Serbia Government

eUprava provided a well-organized website with all government services listed. However, the broad scope and complexity of the business logic made navigating these services difficult to implement effectively, which often left citizens unsure where to begin. The team needed to make government digital services easier to locate, support dozens of services across multiple agencies, launch quickly to meet EU digital readiness deadlines, and build a sustainable system that could scale and be maintained by partners long-term.

Working with Rasa, eUprava launched a new conversational AI agent built entirely on CALM (Conversational AI with Language Models) using Rasa Studio. The agent covers 60 digital services across government functions, helping citizens get routed directly to the correct eUprava website service based on guided questions. The team used Studio to design and manage flows, keeping logic fully visible and versioned while allowing partners to build and iterate without heavy engineering involvement. GPT-4o, hosted within eUprava’s Azure environment, supports language understanding, while service logic remains fully deterministic and transparent.

Additionally, a Retrieval-Augmented Generation (RAG) component was implemented, enabling the assistant not only to start services but also to provide detailed information about them. For example, when a user asked, “What is the end date to apply for child testing for primary school?” the assistant responded, “The end date is August 15.” When the same user followed up with, “I want to sign up my child for primary school,” the AI agent immediately activated and routed to the primary school testing flow.

The initial go-live focused on delivering immediate value with a horizontal approach: covering as many services as possible with simple flows that gather key information and route users to the correct transaction page. Citizens can now schedule passport renewals, renew identification, access healthcare services, or submit job search requests. CALM allows eUprava to control how language interpretation connects to each service flow. LLMs handle intent recognition, while every task runs through defined logic that can be tested, monitored, and updated as new services come online.
With the first phase live, eUprava has established a strong foundation for long-term expansion. Citizens reach services faster without navigating multiple agency websites, government offices receive fewer misrouted or incomplete inquiries, logic remains fully transparent and versioned, partners build and maintain flows directly using Studio, and the system architecture supports future expansion without requiring system redesign.

As the assistant grows, CALM gives eUprava control over where and how LLMs contribute, keeping costs predictable and conversations reliable even as use cases become more complex.