The two jobs you can't hand to AI

The move Hand off the build; keep the deciding and the checking, because they carry the stakes.

The short version: Any task splits four ways: decide, build, check, run. AI is great at building and running. But checking needs a standard for what good means and stakes if it's wrong, and AI has neither. So let AI build it, and never let AI be the one who checks it. That check is the part they pay you for.

The scene

The draft comes back and it's good. Clean structure, right tone, nothing obviously wrong. You're the advisor on this account, and the client update needs to go out before lunch. So you skim it, fix a comma, and hit send. Fifteen minutes, done. A year ago that update took you an hour.

The trap

Here's what actually happened. Any piece of work has four parts: deciding what it should be, building it, checking that it's right, and running it out the door. AI is genuinely great at the building. Fast, tireless, cheap. But you didn't just hand it the building. You handed it the checking too. The skim-and-send was the check, compressed into fifteen seconds because the draft looked finished. And the checking was the part they were paying you for. Anyone can generate a client update now. What the client trusts you for is the judgment that catches the number that's off, the commitment you can't actually make, the tone that'll land wrong with their CFO.

The shift

Let AI build it. Never let AI be the one who checks it.

AI can build it. It can't be the one who checks it.

Checking needs two things AI doesn't have: your standard for what "good" means here, and skin in the game if it's wrong. A polished draft isn't a checked draft. It just looks like one, which is exactly what makes skipping the check feel safe.

Use this

Use AI to check faster, but keep the judgment yours. Before you ship anything it built, run this:

Here's a task I need to get right:
[what it is + who it's for]

Here's the draft:
[paste it]

Before I check this myself, surface what I should scrutinize. Return:
1. The three claims most likely to be wrong
2. Any number, name, or date you're not fully sure of
3. What you assumed that I never told you
4. The one place a sharp reader would push back
Don't fix any of it. Just flag it so I can judge.

Filled in, the top looks like this:

Here's a task I need to get right:
A monthly update to a client's CFO on a project that's two weeks behind.

Here's the draft:
[the AI draft you were about to send]

Why this helps: you still do the checking, so the judgment stays yours, but AI hands you the shortlist of what to check. A real review takes five minutes instead of thirty. "Don't fix it, just flag it" is the whole point.

What stays human

AI can point at what's risky. Only you can decide whether it's actually wrong, whether it matters here, and whether your name goes on it.

Try it this week

Next time an AI draft looks finished, don't let "looks finished" be your check. Run the flag prompt once, then spend five real minutes judging what it surfaces before you send.