AI case study
Ministry of Development, Industry, Trade and Services, and the Secretariat of Foreign Trade of Brazil
With traders waiting up to four days for answers on import and export procedures, Brazil’s Ministry of Development, Industry, Trade and Services (MDIC) and the Secretariat of Foreign Trade (SECEX) deployed Lina, an AI-powered virtual assistant on the Siscomex portal that provides immediate, 24/7 guidance grounded exclusively on official government sources. Lina supports the implementation of the World Trade Organization (WTO) Trade Facilitation Agreement and Brazil’s New Import Process (NPI).

The Problem
Users of Brazil's Siscomex portal, the country's foreign trade single window, have typically relied on a range of information sources to navigate import and export procedures. These included manuals, legislation, FAQs, news releases and a centralized helpdesk service "Comex Responde", Brazil’s national trade enquiry point. These sources are issued and maintained by multiple government agencies. Although these channels provided reliable information, users, including importers, exporters, customs brokers and logistics operators across the country, often faced difficulties identifying the relevant guidance among the large volume of legislation and documents.
In addition, the Comex Responde service relied on manual submission, triage and directing of enquiries. Users were required to identify the competent authority and complete a detailed form before submitting a question, creating barriers to access and increasing administrative workload. Between January 2022 and May 2023, more than 4,100 enquiries were processed, with an average response time of approximately four days, while many users needed immediate operational guidance.
The challenge became more pressing as the nationwide implementation of Brazil’s NPI accelerated, generating a significant increase in the volume and complexity of questions from traders, customs brokers and logistics operators. Providing faster, more accessible and scalable access to official trade information therefore became a priority.
The Solution
Lina is an AI-powered virtual assistant developed by the MDIC and SECEX and made available through the Siscomex portal. It answers questions submitted in natural language on import and export procedures, customs requirements, licensing processes, the NPI, the Single Import Declaration and other foreign trade topics.
This solution was built using Microsoft Copilot Studio and grounded exclusively on official government sources. Users submit questions in natural language through a chat interface; Lina then uses retrieval-augmented generation (RAG) combined with large language model technology to retrieve information from official manuals, regulations, FAQs and guidance documents, and generates a user-friendly response. When the system cannot answer a question with confidence, users are directed to formal government support channels. Lina does not make regulatory, licensing or customs decisions. All official decisions remain under the authority of the competent government agencies and users are clearly informed that Lina's responses are not legally binding.
SECEX selected a virtual assistant after evaluating other alternatives, including simplifying online forms and improving the existing enquiry management system. None of these significantly reduce response times or provide support outside business hours. An AI-powered virtual assistant was selected because it could provide immediate responses at scale, interpret questions written in natural language and guide users through complex trade procedures. This provides a cost-effective way to improve access to information while supporting Brazil's broader trade facilitation reforms to digitalize trade processes and implementation of the WTO Trade Facilitation Agreement.
The delivery model did not require custom AI model development or dedicated infrastructure. Microsoft Copilot Studio is available to Brazilian federal agencies through the federal government’s digital transformation initiative, funded through existing government digital transformation resources and using internal staff. Ongoing costs are limited to maintaining the knowledge base, monitoring performance, implementing improvements and platform licensing.
Lina is not directly integrated with the transactional functions of the Siscomex portal. Instead, the AI assistant operates at the information level, providing a unified access point to official content published by the agencies participating in Brazil’s foreign trade ecosystem. In this way, the solution did not require changing existing operational systems or data exchange standards, or the need for data sharing between agencies. Nevertheless, institutional coordination was required to identify authoritative sources and establish content governance.
The MDIC led the overall project and defined the business requirements. The team responsible for Comex Responde within the Department of Export Promotion and Trade Facilitation, led the design, configuration, testing and deployment, while making continuous improvements, with support from the MDIC’s information technology (IT) team and guidance from Microsoft.
The project was conceptualized in 2024 and moved through planning in the first half of 2025. It was developed in mid-2025 and after a brief test period, was publicly launched in October 2025. The service was temporarily suspended from early July 2026, however, and will resume after the end of the electoral restriction period in October 2026 to comply with legal restrictions on government communications during Brazil’s electoral period. Upon resumption of operations, the project team's priority is to further enhance response quality and user experience.
