alex's follow-through page.

A clean streak, whether 10 runs or 30 days, raises your confidence and still guarantees nothing about the next run. When quarter-end makes the calendar change daily, Alex withdraws the permission, every reminder goes back to Proposed only, and Alex checks that the next run obeys the tighter boundary.

Maya has an afternoon of client visits across town. She has been working on her Meridian proposal in a local project folder on her office laptop, and she wants to keep going between meetings.

If she shuts the laptop and leaves, her phone still shows the conversation, as long as she has turned on sync. Nothing can read the Meridian scope PDF or the saved decision record, though, because those files sit on a laptop that is now offline.

Her company has local computer access set up and she has sync on, so she does this instead:

1. She keeps the laptop switched on and online.
2. She starts a new cloud task for the next step. A task begun on the laptop cannot switch to the cloud partway through.
3. Between client visits, she opens that cloud task on her phone and asks it to check the phased proposal against the latest capacity note.
4. The cloud task reads the files on her connected laptop and brings back the answer.

A dot connected to the same laptop could do the same check from its own cloud computer. Dots are not on every plan or in every country. As of 20 October 2026, Pro plans in the European Economic Area, the UK and Switzerland do not include them, so check what your own ChatGPT screen shows.

A linked app needs two approvals. Appearing in ChatGPT's list makes it technically possible. Your employer's policy decides whether company documents may go through it.

Before relying on any away-from-desk route, Maya runs a 3-part check:

1. Does the task run locally or in the cloud, and is the laptop connected?
2. Can this route read the specific proposal and its current decision record?
3. Which linked apps does her organization approve for client bid documents?

If her company allows neither a connected laptop nor a linked storage app for client bid files, she does not upload them to a personal account. Before leaving, she has the agent prepare a permitted, non-confidential summary in her employer-approved mobile workspace. The same holds if your workplace uses a different AI product on mobile: the method and the permitted summary carry over.

Use a job you already noticed. If a recurring job in your week came to mind while you followed the examples, this is the place for it. It might use Meetings and follow-through, Priorities and decisions, or Communication and people leadership.

In the prepared capstone, Alex runs the meeting follow-through workflow on the next weekly Northstar check-in. Alex supplies the permitted transcript and a one-sentence note on Monday's goal. The agent pulls the Northstar Onboarding project records, checks the confirmed purpose and drafts the blocker-focused follow-ups in the style Alex's project notes describe.

The standing permission is still paused from quarter-end, so the agent proposes the Monday reminder and waits for Alex to confirm it.

After reading the result against Monday's purpose, there are two useful moves:

• A correction found. Ask the agent to capture that refinement into the skill.
• A clean result. Ask it to confirm what made the method work, so those rules stay saved.

Count the whole cost. Counting only the 20 seconds the software spent generating the page hides the 25 minutes spent hunting files, explaining background, reading and fixing.

Alex's figures are made up for this example:

1. Initial setup: about an hour in week one, for the project records and the meeting-follow-up skill.
2. Task explanation: about 10 minutes a week before, under a minute now.
3. Output review: about 15 minutes before, 3 minutes on one organized page now.
4. Manual correction: about 15 minutes before, about 2 minutes for one phrasing tweak now.

That saves more than half an hour a week, so the setup pays back within 2 weeks, and Alex keeps the workflow. One run shows how the workflow did on this task, and it promises nothing about every future week.

Your own result can go 3 ways: keep it, refine the skill where correction was still high, or stop using AI for that task when supervision costs more than it saves. If a colleague later asks for a proven workflow, share only what that one job needs. That means the saved skill, one permitted worked example and the rules for what comes back to a person, with a note that the sender reviews before sending, using employer-approved tools.

You started this course knowing AI could answer a quick question or polish a paragraph. The work that crowds a leader's week still fell back on you: chasing follow-ups, repeating background, weighing risky tradeoffs, writing messages you can stand behind and preparing difficult conversations.

Across six sections, you can now recognize and explain each of these:

1. Sections 1 and 2. An agent turned a transcript and your stated purpose into notes and drafts, and a project and a skill saved the background and the method.
2. Section 3. An attention queue and a decision record kept your choices and their reasons clear.
3. Section 4. A voice skill and spoken rehearsal kept your messages recognizably yours.
4. Section 5. Prepared observations and open questions kept your people conversations fair.
5. Section 6. Subagents, standing permissions with clear limits, checked away-from-desk routes and a complete-effort comparison let you delegate larger jobs without a second job to manage.

That answers Victor's Friday afternoon: one clear summary to decide from instead of 3 conflicting drafts. None of it needed code, because every setup step started with you explaining the work in plain words.

Think back to the client who asks whether your team can still start their project on the agreed date. You can now give an agent the notes from your last project meeting and say what the answer has to settle. It brings back what is ready, what is blocking the start, and drafts asking your team for the missing facts, so your attention stays on the decision and the client.

The principle, you lead the work and the agent prepares it, becomes 4 questions for any new demand:

1. What result would help the next decision?
2. What grounded context and methods does the agent need?
3. Which exceptions or conflicts must return to me immediately?
4. Which human judgment or conversation belongs only to me?

The software prepares the ground. The purpose, the tradeoff and the care for the people involved stay with you.

Your first step: pick one meeting this week whose follow-up you usually end up chasing, run the meeting follow-through workflow on its transcript, and judge the result with the keep-or-change test from the first section.