Inteligência Artificial Generativa nos processos licitatórios do setor público: um estudo de caso no TJPE sobre a engenharia de prompt para artefatos de licitação e seus aspectos éticos e de transparência

Abstract

The implementation of Generative Artificial Intelligence (GAI) in the Brazilian Judiciary emerges as a response to modernization and efficiency challenges given the high procedural volume and bureaucratic complexity. However, its application faces significant obstacles related to lack of transparency and explainability, algorithmic biases, and the hallucination phenomenon - generation of apparently coherent but factually incorrect information. This study focuses on the Court of Justice of Pernambuco (TJPE) and the Tender Exemption modality, considering its quantitative relevance and potential for document standardization. In this context, the following research question was formulated: how can prompt engineering be structured to enable safe, transparent, and effective use of GAI in preparing administrative artifacts for direct contracting through Tender Exemption within TJPE? The general objective was to investigate how prompt engineering can be structured to enable ethical, transparent, and effective use of GAI in preparing administrative artifacts for direct contracting through Tender Exemption, based on a TJPE case study. The methodology is characterized as applied, exploratory-descriptive research with a mixed approach (qualitative-quantitative) and experimental design, based on literature review and document analysis. A hybrid architecture was developed integrating five prompt engineering techniques: Chain-of-Thought (CoT) for structuring legal-administrative reasoning, Retrieval-Augmented Generation (RAG) for consulting updated regulatory databases, Self-Refine for automated quality control, Few-Shot Learning for institutional standardization, complemented by Persona configuration (specialized legal analyst) and low temperature (≤0.2) to ensure precision and regulatory compliance. Although ChatGPT was initially considered due to its predominance in the Brazilian judiciary (96.4% usage among magistrates), limitations in generating complete documents led to selecting Claude (Anthropic) for experimentation. The Technical Report (TR) development followed Design Thinking methodology, structured in five iterative stages. Empirical validation consisted of generating five Terms of Reference representative of TJPE's exemption contracting, evaluated by three experts using a structured instrument covering legal compliance, technical quality, and institutional adequacy. Preliminary results indicate significant GAI potential in automated Terms of Reference preparation, with generated documents demonstrating high legal compliance (94% vs. 87% for manual preparation) and 75% time reduction in the preparation process. Implemented techniques proved effective in mitigating hallucinations, although human supervision remains indispensable. Limitations include the restricted number of analyzed samples and TJPE's specific institutional scope, constraining immediate generalization of findings. The research validates GAI applicability for optimizing administrative processes in the judiciary, contributing to greater operational efficiency and legal certainty in preparing procurement artifacts. The investigation emphasizes the importance of human supervision, algorithmic transparency, continuous validation, and specialized training for responsible and ethical technology use. The resulting RT constitutes a methodological guide for safe GAI adoption in other public institutions, aligned with algorithmic governance principles established by CNJ Resolution No. 615/2025.