Case Studies / Polish Public Institution
GOVERNMENT & PUBLIC AFFAIRS · CITIZEN SERVICES
Multi-agent chatbot replacing manual citizen inquiries across 11 knowledge areas
A leading Polish technical inspection authority receives thousands of citizen inquiries about permits, equipment safety requirements, and certification procedures. Citizens ask in everyday language; the answers live in dense, technical documentation spread across multiple departments. The institution’s previous chatbot couldn’t bridge this gap – it required citizens to choose a category upfront (which most couldn’t), and failed on tasks like date calculations. We designed a new multi-agent architecture and built a working prototype that routes questions intelligently across 11 specialized knowledge areas without requiring citizens to know the right department.
Client: Leading Polish technical inspection authority (under NDA) – responsible for equipment safety, operator permits, and technical compliance across industrial and public infrastructure.
KEY RESULTS
11
Specialized knowledge areas covered by domain-specific agents
Auto
Intelligent routing – citizens ask freely, no category selection
Plain→Tech
Everyday citizen language matched to technical documentation
INDUSTRY
Government / Technical Inspection
USE CASE
Citizen-facing multi-agent chatbot
AI APPROACH
Multi-agent architecture + RAG
KNOWLEDGE SCOPE
11 specialized domains
PLATFORM
Azure, OpenAI, Qdrant, Langfuse
ENGAGEMENT
Architecture design + working prototype
The challenge
The institution oversees safety and compliance for technical equipment across Poland – from industrial machinery and pressure vessels to elevators and amusement rides. Citizens and businesses contact them with questions about what permits they need, which safety requirements apply to their equipment, how to apply for inspections, and what the deadlines and procedures are.
The answers to these questions are spread across dense technical and legal documentation that uses precise, domain-specific language. But citizens don’t ask in that language. They describe their situation in plain, conversational Polish – “I’m installing a new boiler in my building, what do I need?” – and expect the system to figure out which regulations, permits, and procedures apply.
The institution had an earlier chatbot, but it hit fundamental limitations. It required citizens to select a category before asking a question – but most people didn’t know which category their question fell into. It couldn’t handle date calculations (critical for permit deadlines and validity periods). And it struggled to combine knowledge from multiple domains when a question crossed category boundaries.
The core problem: citizens asked questions in everyday language about complex technical regulations. The previous chatbot required them to know the right category upfront, couldn’t compute dates, and couldn’t combine knowledge across domains. The result: citizens gave up and called the office instead, keeping the manual inquiry load high.
What we built
We designed a new architecture from scratch – addressing the specific technical limitations of the previous system – and delivered a working prototype that demonstrated the approach across all 11 knowledge areas.
Multi-agent architecture with intelligent routing. Instead of one monolithic chatbot, we designed a system with a main orchestrating agent that routes questions to specialized agents – each with its own dedicated knowledge base covering a specific domain (permits, equipment categories, safety requirements, procedures, etc.). The citizen simply asks their question in natural language. The main agent determines which domain(s) are relevant and delegates accordingly – no category selection required.
Plain language to technical language matching. The core challenge was bridging the gap between how citizens describe their situation and how the technical documentation is written. A citizen asking about “the big metal tank in my factory” needs to be matched to documentation about pressure vessels, specific safety classes, and the relevant inspection procedures. We engineered the retrieval and prompting pipeline to handle this semantic gap reliably.
Dynamic query handling. The new architecture handles tasks that the previous system couldn’t: automatic date calculations for permit deadlines and validity periods, clarification of ambiguous questions through follow-up dialogue, and combining knowledge from multiple domains when a question spans categories.
Easy knowledge management for non-technical staff. The knowledge base is managed through editable SharePoint Lists – so domain experts within the institution can update, add, and maintain content without needing developers or IT support. This was a deliberate design choice: the system’s accuracy depends on current content, and the people who know the content best shouldn’t need technical skills to maintain it.
Secure deployment on Azure. The prototype was built on Microsoft Azure with Entra ID authorization and Langfuse monitoring – matching the institution’s security requirements and providing full visibility into how the system performs.
The results
BEFORE
Citizens had to choose a category upfront. No date calculations. Couldn’t combine knowledge across domains. Many citizens gave up and called the office instead.
AFTER
Citizens ask in plain language – routing is automatic. Date calculations built in. Cross-domain questions handled. Knowledge managed by non-technical staff via SharePoint.
The working prototype demonstrated that the multi-agent architecture solves the specific problems the previous system couldn’t handle. Citizens can describe their situation naturally and get accurate answers from the right knowledge domain – without needing to know the institution’s internal structure or terminology.
For any public institution or enterprise with a complex, multi-domain knowledge base and non-expert users, this project illustrates an approach that scales: rather than building one massive chatbot that tries to know everything, decompose the problem into specialized agents that each handle their domain well, with an intelligent routing layer that puts the right expert in front of the right question.
Technology used
More Case Studies
See how we’ve helped other companies
AUTOMOTIVE · LEADING LUXURY MANUFACTURER
Intelligent virtual assistant replacing manual planning queries across SAP and cloud systems
Product planners spent hours manually querying SAP BW and multiple data warehouses for every decision. We built a bilingual voice-and-text assistant that retrieves planning data on demand – no system expertise needed.
Hrs → Sec
Data retrieval
DE + EN
Voice & text
SAP BW
Integrated
MANUFACTURING · RADAWAY
Making email-based order processing reliable with LLMs
Staff were manually reading customer emails, identifying products, and entering orders by hand. We turned a promising AI prototype into a production system that handles it end to end, across languages, formats, and attachments.
-90%
Manual intervention
95%+
Match accuracy
LOGISTICS · FR. MEYER’S SOHN
Eliminating manual data extraction from thousands of daily shipping emails
Operations staff were manually reading German and English logistics emails to pull out routing and scheduling data, every single day. We built an AI pipeline that extracts, structures, and delivers the data automatically.
–80%
Manual effort
2 langs
DE & EN
On-prem
Deployed