Can Municipal AI Be Trusted? Lessons from an Agile Pilot in Romania

Artificial intelligence is rapidly finding its way into public administration, promising faster access to information and more efficient municipal services. Yet one fundamental question remains: can citizens trust AI when it provides official information?

This challenge became the focus of the PrimAssist agile pilot, implemented by Geseidl Consulting Group in Măgurele, Romania. Rather than creating another conversational chatbot, the pilot explored how AI can be made reliable enough for use in municipal administration, where accuracy and transparency are essential.

Building Trust Through Verified Information

The pilot adopted a fundamentally different approach from many public AI assistants currently available.

Instead of generating responses from general internet knowledge, the system was designed to answer questions exclusively on the basis of the municipality’s own official documents. Every response is accompanied by references to its source, allowing citizens to verify where the information originates.

To further reduce the risk of AI hallucinations, each response passes through a multi-agent verification process before it is delivered to the user. At the same time, privacy was treated as a core requirement: conversations are not stored, and the solution operates entirely within the municipality’s own infrastructure, ensuring compliance with GDPR requirements.

Promising Technical Validation

The pilot demonstrated that trustworthy AI for public administration is technically achievable.

The validation delivered encouraging results:

• 100% meaningful responses

• Zero connection errors

• No generic “I don’t know” answers

• Average response times between 11 and 14 seconds

While these figures highlight the robustness of the technical solution, they were not the pilot’s most important outcome.

The Real Innovation Lies in Knowledge Management

One of the strongest insights generated by the agile piloting process was that the performance of municipal AI depends less on the language model itself and far more on the quality of the underlying knowledge base.

The team concluded that carefully curated, structured and up-to-date municipal documentation has a much greater impact on the quality of AI responses than simply using a larger or more advanced model. As they summarised:

“Curation quality beats corpus size.”

This finding offers an important lesson for municipalities considering AI adoption. Before selecting an AI platform, local governments should first ensure that their documents, regulations and procedures are properly organised, maintained and accessible.

Agile Piloting Beyond Technology

The PrimAssist pilot illustrates how agile piloting can uncover insights that extend well beyond the technology being tested.

While the solution successfully demonstrated that AI can deliver accurate, transparent and privacy-conscious responses in a municipal environment, its greatest contribution was revealing that digital transformation begins with high-quality knowledge management rather than artificial intelligence itself.

As municipalities across Europe explore AI-powered public services, this Romanian pilot shows that building citizen trust is not primarily about choosing the most powerful model. It is about ensuring that every answer is accurate, verifiable and grounded in official information—the very principles that agile piloting is designed to validate before wider deployment.

Download Report