Overview
| AI type | Generative AI, retrieval-augmented generation |
|---|---|
| Training data | Grounding data for RAG obtained from official government sources: legislation, operational manuals, FAQs and guidance documents |
| Platform | Siscomex portal built on Microsoft Copilot Studio |
| Core functionalities | Natural language guidance on import and export procedures and other foreign trade topics |
| Primary users | Importers, exporters, customs brokers, freight forwarders and other foreign trade operators |
Challenges
Accuracy of AI-generated responses
Early testing showed that the assistant could occasionally generate incomplete answers or prioritize less-authoritative sources. This was addressed through continuous testing, refinement of prompts, improvements to the knowledge base and close monitoring of user interactions.
Building a reliable knowledge base
The knowledge base had to be assembled from a large volume of official documents produced by different government agencies, requiring institutional coordination to identify authoritative sources and content from governance processes.
Limited prior experience with generative AI
The solution was developed by a small multidisciplinary team with no previous experience with generative AI, requiring a rapid learning process supported by technical guidance from Microsoft.
Regulatory and policy requirements
The project had to take into account data protection requirements, government digital service policies, transparency obligations, responsible AI principles and information security requirements. Risks were minimized by relying exclusively on official public information and avoiding autonomous decision-making.
Best practices
Grounding in official sources
Lina relies solely on official government content (legislation, operational manuals, FAQs and guidance documents), promoting consistency and ensuring all users access the same official information regardless of their profile or experience.
Strong business ownership
The Comex Responde team defined business requirements, curated the knowledge base and supervised testing and implementation with the close involvement of experts, ensuring the knowledge base reflected authoritative guidance.
Complement, not replacement
The assistant was introduced as a complement to the existing national enquiry point, preserving users’ ability to seek human support, with escalation whenever a question cannot be answered with sufficient confidence.
Iterative development and continuous monitoring
Extensive testing, monthly review of user interactions, periodic review of AI-generated responses, and regular knowledge base updates improved the quality and reliability of responses over time.
Leveraging existing government platforms
Building on Microsoft Copilot Studio, available through the federal government’s digital transformation initiative, avoided the need for custom software development and additional IT infrastructure.
Internal capability building
The project team developed new skills in prompt design, knowledge base management, AI testing and performance monitoring, and gained experience in identifying and mitigating issues such as hallucinations and source prioritization, with training provided by Microsoft throughout development.
The impact
Immediate, round-the-clock access to official guidance
After its launch in October 2025, Lina recorded more than 2,500 chat sessions in its first months of operation, with an average of approximately 600 sessions per month from October 2025 to June 2026. Users receive information instantly rather than waiting days for a response, as was the case through traditional support channels, and the solution has significantly expanded the availability of support by providing assistance 24/7. Lina has also reduced the need for users to identify the relevant agency before asking a question, improving access to official trade information through immediate, consistent and user-friendly natural language interaction.
Reduced administrative burden and improved public services
By automatically handling a large share of routine enquiries, Lina has reduced the volume of questions requiring manual processing through Comex Responde, allowing public officials to focus on more complex cases and on the continuous improvement of guidance materials. An unexpected benefit was the generation of valuable insights regarding user needs: conversation analytics revealed recurring questions and knowledge gaps, helping SECEX improve its documentation, FAQs, communication strategies and training materials.
Generative AI can significantly improve access to public services, but its success depends on strong governance, reliable content and clear mechanisms for human support whenever needed.
Glossary
| Generative AI | AI that creates new content based on the data it has learned from, producing new outputs in response to a prompt. |
|---|---|
| Grounding | Practice of anchoring an AI model's response to verified source material rather than letting it rely solely on its internal training data. |
| Hallucinations | AI generates incorrect, misleading or fabricated information while presenting it as factual. |
| Large language model | An AI model trained on vast amounts of text data so that it can understand, generate, summarize, translate and answer questions using human language. |
| Retrieval-augmented generation | A technique that lets large language models find and use new information from outside data sources before answering a question. |
| Single window | A digital platform that lets business submit standardized information and documents once at a single entry point to fulfil all import, export and transit regulatory requirements. |
Footnote
This case study is based on the written submission by MDIC and SECEX.
← Back to Electronic Commerce and Digital Technologies
Created in September 2026