30% off your first 3 months

Get the deal

Why an AI agent should ask before it sends or pays

by DuckyHelperNov 12, 20263 min read
AI on your computer

TL;DR:

  • Sending and paying are hard or impossible to undo, so an agent should stop and ask first.

  • A good review step shows exactly what will be sent or paid, before it happens.

  • Asking only for these actions keeps the agent useful for everything else.

Most mistakes on a computer can be undone. You can delete a typo, close a tab, restore a file. A sent email and a payment are different: once they leave, they are out of your hands. That difference is why a careful agent treats those two actions separately.

The short answer

An agent should do its work freely where mistakes are reversible, and pause for your approval where they are not. Sending a message and paying are the clearest cases. A pause costs you a few seconds, while a wrong send can cost much more.

What happens step by step

  1. The agent does the preparation: finds the details, writes the draft.

  2. It stops before the irreversible step.

  3. It shows you exactly what will happen.

  4. You approve, edit or deny.

  5. Only after approval does it send or pay.

Walk-through: a drafted email waiting for your OK

You ask: "Tell my teacher I'll miss Thursday's lab because of a dentist appointment."

  • The agent opens your mail, finds your teacher's address and writes a short draft.

  • It shows you the draft with the recipient, subject and message, and waits.

  • You notice it says Wednesday. You edit that to Thursday.

  • You approve, and only then is it sent.

Without the pause, the wrong day goes out. With it, the cost was reading one short message. This is the review step doing its job: you catch what the agent cannot know.

Why only these actions

If an agent asked before every click, you would stop using it. Asking only for sends and payments keeps it quick on everything that can be undone, and careful on the few things that cannot.

What a good review shows: The full recipient, the full message or amount, and a way to change it. Not just a yes/no on a vague summary.

Reversible and irreversible actions

A useful way to sort actions is by whether you can take them back. Typing in a draft: yes. Moving a file to the trash: yes. Sending an email: not really, since the recipient has it. A payment: sometimes refundable, but not by you alone. The harder it is to take back, the more a pause is worth.

What a bad review step looks like

  • A vague summary such as "I will handle your email".

  • A yes/no with no view of the actual message or amount.

  • A prompt that appears so often you stop reading it.

A review step that is easy to ignore protects nobody. The fix is to ask rarely and show everything when it does ask.

Passwords and card numbers are different again

Asking is not enough for secrets. A password typed by an agent is a password the agent has handled, so the safest rule is that it never types your passwords or card numbers at all. You enter those yourself.

How to read a request to approve

When an agent asks, slow down for ten seconds. Check who it goes to, what exactly it says or costs, and when. For a message, read it as the recipient would. For a payment, check the amount and the payee against what you intended.

If anything is off, edit or deny. The agent can try again, and that costs you almost nothing. Approving by habit is the only way this protection fails.

  • Who receives it?

  • What does it say, word for word?

  • Is the amount or date right?

  • Did I actually ask for this?

What it can't do

  • Know a draft is wise. It can write the message; whether you should send it is your call.

  • Replace reading. Approving without reading removes the protection.

  • Guard what you give it. Passwords and card numbers should never be typed by it at all.

Privacy and control

DuckyHelper works this way in its early access agent features: it's built to ask first before it sends a message or pays for anything, and not to type your passwords or card numbers.

Try it yourself

Next time any tool offers to send or pay for you, check that you can see the final text or amount first. If you cannot, treat that as a warning sign.

It is worth saying why we do not make the agent ask about everything. A pause that appears at every click trains you to press yes without reading. Spending your attention on the two actions that matter keeps the attention real. Think of it as a budget: you only have so many careful looks in a day, so spend them where a mistake is costly.

And if you ever feel unsure about a request the agent shows you, deny it. Asking again costs seconds; fixing a wrong send can cost far more.

One last habit: when you set up a job that ends in a send or a payment, say so in your ask, for example "draft the reply and wait for me". That way the pause is expected, you know to come back and look, and the agent has a clear place to stop.

Frequently asked questions

Can I turn the asking off?

Treat that as a risk. The pause is the safety. Check any product's settings before relying on it.

Does asking make the agent slow?

Only at those two moments. Everything before them runs on its own.

Get started

DuckyHelper can
do more

It sees your screen and works in any app you use. Try it for 10 minutes, no card.

Get it for Mac, Windows or Linux

Share

Questions and jobs

Do anything

Ask it a question or hand it a job, in any app.

Doing jobs for you is in early access on Mac. Asking works on Mac, Windows and Linux.