In partnership with
Yes, that Arnold Schwarzenegger.
He has a newsletter on beehiiv. So do Codie Sanchez, Jay Shetty, David Begnaud, Colin & Samir, and Joanna Stern.
They could publish just about anywhere. They chose beehiiv to build a direct connection with the people who want to hear from them most. And you can do the same, whether you’re world-famous or just at the start of your plan for world domination.
A person should review the AI's work. I say it in every room. Last week, someone in one of those rooms asked me which person.
Last month asked how you would know if a system you trust stopped working. This month turns to the people asked to use them. Here is the question, and what I took from it.
The Question From the Room
I host a monthly lunch-hour AI session for the people I support at a mid-market manufacturer. Last week, someone newer to their role raised a hand.
Part of their day is math in a spreadsheet, and they said plainly that the AI does it better than they do. They knew the standard answer: someone checks its work. Their problem was that they hadn't been in the role long enough to know what a wrong answer looks like. The veterans can tell. They can't.
I contemplated which worry was bigger: losing the job or not being able to check the work. Obviously, it was the job. And they looked a step further ahead: as AI gets more expensive to run, it will sit next to labor on the budget, and a person who keeps handing the AI their own responsibilities, then moves into a thin review role, looks like the line item that can go.
In another room and another meeting, a veteran employee joked to a teammate, a beginner AI user as well: you're AI'ing yourself out of a job. The teammate didn't laugh it off. They answered: no, a human is still needed to review the answers and make sure the AI had the right facts.
Later, in both situations, someone else said the quiet part out loud: teach your assistant your work too well, and have you replaced yourself?
What I Told Them
I told them what I could. I don't have control over staffing, so I can't promise it. I can promise how I build with the tools made available to me. If the AI I help with gives someone back fifteen minutes, I count that as a win. And governance says nothing gets written to production until a person reviews it and approves it, because AI is never perfect. A misplaced decimal in the input goes straight through, and the AI processes it faithfully unless a check catches it.
That answer is true. It's also incomplete.
Which Human?
Both situations play to the whole picture. The joke had the right worry, and the teammate had the right answer. People are still needed to validate what the AI produces, and to check that it had the right facts. That is the job that's left. But validating only works if the person can tell right from wrong, and telling is a skill nobody learns from the output. They learn it from years of doing the work by hand, getting it wrong, and being corrected.
Today's reviewers are the people who did the work before the AI did. The newer person in that room is the first sign of the gap. We are automating the repetitions that teach someone what wrong looks like, and then counting on a human in the loop who needed those repetitions to become one.
I don't have a finished fix. Here is what I would do.
Three Moves
Name the reviewer, and make review teachable. For every AI step, name the person who checks it, a person and not "the team." Then write down what wrong looks like. Have the veterans keep a short list of the times the output looked right and wasn't, and give newer people known-answer cases to check against. Review becomes something you can learn on purpose.
Make the teacher the author. If teaching the assistant your job makes you redundant, nobody will teach it well. If it makes you the named author of how the work gets done now, people will. Put their name on it.
Make it safe to say out loud, and be honest about your limits. The question came up because there was enough trust in the room to ask it. The most useful thing I had was an honest split: here is what I control, and here is what I don't. A promise you can't keep costs more than a limit you admit to.
One caution: this is one room. Treat it as a pattern worth checking for in yours, not a law.
Try This Monday
Thirty minutes, on AI-assisted work you already run.
List every step where an AI produces something a person then relies on. For each step, answer two questions:
Who reviews it? A name, not a team.
Could someone with six months in the role catch a wrong answer? Not "would they look at it." Could they tell?
When you finish, you'll have two numbers:
Named: how many steps have a reviewer with a name next to them.
Could catch it: how many of those reviewers would actually spot a wrong answer.
Here's an example. You list 8 steps. 6 have a named reviewer. Of those 6 people, 3 could tell if the AI got it wrong. Your result is "6 of 8 named, 3 of 6 could catch it."
The gap between those two numbers is where "a human reviews it" is a hope, not a control.
Not sure about the second number? Plant one wrong figure in a copy of last week's output and see who finds it.
Who Checks the Checker?
"Human in the loop" is the line everyone uses and almost nobody tests. Naming the human is the first test. Whether that human can tell is the second.
A reviewer is a skill, not a seat.
ACTION ITEM: How many of your AI steps have a reviewer who could catch a wrong answer? Run the Monday exercise and send me your two numbers: steps with a named reviewer, and reviewers who could catch a wrong answer.
Reply, I read every one.
Go hug your ERP. Then ask who taught the reviewer.
Plant Floor to Cloud: The Newsletter goes out every Tuesday. The podcast runs monthly. Same charter, more room to think out loud. Watch or listen on the website, or wherever you get podcasts: Apple Podcasts, Spotify, YouTube, iHeart, and more.
