It's not that the output is wrong. It's that you never argued with it first.

Not long ago, I read that AI agents aren't just drafting paragraphs anymore. You give one a goal, and it plans the steps, uses the tools it needs, and runs the whole thing to completion. No check-ins. No draft to react to halfway through. Just a finished result waiting for you when it's done.

I sat with that for a while.

Because I remembered a smaller version of this same feeling. A newsletter section I'd asked an AI tool to draft, months ago. It came back clean and usable in seconds. I almost published it as-is.

What stopped me wasn't a flaw I could point to. It was the sense that I hadn't actually decided anything yet. The tool had decided for me, and I'd nearly let it stand in for a choice I never made.

Now the tools don't just draft the paragraph. They run the whole workflow. Which means the ten minutes I almost skipped back then just got a lot more expensive to skip.

Here's the distinction I keep coming back to. Assembled work looks finished. Developed work feels finished. They're not the same thing, and from the outside, you often can't tell which one you're looking at.

An agent can assemble almost anything now. Give it a goal, and it will find the pieces, put them in order, and hand you something coherent. That's real, and it's genuinely useful. But assembly isn't the same as understanding why the pieces belong together.

Developed work carries a different kind of proof. It's been reviewed and reworded several times before it carried the weight intended. You can usually feel that difference in a sentence, even if you can't name why. The agent can construct the shape of the content without ever going through the part that conceives it.

That's the trap. Not that the output is wrong. It's that "finished" and "understood" have quietly become two different things, and only one of them used to require you.

I think most of us have started treating friction like a bug, like something to be removed. The blank page, the false starts, the version you write and then throw away, all of it gets framed as waste now, time you could have saved, if only you'd had the good sense to let a machine handle it.

I used to think that too. I don't anymore.

The friction was never the obstacle between me and the finished piece. It was the process that made the piece mine. Every bad draft I threw away taught me something about what I actually meant. 

Something I couldn't have known until I'd tried saying it wrong first, sometimes more than once, sometimes embarrassingly so. When you skip that, you save time. But you skip the part where the idea becomes yours, rather than just being available to you.

An agent can't feel that friction, because it isn't trying to become anything. It's trying to finish. Those are different goals, and only one of them was ever actually about you. The other one is just very good at looking like it was.

I'm not the only one noticing this.

A recent industry report put a number on it. Seventy-three percent of talent leaders now say critical thinking is the number one skill needed to work alongside AI agents. Not technical skill. Not prompt-writing, which I promise is not the career-defining superpower half the internet is currently selling it as. 

Let that sink in for a minute. 

Critical thinking is the thing you use to decide whether what came back is actually right, actually yours, actually worth standing behind.

That's not a coincidence, and it's not really new information either. It's the market catching up to something a lot of us have felt privately for a while, usually right around the moment we almost hit publish on something we hadn't actually read closely. 

The more the tool can do on its own, the more your judgment becomes the entire point of your involvement. Not a nice-to-have. The job itself. The last job, possibly, that nobody's figured out how to automate, mostly because nobody's figured out how to bottle "actually paying attention."

Seventy-three percent of leaders didn't arrive at that number by accident. They arrived at it by watching what happens when nobody's left to argue with the output.

I don't have a polished system for this, but I have a habit.

Before I let anything, an agent's plan, an AI draft, even my own first pass stand as finished, I ask one question. Did I actually decide this, or did I just approve it? 

Those feel similar in the moment. They are not similar at all.

Approving something means it looked fine and I let it through. Deciding on something says I understood it well enough to defend it if someone pushed back. Most of the time, if I'm honest, the first draft of anything is only approved. It takes another look, sometimes a debate with myself, before it becomes settled.

I've gotten fairly good at catching it in my own writing. I'm still working on catching it everywhere else. The plan I approved in a meeting because pushing back felt like more effort than it was worth; the recommendation I took because the person sounded confident. 

Confidence isn't the same thing as being right. It just sounds like it.

That extra pass is the ten minutes. It's not a productivity tip. It's the difference between work that happened near me and work that occurred because of me.

AI Agents keep getting better at finishing things. That's not going to slow down, and I'm not writing this to tell you to distrust the tools. I use them constantly, and I'm not going back.

But finishing was never the hard part. Deciding was. And deciding still takes something no agent has figured out how to do for you. The ten minutes where you sit with an idea before it's convenient, before it's assembled, before it's handed back to you looking done.

I don't think that's a limitation of the technology that's going to get solved with the next version. I think it's a limitation of the question itself. You can automate a workflow. You can't automate the moment where someone decides a thing is actually true, or actually right, or actually worth their name on it. That moment has always required someone to still be in the room.

That's the part I'm not willing to offload. Not because it's slow. Because it's the only part that was ever actually mine.

© Copyright David Wakeman.  All Rights Reserved