Uso de IA generativa na geração automatizada de relatórios em processos de prevenção à lavagem de dinheiro

Abstract

This work presents the development and implementation of a solution based on generative arti- ficial intelligence for the automated generation of reports in money laundering prevention (MLP) processes. The research is justified by the increasing regulatory complexity and the high volume of financial transactions, which hinder the manual detection of suspicious operations. The main objective was to propose a cloud-based architecture, using AWS services and generative Al mo- dels, capable of processing structured data in JSON format and producing detailed, standardized reports following the guidelines of BACEN and COAF. The methodology involved filtering regis- tration, transactional, and contextual data, designing optimized prompts for interaction with Al, and implementing a serverless architecture with parallel processing of multiple dossiers through Amazon SQS. The results demonstrated that the system is functional, scalable, and efficient, generating encrypted reports in an average of 90 seconds. Furthermore, the application con- tributes to reducing the workload of human analysts, increasing efficiency and accuracy in risk analysis. It is concluded that the proposed solution is feasible as a support tool for financial compliance, although it still requires human supervision to avoid misinterpretations. As future work, it is suggested to integrate anomaly detection models and adapt the formatting of the generated reports to the standard required by COAF, expanding its applicability in regulatory environments. The processing flow employs asynchronous queues to allow parallel execution of multiple dos- siers, while encryption mechanisms guarantee the confidentiality of the documents generated. Specific instructions were adopted in the prompt to avoid hallucinations and to record as "not consulted" any missing data in the input JSON. The results demonstrated that the solution is functional, scalable and technically efficient, achieving an average time of around 90 seconds to generate encrypted reports ready for review. In addition to reducing the operational burden on analysts, the system increases the standardization and precision of analyses, reinforcing its applicability as a tool to support financial compliance.