Chat PaT

How Chat PaT Works: The Virtual Assistant for Provincial Public Services

What it is?

The Chat PaT virtual assistant is a service of the Autonomous Province of Trento designed to help citizens, businesses, professionals, and other entities navigate the public services offered by the Province.

Through a simple and accessible chat interface, the assistant allows you to:

• ask questions in natural language (writing as you would to a human operator);
• get immediate answers drawn from the official content of the Province's Service Catalog;
• receive direct links to detailed pages and online forms;
• change language (Italian, English, German, French, Spanish, Portuguese);
• ask for support from the Information Point, when necessary.

The assistant does not replace office staff but strengthens contact channels, providing accurate and updated information and reducing search times.

It is a quick access point, available at all times, guiding you step-by-step in discovering provincial public services.

What you can ask

The virtual assistant helps you to:

• find information on provincial services, for example, contributions, authorizations, qualifications, concessions, declarations, etc.;
• understand how to access the services and what requirements are needed;
• receive direct links to the official pages of the Province's Service Catalog;

If your question requires further investigation or human intervention, the assistant suggests you contact an operator.

How the chat works

1. Write your question in the chat space, trying to explain clearly and precisely what you are looking for.
2. Receive the answer immediately, with any useful links or references.
3. If you wish, you can rate the answer with a thumbs up 👍 or down 👎, to help improve the service.
4. You can also write your question in other languages: English, German, French, Spanish, and Portuguese.

The assistant is always available, every day and at any time.

Source of information

The assistant's answers come exclusively from official sources of the Autonomous Province of Trento, particularly the Provincial Service Catalog.

The assistant is continuously improving, but may provide incorrect answers: always check official sources.

Security and data protection

Conversations with the assistant are secure and confidential.

No personal data is stored, and the information you write is used only to offer you the best possible support and to improve the service over time.

In compliance with EU Regulation 2016/679 (GDPR) and national regulations on personal data protection, the advanced interaction system will adopt a data minimization approach. This principle provides that the chatbot tracks and archives exclusively the information strictly necessary for service improvement and statistical purposes, avoiding any superfluous processing of personal data.

Minimization Principles

• Limitation of Personal Data Collected
  • The chatbot neither requests nor stores users' personal data, such as names, emails, or phone numbers. Any personal data collected indirectly/involuntarily (i.e., for example, voluntarily provided by the user in the conversation) are autonomously analyzed and intercepted by a dedicated algorithm, in order to remove them before saving the conversation.
  • Interactions are analyzed anonymously and/or aggregated to ensure user privacy and comply with GDPR regulations.
• Tracking of Information Necessary for Statistical Purposes
  • Aggregated and anonymous usage data are collected and analyzed, including, by way of example:
    • Number of daily interactions.
    • Type of requests (e.g., search for specific services, generic questions, support requests).
    • User ratings on the provided answers (through the integrated feedback system).
  • This data is used exclusively to improve the service, optimize responses, and refine the interaction model.
• Data Management and Retention
  • The collected data, i.e., the contents of the conversations between the user and the chatbot, are subjected to an automatic anonymization process upstream: as soon as the user's request arrives, it is immediately anonymized before being processed by the semantic router and topic analysis.
  • All anonymized data is archived on an ACN-certified Azure infrastructure, ensuring security and compliance with national cybersecurity standards.

Technical Architecture

The technical implementation of the advanced interaction system is based on the Microsoft Azure platform, which provides a scalable, secure, and compliant infrastructure.

The key components of the architecture include:

• Certified cloud environment: The infrastructure is hosted on Microsoft Azure, qualified by the National Cybersecurity Agency (ACN), ensuring reliability, data protection, and compliance with Italian and European regulations.
• Hosting and Service Management: The service is managed via Azure App Service, a serverless platform that allows the automatic deployment and scaling of web applications.
• Natural Language Processing (NLP) Engine: Azure AI Language services are used to process user requests.
• Information Retrieval Module (Retrieval): The chatbot employs a Retrieval-Augmented Generation (RAG) architecture, leveraging Azure Document Intelligence for data extraction and Azure AI Search for indexing and retrieving information from the Service Catalog. The connection with the Service Catalog occurs via REST APIs implemented with Azure Functions.
• Response Generation Engine: The system processes responses using Azure AI Foundry with Azure AI Language and LLM models, integrating semantic chunking and extraction functionalities. Information is indexed via Azure AI Search after being processed by Azure Document Intelligence and stored in Azure Storage.
• Continuous Update System: The chatbot automatically synchronizes with editorial changes to the Service Catalog by periodically querying the REST APIs.
• Monitoring: Azure Monitor ensures continuous performance control, collecting data on requests, errors, and response times, with anomaly detection and automatic alerts.

