Two conversations stopped me this week. One with a student. One with a CEO.
They were asking completely different questions. But underneath, the problem was identical.
The Paradox Nobody Is Talking About
The student wanted to know why she should learn to calculate margins by hand when AI could do it faster. The CEO wanted to know why he should spend time mapping his approval workflows before selecting an automation tool.
My answer to both was the same: because the machine can’t tell you when it’s wrong.
This is the central paradox of the AI era I keep running into — in boardrooms, lecture halls, and due diligence calls: AI doesn’t produce more right answers than wrong ones. It produces more polished answers. The formatting, the confidence, the apparent coherence — it’s uniform regardless of accuracy. Which means the person reading the output carries more responsibility than ever, not less.
If you’ve never done the underlying work yourself, you’ve lost the calibration to catch the error. The machine delivers the correct margin and the impossible margin with equal elegance. If you can’t see the difference, you won’t even notice that you can’t.
The Judgment Problem Hiding in Your Processes
Here’s where it gets practical for CEOs and operators.
When I work with SME leadership teams, one of the first things I do is ask them to map where decisions actually live in their operations. Not the decisions on the org chart — the micro-judgments embedded in daily work. The ones that look like routine tasks.
A useful three-question check I use before any process automation:
- What is actually happening here? — Not the label (“approval workflow”), but the specific human action.
- What judgment is embedded in this step? — What does someone need to know to catch a wrong result?
- Should this be done by a person, a tool, or eliminated entirely? — In that order.
Most teams skip questions one and two and jump straight to three. They call it efficiency. What they actually do is automate a process nobody fully understood. And then they’re surprised when the output looks clean but the numbers don’t hold.
The cheapest hour you can spend this quarter is walking through a department’s ten or twenty tasks with these three questions. What you almost always find: several steps exist only because, years ago, someone had to sit between two systems. The logic was never in the process design — it was in the person’s head.
That person is often still there. The question is whether they always will be.
The Knowledge That Walks Out the Door
This brings me to a more urgent problem that gets almost no airtime in AI conversations.
In the Mittelstand, institutional knowledge is a quiet crisis. In almost every company I engage with, there’s someone — the export specialist who has been doing it for over two decades, the estimator who can feel when a bid is wrong before running the numbers, the operations lead who knows exactly which supplier holds under pressure. This knowledge isn’t documented. It lives in one person’s nervous system. And when that person retires, it doesn’t go to the wiki. It goes with them.
AI doesn’t solve this. In fact, it can accelerate the loss — because junior people, believing the tool will compensate, stop asking the experts. They stop creating the friction that actually transfers knowledge.
One of the most practical things a leader can do right now isn’t deploy a new AI tool. It’s send a junior employee to spend three hours with whoever holds the deepest tacit knowledge in the business. Not for a formal knowledge-transfer project — just to ask, listen, and ask again. That’s the raw material for everything that comes after.
The Judgment Stack: A Mini-Framework
When I put these two ideas together — building judgment through doing, and capturing judgment before it disappears — I get what I’ve started calling the Judgment Stack. Three layers, in order:
Layer 1 — Do it first. Before delegating a task to an AI tool, do it manually enough times to build calibration. You’re not being inefficient. You’re building the ability to audit the tool’s output.
Layer 2 — Map where judgment lives. For every process you’re considering automating, run the three-question check above. Know what’s being automated — not just what’s being accelerated.
Layer 3 — Extract before it exits. Identify the three people in your business who carry irreplaceable tacit knowledge. Create low-friction ways for that knowledge to transfer to the next generation — before the decision is forced on you.
The stack is deliberately sequential. Skipping a layer doesn’t make you faster. It makes you fragile.
A Brief Human Moment
A few weeks ago, I was reviewing a financial model with a management team preparing for a potential transaction. The model had been built with significant AI assistance. It looked immaculate.
Nobody in the room had questioned one core assumption — a margin figure that, if you knew the sector, was quietly implausible. Not obviously wrong. Just slightly off in a way that compounded badly by page twelve.
We caught it. But only because one person in the room had the judgment from years of operating in that industry. Without that person, that number would have sailed into the data room unchallenged. Clean formatting is not the same as correct analysis. That gap — between polish and precision — is exactly where judgment earns its keep.
From a transaction-readiness perspective, this matters doubly: a business whose critical judgment lives in one person’s head is a concentration risk. Buyers see it. Valuations reflect it.
What You Can Do Next Week
- Run the three-question check on one process your team complains about. What judgment is embedded? Does a human still need to be in that loop?
- Name your three “walking libraries” — the people in your business who hold the most irreplaceable tacit knowledge. Write their names down. That list is a risk register.
- Pair one junior employee with one expert for a single unstructured conversation this week. No formal agenda. Just questions and time.
- The next time you review an AI-generated output — a report, a forecast, a proposal — ask explicitly: “What would have to be true for this to be wrong?” Make your team answer it out loud before you approve anything.
- Before your next automation project, spend sixty minutes with the three questions above before you touch tool selection. The clarity will surprise you, and it will save you from automating the wrong thing.
What’s one area in your business where you suspect the judgment lives in a person’s head rather than in any system — and what are you doing about it before they walk out the door?