AI for public services: accuracy, empathy, and full traceability in digital tax advisory

Overview

From Human Overload to Intelligent Assistance

A provincial tax agency was overwhelmed by thousands of repetitive taxpayer questions each month. Answers varied by agent, consistency was low, and knowledge was siloed.

Together, we developed AsTri, an AI-powered conversational assistant capable of delivering legally sound answers based on an up-to-date normative database — complete with verifiable citations and a professional yet human tone.

More than a chatbot, a trusted digital legal advisor built for the public sector.

The Challenge

Democratizing Legal Knowledge Without Losing Precision

Human teams spent countless hours interpreting decrees and laws to answer similar queries. The goal was to design a system that could provide legally accurate, traceable, and consistent responses — while improving efficiency and trust. 

Main pain points:

• High volume of repeated questions.

• Risk of human error and inconsistent replies.

• Lack of transparency in how answers were generated.

• Heavy dependency on individual expertise.

The Solution

From Regulation to Conversation: AI That Speaks Legal Language

We built AsTri to validate whether an AI agent could safely assist taxpayers while staying within the boundaries of official legal frameworks.

Key features:

• Smart legal advisor: interprets laws, decrees, and resolutions, quoting sources directly.

• Natural and warm tone: conversational, yet technically rigorous.

• Traceability and control: every response includes cited sources and ranked relevance.

• Advanced document support: interprets complex tables and cross-references.

• Scalable, secure infrastructure: FastAPI + Vercel + Qdrant + MongoDB, ready for private cloud.

Key Results

Validating the Future of Digital Public Services

• +80% accuracy validated by the legal review team on complex tax cases.

• Advanced semantic understanding: maps equivalent legal terms (e.g., “candies” → “confectionery”).

• Reliable extraction of normative tables and structured data.

• Human-approved tone: warm, clear, and professional — indistinguishable from a live agent.

• 100% validation of initial hypotheses, including:

1. Answers are strictly limited to the knowledge base provided.

2. Source citations and order of legal relevance are accurate.

3. No hallucinations or speculative reasoning.

4. Readable and structured format for improved comprehension.

5. Consistent tone and behavior across all responses.