Before You Let AI Click, Send or Buy on Your Behalf

I'm happy to let AI do the clicking, but any button that can't be undone, I press myself.

For a couple of years, the worst AI could do was give you a bad answer. Now AI can open your web browser, fill in a form, send the email and move the file. In some tools, it can even buy the thing.

That's useful, but it's also the first time a mistake costs more than a second read. The question used to be "is this answer right?" Now the question is "what happens if it's wrong, and will I find out before it matters?"

The best picture I know is a driving lesson. The learner is at the wheel, and the instructor has a second brake pedal on their side of the car. Nobody thinks that pedal makes the learner useless. The pedal is there because some mistakes can't be taken back.

Every job you give AI sits somewhere on that road.

Reading needs no brake at all, and a job you can undo only needs a look afterwards. Anything you can't undo needs a human foot on the pedal every time, and the foot is yours.

Leave the approval setting on

Every tool that acts for you has a setting for how often it asks you first. The names vary, for example "ask before every action", "review the risky ones" or "run without stopping". Claude and ChatGPT both have a version of this setting, and both start out set to asking.

My advice is to leave it asking. On a long job, the AI will stop every few minutes to ask "can I send this?" Each stop is the brake pedal, which gives you a chance to see what's coming and say no.

People switch the setting off because the stops annoy them. Then one afternoon the AI sends a half-finished reply to a client, and they switch it back on. Skip that afternoon.

Write the rule down in your preferences as well. Preferences are the standing notes you give the AI about how you like things done, and writing the rule there means it reaches places the setting doesn't.

When you're acting on my behalf, show me anything you're about to send, submit, buy, post, delete or share. Then wait for my yes. If you're not sure whether something counts, it does.

A written rule is only an instruction, and AI can misunderstand an instruction. The setting is the lock. That's why you use both the rule and the setting.

Practise in an empty car park

Before anything real, give the AI a job where the worst case is nothing. You could copy a folder of documents and ask it to rename and sort the copies. You could ask it to compare three product pages and write up the differences, or to draft replies to five emails and send none of them.

Watch where it pauses, what it asks and where it surprises you. You're learning what its mistakes look like, so you'll spot one when the job is real. After that, give it one small real job with the setting on while you sit and watch, and then give it two.

Trust builds the way it does with any learner. You watch for a long while before you let them drive alone.

Write your never-list

Some roads the AI doesn't go down, however good it gets, and these are mine.

• Passwords, card numbers and bank logins. The AI may type into those forms, but it should never know the details. Some tools can hand over to your password manager so the AI never sees the password, and that's the only version I'd use.
• Moving money, trading, or anything with a "confirm payment" button. That button is always yours to press.
• Deleting anything that has no backup copy.
• Anything in a work system your employer hasn't said you can connect.
• Running alone, with nobody watching, on a job you haven't seen it do at least once.

Press stop when something looks wrong

Sometimes the AI lands on a page you didn't ask for, starts doing something odd, or quotes an instruction you never gave. If so, stop the AI. Every tool has a stop button. Once you've pressed it, read what the AI did, and then decide.

Text on a web page or in a document can sometimes look like an instruction to the AI. Good tools guard against this, and your approval setting is the second guard.

But no brake pedal works if you're asleep. After the twentieth "can I send this?", people start clicking yes without thinking, and that's the moment the twenty-first request is wrong.

When you notice you've stopped reading the requests, take that job back.

By the time you read this, the settings I've described will have moved, because the tools change every month. The road hasn't moved, though. Anything you can undo gets checked, and anything you can't undo waits for you.

AI fills in the form. You press pay.