The Azure AI Foundry platform offers access to a wide range of Large Language Models (LLM) from different providers (OpenAI GPT-4, OpenAI GPT-3.5, Meta Llama 2, Meta Llama 3, Mistral, Azure AI Models, Cohere), allowing comparative tests to optimize performance and costs.

The chatbot's user interface is developed with Next.js and hosted on a dedicated minisite, accessible via a custom domain and integrated with the digital ecosystem of the Autonomous Province of Trento. Azure App Service (Azure Static Web APP) offers a flexible environment for application development.

For environment management, an approach based on two distinct environments is adopted: development (DEV) and production (PROD). The DEV environment includes Azure AI Search with test indexes, Azure OpenAI Service with dedicated deployments for testing, Azure Blob Storage for test document corpora, Azure Functions in development, and Azure App Service with staging slots.

The PROD environment is configured with production indexes, production deployments for Azure OpenAI Service, Azure Blob Storage with geographical replication, Azure Functions, and Azure Monitor/Application Insights for continuous monitoring. Deployment between DEV and PROD is managed via Azure DevOps for CI/CD, Infrastructure as Code with Azure Resource Manager, and Automated Testing.

In application of the risk-based approach adopted by the AI Act (EU Regulation 2024/1689), a risk assessment was conducted for the ChatPaT advanced interaction system considering numerous aspects, including: the technology employed, the level of autonomy, data management policies, the type of target users, as well as all the technical specifications listed above. The analysis led to the final classification of ChatPaT as a low-risk system.

System Improvement

The Chat PaT virtual assistant was designed and implemented as a system based on a Retrieval-Augmented Generation (RAG) architecture to simplify access to provincial digital services. Although the system is active and usable by users, it is currently in a version subject to continuous improvement and evolutionary maintenance.

In this phase, particular attention will be paid to user questions and feedback to refine and optimize responses.

The advanced interaction system enables the collection of metrics, feedback, and conversation logs in anonymized form for monitoring and continuous improvement activities.

1. Data Collection and Anonymization: User questions, along with conversation logs, are recorded and saved exclusively in anonymized form. The system respects the principle of data minimization (Privacy by Design and by Default). Conversation contents are subjected to an automatic anonymization process: a dedicated algorithm analyzes and intercepts any personal data involuntarily provided by the user to remove them before saving the conversation.
2. Pattern Identification (Gap Analysis): Aggregated and anonymous usage data – including the number of daily interactions, the type of requests, and user ratings on the provided answers (via the integrated "thumbs up/down" buttons) – are analyzed. This analysis seeks to identify interaction patterns and the most critical thematic areas.
3. RAG Performance Refinement: The goal of the refinement process is to diagnose any problems related to specific categories of documents and apply targeted and cyclical interventions. This work is essential because the system operates in an environment of ** high semantic confusion** (the "administrative labyrinth"). The analysis of the domain gap between formal and colloquial language confirmed the intrinsic complexity of the administrative linguistic domain. The refinement process aims to strengthen the RAG's ability to translate the user's natural language into the actual administrative taxonomy.

Released system versions

Below is a list of the main releases that defined the system's activation and evolution.

• Version 1.0.0 (Release Date: August 28, 2025): First prototype of the Advanced Interaction System (AI Chatbot). This version marked the beginning of the AI Chatbot development, with the initial integration with the Service Catalog via API and the realization of the necessary support infrastructure for the system's operation.
• Version 1.1.0 (Release Date: 25 November 2025): Experimental version of the Advanced Interaction System (AI Chatbot) available to the public. RAG system operationalization on ACN-qualified Azure infrastructure. Partial compliance with Accessibility requirements (WCAG 2.1 AA), full compliance with GDPR and AI Act Transparency Obligations (Art. 50).

Context and objectives

The virtual assistant of the Autonomous Province of Trento was created within the framework of the PNC-A.1.3 Flagship Project “Digitalization of the Public Administration – Citizens and Digital Businesses Area (CID)”, financed by the Complementary National Plan (PNC) as an integration of the PNRR – Mission 1, Component 1, Investment 1.4 “Digital services and citizen experience” - CUP C49G22001020001.

The general objective is to simplify access to digital public services, improving:

• the findability of provincial services;
• the understandability of information and requirements;
• the ease of use and user experience on the Province's digital channels.

The virtual assistant fits into the “Trentino Digital District 2026” strategy as an immediate and continuous support tool for citizens, businesses, and professionals.

System Governance

The service was developed in collaboration between:

• Autonomous Province of Trento – Strategic Mission Unit for Digitization and Networks, project owner;
• Trentino.AI Scarl, subject in charge of the design, development, and maintenance of the system, contact email: info@trentino.ai;
• Tangible S.r.l. SB and Fondazione Hub innovazione Trentino, who conducted research and tests with citizens to understand how to develop the system's user experience.

The service governance is ensured by the working group composed of the Province's Strategic Mission Unit for Digitization and Networks and Trentino.AI, which monitors the system, defines update criteria, and manages user feedback.