Everything so far has been about getting an agent to do more. This lesson is about staying firmly in charge while it does. Good news: it comes down to one simple rule you can hold in your head, not a security course.
The two-list rule
Sort every action into one of two lists. On the left, things an agent may do on its own. On the right, things that always need your click — no matter how much you trust it, no matter how good it's been. This single sort prevents almost every way an agent can embarrass or cost you.
May do alone
- Draft (emails, posts, copy)
- Summarize and brief
- Organize and categorize
- Research and compare
- Read your files
- Propose plans and options
Always needs your click
- Send (email, messages)
- Publish or post
- Pay or purchase anything
- Delete anything
- Promise or commit on your behalf
- Share private data
Notice the pattern: the left list is reversible, the right list is public or permanent. A bad draft costs you nothing — you just don't send it. A bad send, a wrong payment, or a deleted file costs you real trust, money, or work. So the right-hand actions stay on your side of the desk, always. Every job card in this course already bakes this in with the line "draft, don't send."
Data privacy, in plain terms
Two habits cover most of it. First, don't hand an agent private information it doesn't need — client Social Security numbers, unreleased financials, other people's personal details. If a job doesn't require it, keep it out. Second, only connect tools you're willing to have read. Connecting your email gives the agent your email; that's fine for an inbox summarizer you trust, and not fine on a whim. When in doubt, give less access and add more only when a real task needs it.
Reviewing like a manager
Here's the shift that makes agents actually save time: spot-check, don't re-do. A nervous new manager redoes their hire's work and saves nothing. A good one samples it — reads two of the ten drafts, clicks a couple of the cited sources, checks the totals against a few receipts — and trusts the pattern if the sample holds. You're reviewing to catch problems, not to re-perform the task. If the sample's clean, ship it.
It will, sometimes. The manager's move isn't to fix the output and move on — it's to fix the job card. Wrong tone? Your voice guide or a "match this" line was missing. Touched the wrong files? Your inputs weren't fenced tightly enough. Went too far? Add a rule. That's how an agent "learns" under your management: not on its own, but through a sharper brief next time. The output is today's problem; the job card is every future day's fix.
The 80% rule
Hold this whenever you're tempted to do it yourself "because it'll be better": an agent that gets you 80% of the way in five minutes beats your perfect version in two hours — because the perfect version usually never gets made. The goal isn't a flawless draft from the machine; it's a strong draft in minutes that you finish with your judgment. You stay the editor, the boss, the taste. The agent just clears the blank page and the busywork so your 20% is all that's left to give.
Let AI do the reversible work alone; keep the permanent and public work on your click; review by sampling, not redoing; and when it's wrong, fix the brief. Do that, and you can delegate freely without ever handing over the wheel.
- Look at the job card you filled in Lesson 3 or 4.
- Check every action it lets the agent take against the two-list rule. Anything on the right-hand list? Rewrite it as "draft it and show me."
- Confirm the inputs are fenced to only what the job needs — nothing more.