The adoption of Artificial Intelligence (AI) in the Brazilian public sector has moved beyond the experimental phase, positioning itself as a strategic vector for digital transformation and the delivery of citizen services. Large-scale institutions such as Serpro, Caixa, and Banco do Brasil have articulated intensive journeys dating back to 2017, aiming not only for innovation but also for increased productivity and improved public service. This evolution is supported by ambitious national plans, such as the Brazilian Artificial Intelligence Plan (PEBIA), which outlines five strategic axes, including infrastructure, capacitation, and AI for public services. The challenge is twofold: achieving technological excellence in a scenario of rapid disruption while ensuring this transformation is ethical, inclusive, and sovereign.

Effective structuring begins with defining clear strategies that balance competitiveness and public purpose. Many organizations have established Centers of Excellence (CoEs) in AI and Data Science, often linked to Technology but with a strong bridge to business operations. Continuous capacitation is essential for the democratization of AI, allowing all employees, from the frontline to top leadership, to use tools coherently and understand associated risks. Programs like the "Serpro AI Journey" and the ENAP/MGI training tracks aim to train tens of thousands of civil servants, addressing the barrier of lacking specialized personnel and ensuring the digital literacy necessary for transformation.

Technically, AI and data governance form the bedrock of sustainability. Public sector systems handle massive volumes of sensitive and classified data, necessitating rigorous compliance protocols with the General Data Protection Law (LGPD). Governance must be by design, integrated into the solution's lifecycle, from prospecting (like Serpro's Ethics by Design model or BB's Prisma matrix) to post-deployment monitoring, which must be continuous and perennial to mitigate algorithmic biases and ensure decision traceability. The Federal Government’s AI Ethics Framework, aligned with OECD principles, establishes guidelines for risk assessment (excessive, high, medium, low risk) and the need for effective human supervision, even in automated decisions.

Digital sovereignty is a strategic pillar that demands robust infrastructure. The exponential data volume, compared to a "tax tsunami", exceeds human capacity, making AI systems a filter and a compass. This necessitates constant investment in high-performance hardware (GPUs, HPCs) and infrastructure supporting heavy Machine Learning and Inference processing. There is a coordinated effort to build a Brazilian "sovereign AI," including the development of Large Language Models (LLMs) trained in Portuguese with national data, ensuring that the processing of restricted data remains within national territory, in compliance with regulations such as GSI's IN 05/08.

In the Generative AI field, resource rationalization is essential. Large Language Models (LLMs) demand infrastructure and incur continuous inference costs, making model selection an efficiency decision. Serpro LLM, for example, focuses on providing Small Language Models (SLMs) or medium-sized models (under 20 billion parameters) which, in many use cases, demonstrate comparable efficacy to larger models but with significantly lower cost and memory demand. A crucial technique for injecting domain-specific knowledge into LLMs without the high cost of Continuous Pre-training (CPT) or Fine-Tuning is Retrieval-Augmented Generation (RAG). RAG uses embedding models and vector databases to dynamically enrich the user's prompt with relevant context, enabling the LLM to accurately respond to queries about internal documents or fixed knowledge bases.

Technical applications in government are diverse and complex. Computer vision and biometric systems, such as the Liveness service, are crucial for continuous authentication and preventing Deep Fake fraud, a growing challenge that requires constant model retraining and human curatorship to identify new attack patterns. At the Federal Revenue (Receita Federal), AI is employed for network analysis, clustering, predictive selection of taxpayers, and supporting self-regularization, optimizing auditor work and fiscal accuracy. Additionally, behavioral analysis in system logs and accesses, utilizing AI for anomaly detection, creates a robust cybersecurity layer, identifying atypical behaviors that could be vectors for social engineering attacks.

In conclusion, AI in the Brazilian public administration is characterized by a complex articulation between high technology and social responsibility. Success lies not only in adopting cutting-edge models but in the capacity to build a governance infrastructure that handles data quality, ensures technological sovereignty, and rationalizes investments. The future path involves consolidating specialized LLMs (such as the specialist model for the Official Gazette under development at Serpro) and expanding platforms that democratize AI use by all developers and business areas, ensuring the technology serves the ultimate purpose of a more intelligent, empathetic, and human State.