AI Gets More Capable and Less Readable at the Same Time — Here's What That Means for Mittelstand Leadership

I watched fifteen approvals happen in under four minutes. Not one of them was actually a decision.

Each person in the chain had an AI summary in front of them. Each clicked approve. Nobody had read the underlying document. Nobody could have caught an error if one existed. The process looked like governance. It was theater.

That scene has stayed with me — not because it was an outlier, but because it wasn’t.

The Real AI Question Isn’t “Which Jobs Disappear?”

Every room I walk into wants to talk about displacement. Fair enough. But I’ve noticed the harder question gets skipped: What does a human need to bring to the table for their judgment to actually matter?

This week I had the same conversation from three directions at once — with a Mittelstand CEO before we started recording our podcast, with my students at DHBW in my AI Catalyst classes, and at home as a father. The question is everywhere right now.

The displacement framing leads to anxiety and paralysis. The judgment framing leads somewhere useful.

Here’s where I’ve landed: as AI becomes more capable, the human contribution isn’t disappearing — it’s concentrating. It’s narrowing down to three acts that cannot be delegated, no matter how good the model gets.

The GCL Model: Goal, Context, Limits

In every AI-assisted process — a contract review, a financial analysis, a junior’s first onboarding task — the quality of the output is determined by three things the human must own:

Goal — What are we actually trying to achieve? Not “summarize this document,” but why, for whom, and to what decision does it connect?

Context — What does the model need to know that isn’t in the document? Your industry norms, your specific situation, the non-obvious constraints, the thing your most experienced colleague would have mentioned in the first five seconds.

Limits — Where does the AI’s output stop and human accountability begin? What can it not decide? What requires a real sign-off, not a click?

I use this model in my AI Empowerment workshops because it survives every technological shift. The slides update every few weeks — a multi-agent setup that took months to run in early 2025 takes minutes now. But whoever owns Goal, Context, and Limits is still in the driver’s seat.

AI is not an IT project. It is leadership work. I say that line out loud, early, in every workshop. Later I can just point back to it.

Why Clever Prompting Became Irrelevant by Lunch

This July, I ran a full-day workshop with an industrial group. By noon, something had shifted. The group had stopped arguing about how to phrase their prompts. They were arguing about their own internal processes — who hands off what, what “done” looks like, who actually owns the decision.

That’s the right argument to have.

Researchers at Stanford and Northeastern have documented what they call “typicality bias” — models trained on internet-scale text trend toward the average answer. The way you pull high-quality output back out isn’t a magic phrase. It’s a clear goal and rich context. That group reached the same conclusion themselves, which is why it landed harder than any paper I could have cited.

Context engineering has replaced prompt engineering. And context is a leadership task, not a technical one. It requires you to understand your own business well enough to brief a very capable but contextually blind assistant.

The Transparency Trap No One Is Talking About

Here’s the shift I’m watching most carefully right now.

Current AI models think by writing — you can read the chain of thought, follow the reasoning steps, catch an error before it propagates. That’s a useful property, and many governance frameworks have been quietly built on top of it.

But that transparency is becoming a construction choice rather than a given. Newer architectures run reasoning through hidden states rather than written-out steps. The gain in capability is real. The side effect is that the reasoning becomes unreadable.

If your control system is built on being able to watch every step, you have a fragile control system.

The answer isn’t deeper technical oversight. It’s cleaner ownership at the top and bottom of the process: sharper objectives going in, clearer accountability coming out. The GCL model holds precisely because it doesn’t depend on reading the middle.

This is where I see Mittelstand leaders most exposed — not because they’re using AI carelessly, but because they’ve built their oversight on a transparency that won’t last.

The Junior Question Is Really a Judgment Question

Back to that CEO who told me, just before the camera started: “I don’t have any juniors — and I don’t know how to hire them anymore.”

I hear the mirror version from my DHBW students: Will there be a job for me?

My answer to both: the entry-level role isn’t gone. It’s been reassigned.

A junior who has never worked through a calculation themselves cannot judge whether an AI’s output is correct. A junior who has never read a contract cannot evaluate a summary. When they hit approve, it isn’t a decision — it’s the same theater I described in the opening. Fifteen sign-offs, zero decisions.

The practical entry-level job I keep coming back to sounds unglamorous: go into the business and describe it. Map the knowledge that lives only in people’s heads. Find the person who has run the export department for decades, whose phone rings during vacation, whose succession has failed three times. Sit with them until what they know can survive their absence.

That’s not junior work in the old sense. It’s judgment development — and it’s essential groundwork for any AI deployment that is supposed to genuinely transfer to the next owner, team, or buyer. Which, if you’re building a business that is fit for transaction, matters more than most leaders currently appreciate.

What To Do Next Week

Five posts, one week, one conclusion: AI gets more capable and less readable at the same time. The right response is not to understand it more deeply — it’s to own your judgment more clearly.

Here’s what that looks like in practice:

Audit your approvals. Pick one workflow where your team approves AI-generated outputs. Ask honestly: is each sign-off a real decision, or a click? If it’s a click, fix the process before you scale it.
Run the GCL check on your three most AI-assisted processes. Who owns the Goal? Who supplies the Context? Who defines the Limits? If the answer is “the AI figures it out,” you have a governance gap.
Stop optimizing prompts. Start briefing. Treat AI the way you’d brief a sharp external consultant — on your industry, your situation, your non-obvious constraints. That context is worth more than any clever phrasing.
Map one critical knowledge holder this week. Who in your business carries irreplaceable expertise that exists only in their head? Assign someone — ideally a junior — to start turning that into documentation. This is urgent work regardless of AI.
Pressure-test your oversight model. If any part of your AI governance depends on readable chain-of-thought, document that assumption explicitly. Design the fallback now, before the model changes and the transparency disappears.

No doom and gloom. But act while you can act.

Which of your current approvals would you be uncomfortable defending if someone asked you to explain the underlying judgment call — not the output, but the actual decision behind it?