Ten conclusions. One held without qualification.
I’ll be honest — I didn’t expect that number when I ran a counter-agent during a live workshop at DHBW Lörrach. We’d built a price-monitoring system in real time with students: AI agents scanning dealer websites across ten countries, surfacing nearly 180 product offers, auditing the data, verifying sources. The research was clean. The numbers checked out.
Then the counter-agent went to work. Not on the data — on the premises behind the conclusions. Six of our ten findings held only with conditions. Three didn’t hold at all. One was clean.
We rewrote the output.
The Wrong Conversation
Most AI conversations in the Mittelstand are still about tools: which model, which platform, which use case. I understand why — it’s a concrete question with a deliverable answer.
But it’s the wrong starting point.
The harder question — the one that determines whether AI makes your business stronger or just faster at being wrong — is this: what’s the judgment layer you’ve built around your AI?
The models are over-confident by design. I told the students exactly that during the workshop. They don’t know what they don’t know. They produce plausible-sounding output even when the underlying assumptions are broken. Clean data doesn’t protect you from a flawed premise. Verification agents don’t protect you from a flawed premise. Only deliberate challenge — built deliberately into the architecture — does.
That gap between what AI produces and what’s actually true is a leadership problem, not a technology problem.
The Conductor Shift
A retail CEO framed it cleanly at a summit earlier this year: he’s no longer the player — he’s the conductor. The orchestra is capable. Somebody still has to decide what gets played.
In practice, that means three things that remain entirely human: which question you ask, which answer you let through, and where you say stop. None of those is a technology decision. Your calendar reflects this shift whether you’ve acknowledged it or not.
If AI is working properly in your business, you should be spending less time on tasks done by hand and more time on review, decision, and direction. If you’re doing three times as much of the same work, faster — you’ve automated the treadmill, not escaped it.
The goal isn’t 10x output. It’s 10x your lifetime back for the work only you can do.
Your Moat Is Not the Model
A story from this week stayed with me. A colleague was invited to a large consulting firm who wanted to see how he works with his agents. He showed them his screen. Their response: they wanted it. The infrastructure, they said, was worth real money.
He declined. And then he said something more important: building the infrastructure is the easy part. Anyone can do it, and they can do it faster every day.
The question that remains is uncomfortable: what in your setup can they not copy?
My own list is short. Trust built over years. Data that only you have. Real customer access. Operational knowledge — knowing how your business actually runs, not how it looks on a slide. Regulatory understanding specific to your sector.
None of those live in the product. None come from the model. They come from what you’ve built, what you know, and what your customers believe about you. AI can compound that moat. It cannot create it.
The Workaround Trap
Two days later at DHBW Ravensburg, I asked a room of students: what’s the worst thing a CEO can say about AI?
Answers came fast. “Copilot handles that.” “The IT department takes care of it.” “The young hire just came from university — she manages it.” Then a student added a fourth I genuinely hadn’t heard before: a firm banning all AI in the building on EU AI Act grounds. My reaction wasn’t polished. It landed anyway.
The problem with banning AI isn’t legal risk — it’s that the ban doesn’t work. Your people are already using it privately. You haven’t removed AI from your company. You’ve made it invisible to you. That’s Shadow IT, and it’s a policy failure before it’s a technology problem.
We worked through two real cases in the room. One company transcribes every Teams meeting and has one employee manually copy the summaries into a shared notebook — one page per day. An agent could answer questions from that archive, but the thread gets lost as the archive grows with no structure behind it. Another marketing team pulls CSV exports from their ad platforms every month and pastes them into Excel.
Both are human workarounds wearing the costume of a process. The fix isn’t a better prompt or a new tab. AI needs a database, not a document. It should build the program once — then run it without anyone touching it each month.
The Conductor’s Stack
Here’s how I think about the judgment layer in practice. I call it the Conductor’s Stack — four layers that separate AI-augmented leadership from AI-augmented noise.
| Layer | The Question | Who Owns It |
|---|---|---|
| 1. Question Design | Am I asking the right thing? | You — the leader |
| 2. Premise Audit | Are the assumptions behind the conclusions sound? | Counter-agent or human reviewer |
| 3. Output Filter | Which answer do I let through? Which do I set aside? | You — the leader |
| 4. System Design | Is this a database and a program, or a workaround? | Operational decision |
Most Mittelstand AI projects operate only at layers 1 and 3 — ask and filter. Layers 2 and 4 are where deployment either compounds your edge or quietly erodes it. The counter-agent lives at layer 2. The decision to build a database instead of a new tab lives at layer 4. Both are leadership calls, not IT calls.
What You Can Do Next Week
- Audit one AI output your team relies on. Ask: is there anyone — human or agent — challenging the premises, not just the data quality?
- Review your own calendar. Has the balance shifted toward review, decision, and direction over the last six months — or are you doing the same tasks, faster?
- Write down your moat. What three things in your company can AI genuinely not replicate? If the list is short or vague, that’s important information worth sitting with.
- Find one human workaround. The monthly CSV paste. The meeting summary typed up by hand. Ask: does this need a better prompt, or a database and a program that runs itself?
- If AI is banned or informally discouraged in your company, ask what your people are actually doing. The answer is already there — it’s just outside your visibility.
Where in your business does AI produce something — and nobody ever challenges the premise behind it?
That’s where the real exposure lives. Not in the model. In the gap between confident output and sound judgment.