The Unnamed Position
In spring 2025, a mid-sized mechanical engineering company in the Landshut area — referred to here as Präzicon — received an unexpected invoice. Not from a supplier or a customer. From its own production.

Eighteen months earlier, Präzicon had deployed an AI-based quality control system. The model inspected components for dimensional accuracy — more precisely and more quickly than the previous manual inspection. The pilot went smoothly. The rollout too. The model ran stably. The reject rate fell.

The Unspoken Hope
When companies talk about AI adoption, one hope is almost always implicit: the model will solve our data problems.

Machine learning approximates a function. The function it approximates is that of the training data. If the training data is contradictory — the model gives contradictory outputs. No algorithm, no architecture, no volume of data overcomes a definition problem.

In spring 2026 the Berlin tax advisors' chamber files suit against the company Accountable. The subject of the proceedings is a single word on the website: AI-tax-adviser.

What a Lawsuit Actually Attacks
The first layer is professional law. Tax advisers, lawyers, doctors, and auditors operate under protected titles. Whoever uses the word tax adviser in a product description collides with a layer that applies regardless of whether the product factually serves its clients better.

EU AI Act in Practice
The EU AI Act entered into force in August 2024. Transition periods for most high-risk categories end in August 2026 — sanctions of up to €30 million or 6 % of global turnover.

Product safety. AI systems as safety components under EU regulatory regimes (Machinery Regulation 2023/1230). Influence on product safety = most likely high-risk.

Market Observation: Salary vs. Cloud Costs
Industrial plant 200–500 employees: cloud expenditure for AI €80,000–€250,000/year. AI operator salary: €45,000–€75,000 gross. Senior: up to €95,000.

For comparison: senior software architect €95,000–€130,000, ML engineer €85,000–€110,000, SAP architect €100,000–€140,000.

Introduction: Where Systems Really Break
The four cases described are anonymised but real. They share one thing: the technical part functioned within its specifications. The problem lay in the anthropological layer.

Case A: Mechanical Engineering, Predictive Maintenance
Upper Bavaria, 320 employees. Friday evening 21:30: system reports "Critical" — predicted spindle malfunction in 12–36 hours, probability 91 %. Potential damage: €45,000.
After one hour of discussion: decision against stopping. The machine stopped itself the next morning. Damage: €78,000 plus downtime.

The Layer Model: Technical and Social Layer
The 70/30 architecture requires two diagrams: one for the technical layer (model, pipeline, infrastructure, logging, monitoring) and one for the social layer (roles, escalation paths, mandates, thresholds). Both layers are equal system components. One without the other is not an AI system.

EXPECTED vs. REAL CONTROL
▲ Expectation layer
Expert reads report · follows the reasoning · identifies errors
Assumption: human control detects AI errors