Machine Learning and Generative AI Applied to the Classification of Ombudsman Manifestations

The classification of ombudsman manifestations is essential for the screening and forwarding of demands, but its manual execution tends to be slow, costly, and subject to inconsistencies. Although classical Machine Learning (ML) models are widely used in this task, the advancement of Large Language Models (LLMs) has increased the interest in prompting-based approaches, still with few comparisons under rigorous protocols and with operational robustness evaluation. This work compares ML models and LLMs in real, synthetic, and hybrid scenarios. The pipeline included hyperparameter optimization with Grid Search, Optuna, and Evolutionary Computation, in addition to the final evaluation in shared holdout. The results showed ML superiority in the three scenarios. Parse failures and prediction collapses were also observed in LLMs, indicating that the comparison should consider predictive performance and operational robustness.