AI Agents That Multiply Your Expertise

What AI replaces is execution. What it does not replace — and what gains dramatic value through it — is judgment. The deep understanding of what a good result looks like, what a correct quote contains, which phrasing holds up legally and which doesn't. The lever shifts: from execution to judgment.

The competence that matters shifts accordingly: away from technical mastery of a tool, toward clarity of intent. Those who know precisely what they need — and can articulate it the way they'd speak to a competent colleague — can now achieve things that previously required entire departments.

The Mittelstand has what AI does not: deep, specific domain expertise. The ability to judge whether a kitchen quote is correctly calculated. The knowledge of which DIN standard applies to a specific construction type. This knowledge lives in the minds of employees who've been with the company for decades. Agentic AI multiplies precisely this combination.

Anthropic, the company behind Claude, chose to steer its systems not through rule-following but through cultivated judgment. Whether they know it or not, this is the Aristotelian program: the machine should not follow rules. It should do the right thing at the right moment.

Phronesis is built on this insight. Our agents don't follow rigid decision trees. They possess domain knowledge, context, and the ability to apply both situationally. What makes a good quote depends on the client. What constitutes an appropriate complaint response depends on the case. No rule can decide this in advance — but judgment can.

The humanities have been working on this problem for 2,400 years. The ability to judge in concrete situations rather than execute rules — to do the right thing at the right moment rather than follow a decision tree — is not a technical problem. It's a philosophical one. And that's precisely why we're called Phronesis.

Between 1500 and 1800, European hermeneutics developed a tradition that was buried in the 19th century: probabilitas hermeneutica. Its answer to total doubt about understanding was not: We'll find certainty after all. But rather: We don't need it. What took its place were degrees of probability — sufficient certainty for practical purposes.

A language model operates on probability distributions over token sequences — it does, at a mathematically precise level, exactly what early modern hermeneutics described conceptually: treating meaning as a probability space in which one moves approximatively. The language model is not a deviation from the hermeneutic tradition. It is its technical realization.

And the agent loop — the iterative cycle of searching, finding, revising, searching further — is nothing other than the hermeneutic circle that Gadamer described as the fundamental structure of understanding: one understands the part only from the whole, but the whole only from the parts. One must always already bring a pre-understanding that changes through each iteration.

Foundation models are becoming a commodity. Performance gaps between GPT-5, Claude, and Gemini are shrinking. Token prices fall with a half-life of months. In a world where cognitive capacity becomes commodity, value shifts to what sits above it: orchestration, domain knowledge, integration, governance.

The productivity paradox — why AI isn't showing up in macro data — has a simple explanation: the bottleneck isn't model capability. It's the last mile. The translation of raw AI capacity into domain-specific, compliance-ready, immediately deployable tools. That last mile is precisely what Phronesis solves.

Europe has lost the infrastructure layer. But at the application level, it has genuine differentiation potential: regulatory competence, deep domain knowledge, a Mittelstand structure of 3.5 million companies too small for their own AI departments but large enough for substantial productivity gains. Germany doesn't need to lay its own tracks. It needs to run the right trains on them.