Inteligência artificial na gestão de contratos de obras públicas: previsão de riscos e disseminação de conhecimento para o público leigo.

Abstract

Approximately USD 13 trillion are contracted annually by Public Administration, with around 20% of this amount allocated to public works, including infrastructure construction projects. Despite existing regulations, it is not uncommon for public construction projects to experience delays or even abandonment before completion. Such issues not only lead to the waste of public resources but also negatively impact the economic and social development of the affected regions. Although citizens can contribute to Public Works Contract Management (PWCM)—for instance, by monitoring project progress and identifying potential issues—they might lack the necessary understanding of complex concepts required for effective participation. This limitation may result in missed opportunities for timely societal intervention and oversight. This thesis investigates how Artificial Intelligence (AI) can support PWCM. Specifically, it explores how AI can model risks, simplify complex project concepts, and communicate them to a non-specialist audience. The study employs a comprehensive approach, encompassing domain understanding, risk prediction, and knowledge dissemination to enhance AI-based support for PWCM. The domain understanding phase involved a Systematic Literature Review (SLR), analyzing 3,252 studies published between 2010 and 2024 to identify primary AI approaches applied in PWCM. This review mapped current AI applications, highlighted gaps related to public contracting phases, application objectives, algorithms used, and features considered in building Machine Learning (ML) models, providing a broad overview of predictive and generative AI techniques in the public works sector. Based on selected findings from the SLR, four ML models were developed and evaluated for predicting contractual failure risks in public works, using data from Paraíba, Brazil. Among these models, Random Forest, employing oversampling for data balancing, showed superior performance, achieving a ROC-AUC value of 0.75. To assist lay users participate in PWCM, an integrated system featuring a chatbot interface was developed to display information about public works and their associated risks. This system utilized a large language model (LLM) to answer user queries in simplified terms, facilitating understanding of project risks and details. Tested with 35 participants of varying PWCM knowledge levels, the chatbot demonstrated notable approval regarding clarity, practical usefulness, accuracy, ease of use, and reliability, thereby improving understanding of PWCM processes. This study advances the knowledge of AI applications in PWCM and provides guidelines for integrating such tools into government systems, thereby promoting more informed citizen participation in public affairs.