The AI pilot that puts operations first

Twenty municipal AI pilots only count as a success when they deliver more than just a compelling demo. The Agentic AI Hub sets an interesting benchmark: operations, comparability, and reusability are already part of the pilot phase.

Key Takeaways

Related:Europe’s first exascale supercomputer: who gets to compute?  /  BSI C3A: Cloud sovereignty becomes verifiable

A pilot must factor in day-to-day operations

AI pilots rarely fail at the first demo. The friction starts later: Who monitors data flows? How do you keep a solution interchangeable? And how can one use case be compared to the next? The Agentic AI Hub tackled these questions head-on. On June 16, 2026, DigitalService announced the successful completion of a three-month phase involving 20 projects.

Nine start-ups collaborated with 19 municipalities. The scale was large enough to test more than just a single product-it had to reveal the realities of different administrative environments side by side.

Demand wasn’t the issue

According to DigitalService, around 400 applications from start-ups and nearly 200 from municipalities were submitted for selection. The number signals interest, not proof of impact. But it does show that the bottleneck in public administration isn’t a lack of ideas-it’s determining which ideas can be sustainably operated under real-world conditions.

That’s why the Hub deserves a different classification than a typical AI program. Participants didn’t just evaluate individual applications; they gathered insights into the prerequisites needed for broader future deployment.

Borken builds the layer beneath the use case

In Borken, deepset implemented a technology-agnostic AI orchestration layer using Haystack. The goal, according to DigitalService, is to ensure scalable and sovereign AI applications. It’s less flashy than a visible chatbot-but far more critical for operations.

An orchestration layer cleanly separates business applications, models, and data access. This keeps municipalities more agile when a model, provider, or process changes. The long-term architecture remains open, but the direction is right: first the operational foundation, then the growing number of use cases.

Comparability prevents pilot theater

Twenty projects only create impact if participants derive transferable criteria from them. These include data quality, roles, approvals, interfaces, and the effort required for ongoing operations. Without this perspective, a pilot remains a collection of isolated cases that are hard to compare.

The Hub identifies testing and scalability as its goals. Reliable metrics on processing time or cost savings aren’t published on the project page. That’s not a weakness to gloss over with estimates-it’s the point where the next phase must deliver robust operational data.

Reuse Starts Before Rollout

Reboot Germany thrives on projects that others can not only admire but also adopt. The Agentic AI Hub demonstrates a practical model for this: municipalities and start-ups test together while the platform question remains visible early on. This turns a local experiment into a foundation for further processes.

The same lesson applies to cloud and IT teams. A pilot needs an exit option, an operational model, and a clear description of data flows from the outset. DigitalService documents the pilot structure and the application areas involved.

In practice, this means: before the first test case, responsibilities, logging, and a path for corrections must be included in the plan. Only then can it be decided whether a model needs to be replaced, a new business process added, or an application discontinued. This discipline distinguishes a demo from infrastructure that can be sustainable across multiple municipalities.

Frequently Asked Questions

Image source: AI-generated (July 2026)