The Question That Stopped the Room
Last week, someone asked me a question that sounded perfectly reasonable—and revealed everything about why most companies struggle with AI.
We had just presented Pheraxia GmbH publicly for the first time at a summer event in Freiburg. The energy was good. The concept landed. Then came the question: “So you’ll need to hire for accounting, marketing, simulation… and raise capital for headcount and lab infrastructure, right?”
It was asked in good faith. By someone sharp. And that’s exactly what made it so instructive.
Because in our world, that framing is the old playbook—and the old playbook no longer applies.
The Real Resistance Isn’t Technical. It’s Mental.
I’ve spent two years helping DACH companies implement AI. I’ve seen the full spectrum: skeptics who became builders, operators who thought AI was someone else’s problem, teams that ran pilots and then shelved them when the pilot ended.
But this week crystallized something I’ve been sensing for a while:
The barrier to AI adoption in most SMEs isn’t capability. It’s imagination.
The mental model of “company = people + capital + office + scale” is so deeply embedded that an AI-first approach isn’t perceived as different—it’s perceived as incomplete. Unfinished. Like something is missing.
At a graphite electrode manufacturer in Grevenbroich—a 100-year-old industrial business making materials for green steel—an AI workshop day began with a senior team member making a quiet admission: “I always saw AI as just a tool. But that’s the wrong way to look at it.”
No fanfare. No slides. Just a moment of honesty that unlocked the entire day.
By end of day, every department had built something they owned. That’s the pattern I keep seeing: the turning point isn’t a demo. It’s a confession. It’s the moment someone stops defending the old model and starts experimenting with the new one.
What AI-First Actually Means (A Framework)
Let me be precise, because this term gets abused.
AI-first does not mean:
- Replacing your team with bots
- Running without governance
- Skipping domain expertise
- Waiting for a perfect tech stack
AI-first does mean:
- Redesigning your operating model before you hire or scale
- Treating AI agents as roles, not features
- Keeping humans in command of every external decision
- Iterating in weeks, not quarters
At Pheraxia, we’ve built 55 AI roles across six departments—Science, IP, Commercial, Brand, Diligence, and Ops—with a two-person human team. Every external artifact gets human sign-off. Not because we don’t trust the models, but because governance is the product. Capital goes into proof-of-concept and productization. Not headcount. Not our own lab infrastructure.
This isn’t a cost-cutting story. It’s a different architecture entirely.
The Three-Layer Model I Call “Leverage Before Headcount”
Layer 1 — Clarity: Map what your business actually needs to produce (outputs, decisions, external artifacts). Don’t start with tools. Start with function.
Layer 2 — Assignment: For each function, ask: human judgment required, or AI-executable with human review? Most SMEs discover that 60–70% of recurring work falls into the second category.
Layer 3 — Governance: Every AI output that touches the outside world—customers, regulators, partners—needs a human sign-off protocol. This isn’t overhead. It’s your credibility layer.
Leverage before headcount. Not instead of people—before the reflex to hire.
Sovereignty Doesn’t Start in Brussels
There’s a think tank report circulating that warns Europe will fall behind in AI by 2031—not because of missing talent, but because of three misjudgments: underestimating speed, underestimating depth of transformation, and overestimating the ability to catch up later.
I read it with interest. The diagnosis is largely right. The prescription—billions in compute, top-down coalitions, state-backed infrastructure—is where I part ways.
Because I watched a two-person team build a model-agnostic, multi-vendor, EU-compliant AI stack without waiting for Brussels, without a nine-figure budget, and without a government program. The stack is sovereign not because of where the servers sit, but because of the decisions made about how to build it.
Sovereignty, in my experience, is an attitude before it is an infrastructure.
The SME leaders and operators I work with across DACH often ask: “What’s left for us in this new world?” My answer is always the same: more than you think, but only if you start today. The gap between those acting and those observing is compounding faster than most realize.
The edge isn’t technology. It’s the decision to stop waiting for the perfect moment and start shipping in this one.
The Agents That Check the Agents
One detail from the Grevenbroich workshop that I keep thinking about:
Two departments ran the same data through the same model and got different outputs. Frustrating at first. Then we added a second AI agent to validate the first. Outputs aligned. The team understood immediately what had happened—not because I explained it, but because they built it.
This is the real lesson about AI systems: it’s not about prompts. It’s about architecture. Agents that check agents. Workflows that replace guesswork. Domain experts who know their process well enough to build meaningful validation into the system.
Deep domain expertise isn’t a liability in an AI-first world. It’s the competitive moat. The teams that drive real AI adoption aren’t the ones with the fanciest tools—they’re the ones who know their own business cold enough to design systems around it.
What You Can Do Next Week
If any of this resonates, here are five concrete moves—no budget required, no consultants needed:
-
Run an output audit. List the 10 most recurring outputs your team produces each week (reports, emails, analyses, summaries). For each, ask: is human judgment genuinely required, or is this executable with AI and a human review step?
-
Name your governance layer. Before your next AI experiment, define one rule: which outputs require a human sign-off before going external? Write it down. Make it explicit. This is your credibility infrastructure.
-
Pick one confession to make. In your next team meeting, name one assumption about how your business operates that AI might make obsolete. You don’t have to have the answer. Just naming it starts the conversation.
-
Try the two-agent check. On any AI task that matters, run a second model to validate the output of the first. Note where they diverge. That divergence is information—and it’s usually more valuable than either output alone.
-
Spend one hour on your operating model, not your tool stack. Don’t start with software. Start with the question: If we were designing this business today, what would we build versus buy versus delegate to AI? One hour of that thinking is worth ten hours of tool evaluation.
The hardest part of leading in this moment isn’t understanding AI. It’s unlearning the reflexes—hire to grow, raise to scale, wait for certainty—that made sense in a different era.
The new playbook doesn’t have a published edition yet. We’re all writing it in real time.
What’s the one assumption about your operating model that you suspect AI is already making obsolete—but haven’t yet said out loud?