I Use AI to Attack My Work, Not Write It.

I've started using AI the opposite way most people do. Not to write the thing. To attack it.

Before a deliverable goes out, a plan, a proposal, a recommendation, I hand it to an AI and tell it to be the most skeptical person in the room. Not to polish it. To find the one place it falls apart. Play the executive who is looking for a reason to say no. Find the weakness my own review was protecting.

Why my own review misses it

My review has a blind spot, and so does yours. I read my work the way I meant it, not the way a stranger will. My checklists catch what I already know to look for. They can't catch the assumption I didn't know I made. The skeptical reader has no loyalty to my intent, so it reads the words that are actually on the page.

Last week it caught a load-bearing claim in a deliverable that every pass of mine had waved through. One sentence. The whole argument rested on it, and it didn't hold. I'd read past it a dozen times because I knew what I meant.

The part teams get backwards

Teams point AI at production and call that the win. Generation is the cheap half. The expensive half is judgment, and judgment is where most teams still won't let the machine near their work.

It is uncomfortable, which is the point. You are inviting a critic in before you are ready for one. The alternative is that the first skeptical read happens in the room, by the person you most needed to convince.

The teams getting real value aren't asking AI to do more of the work. They're asking it to find where the work is wrong, before anyone else does.