Some AI Mistakes You Can Undo. Some You Can't.

Watch how people actually use AI and you'll notice they tend to pick a side. One group trusts it with everything. They paste in the draft, skim it, ship it, and move on. The other group doesn't trust it with anything. They run every answer through three rounds of second-guessing, and half the time they wonder why they bothered opening the thing at all.

Both camps are asking the wrong question. The question isn't "can I trust the AI." The question is "what happens to me if it's wrong here."

Because here's the thing nobody says out loud: not every mistake costs the same. Some AI mistakes cost you four seconds and a redo. Some cost you a client, a bank balance, or a reputation you spent a decade building. If you treat those two situations the same way, you're going to be either reckless or painfully slow. Usually both, on the wrong days.

There's an old idea about doors

Long before any of this, people who had to make a lot of decisions sorted them into two buckets. Some decisions are like walking through a door you can walk back out of. Don't like what's on the other side? Turn around. No harm done. Other decisions are like a door that locks behind you. Once you're through, you're through.

Reversible and irreversible. That's the whole idea, and it's the most useful filter I know for deciding how much to trust an AI with any given piece of work.

Ask it on anything you're about to hand off:

• If the AI gets this wrong, can I take it back?
• If the answer is yes, and taking it back is cheap, let it run.
• If the answer is no, slow down and put a human in the way before anything goes out.

That's it. You don't need a risk committee. You need to know which kind of door you're standing in front of.

The question was never whether AI can do the task. It's what happens to you if it does the task wrong.

Why this matters more now than it used to

For most of history, the thing that slowed your work down was also the thing that caught your mistakes. Writing a proposal took two hours, and somewhere in those two hours you noticed the number was off. Sending a letter meant printing it, which meant reading it one more time. The friction was annoying, but it was also a safety net.

AI removed the friction. That's the point of it, and it's genuinely good. You get the proposal in ninety seconds. You get the email, the summary, the spreadsheet formula, the plan, all of it, faster than you can check it.

But speed is the feature and the risk wearing the same coat. When the work comes out that fast, the old pause that used to catch the mistake is gone. So you have to put the pause back on purpose, and only where it counts. Put it everywhere and you've thrown away the speed you were paying for. Put it nowhere and you're one confident wrong answer away from a bad day.

The reversible-or-not filter tells you where to put it.

Where to let AI run full speed

If the worst case is "that wasn't quite right, let me try again," you're standing in front of a door you can walk back through. Go fast. Fix it later if you need to. The cost of a wrong first draft is a redo, and a redo is cheap.

This is most of your day, honestly:

• First drafts of anything. The email, the blog post, the job description, the outline. A clumsy draft costs you nothing. You were going to edit it anyway.
• Brainstorming and options. Ask for ten names, five angles, three ways to structure the meeting. Bad ideas in a list hurt no one. You're the one choosing.
• Summaries you're going to read yourself. If the AI's recap of a long document is a little off, you'll catch it, because you're the only one reading it.
• Rough math and quick research you'll verify before it matters. Ballpark numbers, first-pass comparisons, a starting point you'll check against the real source.
• Reformatting and cleanup. Turning your messy notes into a tidy list. Worst case, it looks wrong and you tweak it.

Notice the pattern. In every one of these, you see the mistake before anyone else does, and undoing it is a keystroke. That's a two-way door. Treating this work with heavy suspicion isn't being careful. It's just being slow for no reason, and it's the fastest way to decide AI isn't worth it. We've written before about consistency being where AI actually pays off, and this is part of why. The speed only pays if you let it run where it's safe to run.

Where to slow down and put a human in the way

Now the other door. If the mistake would be hard, expensive, or embarrassing to walk back, stop treating the AI's output as finished. Treat it as a draft that a person has to sign off on before it leaves the building.

The one-way doors tend to share a few features. The output is about to:

• Go out to people. The email blast, the message to a client, the post that goes public. Once it's sent, it's sent. You can apologize, but you can't unsend.
• Touch money. An invoice, a quote you're committing to, a price, a payment, a refund, anything with a dollar figure a customer will hold you to.
• Make a promise. A contract, a proposal with terms, a deadline you're agreeing to, a commitment someone will plan around.
• Delete or overwrite something. Clearing out records, replacing a file, running a bulk change you can't easily reverse.
• Carry your name in public. Anything with your brand on it where being wrong doesn't just cost a redo, it costs trust.

For this work, the AI is still useful. Let it do the drafting, the first pass, the heavy lifting. Just don't let it be the last step. A person reads it, catches what the machine confidently got wrong, and makes the call to send.

This isn't paranoia. It's the same instinct behind using AI as a second set of eyes right before you hit send, except pointed the other direction. Before, AI is the reviewer. Here, AI is the one being reviewed, because the stakes are high enough that you want a human's judgment to be the thing that commits you.

And it matters because AI is wrong in a particular way. It isn't wrong like a nervous intern who hedges. It's wrong like a confident expert who isn't. It will give you a made-up number in the same calm, certain voice it uses for a real one. We've made this point on its own because it's so important: AI being usually right is exactly what makes it dangerous. On a two-way door, a confident wrong answer costs a redo. On a one-way door, it costs you.

Three questions that sort it in about ten seconds

You don't need to overthink this in the moment. Before you let any AI output go forward, ask:

• Can I take it back? If undoing it is cheap and quick, run fast. If undoing it is hard or impossible, slow down.
• Who sees it if it's wrong? Just me means low stakes. A client, the whole list, or the public means high stakes.
• What does the worst case actually cost? A redo, or a relationship? Be honest about the real number, not the one that feels comfortable.

Most of what crosses your desk will come out clearly reversible, and you should feel good about letting AI fly through it. A smaller slice will be one-way doors, and those are exactly where your attention is worth the most.

The mistake is having one setting

Here's the real trap, and it's the thing both camps from the top of this post fall into. They've got one setting. The trusting crowd treats every door like a two-way door, which works fine right up until the day it doesn't. The suspicious crowd treats every door like a one-way door, so they move at a crawl and get a fraction of the value they're paying for.

The skill, the actual skill worth building this year, is matching your caution to the cost of being wrong. Fast where it's safe. Slow where it's not. Different settings for different doors.

That's also a cleaner way to think about what AI is doing to your work in the first place. It isn't replacing your judgment. It's handing you a pile of fast, cheap first passes, and leaving you the job that was always yours: deciding which ones are safe to let go and which ones you need to stand at the door for. That's the same line we keep coming back to, that AI is coming for the tasks, not the judgment. The tasks are the drafting. The judgment is knowing which door you're standing in front of.

And it goes double for anything you've set to run on its own. An automation is a door you propped open and walked away from, so a wrong answer repeats quietly until someone finally notices, which is its own argument for why set it and forget it is the biggest lie about AI. If it touches money or sends things to people, put a human glance before the irreversible step.

So here's the only homework. Next time you reach for AI, don't ask whether you trust it. Ask whether you can take it back. Then act accordingly, and stop spending the same worry on a throwaway draft that you'd spend on a contract. One of those deserves your attention. The other one just wants to get out of your way.