Every Friday: One AI topic. One clear verdict. No bullshit.
What happened
This week, Germany reached another milestone on the path to a digitized public administration—or at least the next step in a very long march. Digital Minister Karsten Wildberger (CDU) presented the first 18 pilot projects of the so-called “Agentic AI Hub”: AI agents deployed in real government agencies to verify housing eligibility certificates, process care applications, and automatically categorize incoming government mail. The name of the program: “Colleague AI.”
Specifically: Forml is working in Frankfurt and Düsseldorf on the automated review of housing eligibility certificate applications. Formfix supports the processing of care applications in Cologne. Lector.ai processes scanned documents in the Neckar-Odenwald district via Vision-LLM. Wildberger promises an 80 percent acceleration in approval procedures. Behind this lies demographic pressure: The baby boomer wave is hitting the public sector with full force, and there simply aren’t enough young people coming through.
The timing is ironic. The EU AI Act classifies AI in government decision-making processes as high-risk—with corresponding transparency, explainability, and audit requirements. And Germany is taking this leap from a foundation where some municipalities still send official notices via fax.
What the world is saying about it
The political framework is clear: acceleration at any cost. The narrative goes like this—AI as the answer to a staffing crisis and public expectations that are chafing against Germany’s digital lag. The comparison with Estonia and its X-Road system has become so hackneyed that it obscures the real question: What happens when AI agents act autonomously before the underlying infrastructure is digital?
Austria is in a slightly better position—the ELAK at least provides digital record-keeping as a starting point. Switzerland takes a methodical approach: It checks for GDPR and DSG compliance before anything goes into production systems. Internally, this is seen as tedious; externally, as caution that pays off.
In Germany, however, the debate has not yet reached the point that matters: Who is liable if an AI agent provides an incorrect basis for a decision and a citizen suffers as a result? The Federal Constitutional Court has clear positions on algorithmic infringements of fundamental rights. The EU AI Act requires high-risk systems used in administrative decisions to undergo audits. The “Agentic AI Hub” has not yet made any public statements on how these requirements will be specifically implemented in the 18 pilot projects.
What we think about it
The program does not warrant skepticism regarding its goal—but rather regarding its methodology. The figure of an 80 percent acceleration is a political statement, not a metric. Without a defined baseline process, without a control group, and without independent evaluation, it cannot be verified. This is not a minor academic detail. It is the fundamental prerequisite for determining whether we can learn from the pilot projects—or whether we are merely producing press releases.
What concerns me structurally: The federal government is building an AI layer on top of non-digital infrastructure. This is not only inefficient—it is risky. AI agents that analyze documents and provide decision recommendations need clean, structured input data. Anyone who has to scan before they can evaluate has a media conversion problem that no LLM performance can solve.
For the DACH region as a whole, this is a learning moment: The question “How do we deploy AI in government agencies?” is secondary to “What fundamental requirements must be met for AI to function effectively in government agencies?” Austria and Switzerland have a different starting point here—and should leverage it instead of copying Germany’s reflex to rush ahead.
The most remarkable thing about “Colleague AI” is not what is being communicated. It is what is being left unsaid: liability, audit readiness, training for case workers, and a clear framework for human review in decisions relevant to fundamental rights. This silence is not a gap in the communication strategy. It is a substantive problem.
The stakes are high: If these pilot projects fail—due to inadequate infrastructure, lack of acceptance, or compliance issues with the EU AI Act—the failure will not merely affect “AI in public administration.” It will set back the entire public sector’s digitalization narrative by years.
The Verdict
| Criterion | Rating |
|---|---|
| Substance | ⭐⭐⭐⭐⭐ |
| DACH Relevance | Very High |
| Timeframe | Now |
Conclusion in one sentence: Germany wants AI agents in government agencies — the ambition is there, but the infrastructure isn’t, and questions regarding liability and EU AI Act compliance are being swept under the rug politically.
The AI Report is published every Friday. Subscribe for free:
👉 aisyndicate.ch/#/portal
