Two weeks ago, in front of more than forty executives, one of my agents flagged that my own comment in a monthly report pointed the opposite direction from the figure I was commenting on. The agents were right. I said so out loud — “that was complete stupidity” — and the whole room watched it happen.
That moment was worth more than any demo I could have run. Not because being publicly wrong is a strategy, but because a system that catches your mistakes and correctly issues a NO-GO is one being governed properly. Not just used.
That’s the distinction most leaders are still missing.
The Tool Mindset vs. The Workforce Mindset
A tool is something you pick up, operate, and put down. You judge it by how easy it is to use.
A workforce is something you brief, direct, verify, and hand off to. You judge it by the quality of what it produces — measured against a bar you set, checked independently.
Most AI deployments I see in Mittelstand companies are stuck in tool mode. Someone has a license, uses it to rephrase emails or summarize PDFs, and calls it adoption. The value ceiling becomes visible almost immediately.
The shift begins when you treat your agents the way you’d treat a new team member: give them a role, a context, a quality bar, and an explicit handover that outlives the session. An agent that reports success is not proof of success. That’s true of any new hire too.
The Four-Point Workforce Contract
I run a fleet across several machines — a working mix of people and agents handling research, financial drafting, monitoring, and reporting. The management logic is the same as any small team. I’ve reduced it to four moves:
Brief → Target → Check → Handover
- Brief: Does this agent have a role description and genuine context about your business? An agent without context is a brilliant intern with no clue what your company actually does — and the damage it can cause is roughly the same.
- Target: Is there a quality bar — a definition of done — that you own, not the vendor? If you can’t articulate what good output looks like, you can’t catch bad output.
- Check: Do you verify outputs independently, against your own standard? One agent watching another is not a control system. It’s outsourcing your judgment twice.
- Handover: Does the run produce something that outlives the session — a document, a flag, a decision — that a human can pick up and act on? If the value disappears when the chat closes, it was never a process.
Past a certain point, scaling agents isn’t a technology problem. It’s a leadership problem. The question isn’t how many agents you have. It’s whether any of them have a job description.
Your Knowledge Is the Asset — Not the Model
Here’s a claim I stand behind: AI without your organizational knowledge is worth very little. The software becomes valuable only when your context flows into it.
You didn’t build your business on your ERP system. You built it on decades of judgment, relationships, process knowledge, and hard-won calibration — most of which currently lives in a few people’s heads and walks out the door when they retire or leave. That’s not just a succession risk. It’s a valuation risk.
AI can turn that institutional knowledge into a scalable, transferable asset. But only if you deliberately feed it in. The sequence I give to every leadership team is consistent: export, do the work manually, put the result back, validate the loop by hand — only then bring in automation. Don’t start with the system. Start with understanding what the system should do.
With a traditional ERP, you adapt to the software. With AI, the software adapts to you — but only if you do the work of putting your knowledge in first.
Sovereignty Is a Leadership Question, Not a Model Question
This week, a new open-weight German-language model entered the conversation around AI sovereignty for European enterprises. The framing from most commentators: open weights plus local hosting equals sovereignty.
I’d push back on that.
Open weights give you the right to run a model. Sovereignty is something different: it means you can test it, compare it, and swap it when a better option appears. That requires three things — and only one of them lives in the model itself.
First, you need your own test cases. Four or five real problems from your own business, run against every model you consider. Vendor benchmarks tell you nothing about your documents, your edge cases, your risk profile.
Second, you need a clean operating environment — one where no external party reads your data — and people in-house who can manage it.
Third, you need a model that explicitly says when it doesn’t know. And you need to verify that claim through your own tests, not the datasheet.
We ran a new model through our own evaluation tasks the night it launched. Better on long documents, cleaner on logical reasoning, noticeably faster. Five literature citations without internet access: two were correct. Our smaller incumbent model cited fewer total, but explicitly flagged its uncertainty. Which matters more depends entirely on the task. The answer is never simply “use the newer model.”
Sovereignty isn’t a product you buy. It’s a practice you build — one test case, one verified loop, one governed process at a time.
What You Can Do Next Week
If you’re ready to move from tool-user to workforce-manager, here’s where I’d start — in order:
- Audit what’s running unsupervised. For every AI tool in active use, ask: does it have a Brief, a Target, a Check, and a Handover? If not, it’s operating without governance.
- Write four test cases from your own business. Real decisions, real documents, real edge cases. Run every new model or AI feature through them before adoption. Own your quality bar — don’t rent the vendor’s.
- Map where your knowledge actually lives. Identify one area where critical business judgment currently sits in someone’s head. Design a way to capture it and feed it into a workflow — manually first, before any automation.
- Build one deterministic process, not another chat prompt. Choose one recurring task — monthly close, budget review, pipeline report — and design it as a system: inputs, logic, output, validation. The goal is a production line you can run daily, not a query you run occasionally.
- Normalize being wrong out loud. The moment I said “that was complete stupidity” in front of forty people was the moment adoption actually started in the room. Your team needs to see you engage with AI as a learner, not a performance.
One question I’d leave you with: where does the knowledge that actually runs your business live right now — in a system you own, or in someone’s head?
That’s not rhetorical. It’s the starting point for everything else in this conversation. If you’ve started mapping it in your own company, I’d genuinely like to know what you found.