TL;DR:
Sending, paying, deleting and sharing are the actions that are hard or impossible to take back.
An agent should show you what it is about to do and wait for a clear yes.
Passwords and card numbers are for you to type, always.
Letting an agent work in your apps is useful because it saves you steps. But some steps are special. Once they happen, there is no taking them back, or they affect other people, or they involve your money. For those, the right setting is a person in the loop. Not because the agent is bad, but because the cost of one wrong guess is too high.
This article lists those actions, explains what makes them different, and shows what a good approval looks like.
What makes an action need a person
Ask three questions about any step. Can it be undone? Does it reach other people or other systems? Does it cost money or give away something private? If the answer to any is yes, it deserves a pause. The rest, like reading, sorting and drafting, can run on its own.
Walk-through: a list of actions that need a person
Action | Why it needs a person | What the agent should show you |
|---|---|---|
Sending a message or email | Others receive it and cannot unsend it | Recipient, subject and the full text |
Paying or buying | It moves your money | Item, price, total and who is paid |
Deleting files or messages | May be permanent | Exactly what, and how many |
Sharing a file or changing who can see it | Exposes private material | The file and the people or link |
Accepting terms or consent | It is a commitment in your name | The terms or a summary, and the choice |
Posting in public | Anyone can see it | The post and where it goes |
Changing account settings | May lock you out or change security | Which setting and the new value |
Look at the third column. A good approval is not "OK to continue?" It is a concrete preview that lets you decide in a few seconds. The more specific the preview, the better the decision.
Things only you should type
Some entries are not even on the approval list, because the agent should never do them. Passwords and card numbers are typed by you. If a task reaches a sign-in or payment screen, the agent stops and hands the keyboard back. See why an assistant should never type passwords.
What a good approval looks like
Specific. It shows the actual message, price or file, not a summary.
Before, not after. The action has not happened yet.
Easy to refuse. Saying no is as easy as saying yes.
Easy to edit. You can change the text or recipient and then approve.
Recorded. The record shows what you approved.
Why not ask about everything
A tool that asks for permission at every click soon trains you to click yes without reading. That is worse than asking less often. So approvals are saved for the actions in the table, and everything else runs quietly. A small number of meaningful questions gets read; a flood does not.
Watch out: If an agent ever offers to skip approvals for sends or payments "to save time", decline. The time saved is seconds, and the risk it removes protection from is large.
A worked decision: should this one ask?
Take a case that is not on the table: "rename these twelve files". Can it be undone? Yes, you can rename them back, and a record of old and new names makes that easy. Does it reach other people? No, unless the folder is shared. Does it cost money or expose anything? No. So it can run without a pause, provided the agent keeps the list of old and new names. Compare "empty the trash": it cannot be undone, so it deserves a question.
Walking a new action through the three questions takes ten seconds, and it is how you decide where your own approval points should go.
Keep in mind that the answer to these questions can depend on context. Deleting a draft you made five minutes ago is minor. Deleting a shared folder is not. The same verb can be cheap or costly, so the agent should look at what it is acting on, not only the action.
Why this also protects against tricks
Text on a web page or in an email can try to steer an agent into doing something harmful. If sending and paying always need your yes, a planted instruction still has to get past you. Read why a web page can try to give an AI agent orders for how that works.
Privacy and control
This is how DuckyHelper is meant to work. It's built to ask first before it sends a message or pays for anything, and not to type your passwords or card numbers. Its agent features, such as working in your apps, are in early access.
Try it yourself
Look at the tools you let act for you, such as email rules, auto-pay and shared-folder settings. For each, ask if a mistake could be undone. Where the answer is no, check whether a person must approve it.
Frequently asked questions
Does asking first make an agent useless?
No. Most of the work, like reading, sorting and drafting, needs no approval. The pause comes only before the risky step.
Can I allow some actions automatically?
That is your call, but keep a person in the loop for anything that cannot be undone.
What if I approve something by mistake?
Act quickly to undo or contain it, and tell anyone affected.
Get started
