Vinicius Santana Gomes, a special advisor in Brazil's Federal District government, arrived at the Sectoral Internal Control Office of the Department of Health with a simple but pressing question: if artificial intelligence tools were free and available to any public servant, why wasn't anyone using them to work faster?
The answer, he discovered, revealed something counterintuitive about the barriers holding back AI adoption in government worldwide. The problem was never the technology itself—it was training. Across two Brazilian public agencies, Gomes developed and tested a structured pedagogical methodology that separated what the AI tool does, what the operator needs to do, and what the institution requires before any document leaves the unit. The results suggest that generative AI's potential in the public sector has less to do with breakthrough algorithms and far more to do with systematic human preparation.
When Gomes took leadership at the Health Department's Sectoral Internal Control Office in 2024, his team faced a concrete bottleneck: case processing averaged 17 days, 22 hours, and 11 minutes, according to the Federal District's Electronic Information System. It was a multidisciplinary team with no homogeneous legal background, manually grinding through a mounting caseload. Over the course of that year, Gomes implemented a four-layer structured training methodology using free AI models. By the end of 2024, average processing time had dropped 18.2%—a significant gain achieved without a single information-security incident, sensitive-data leak, or compliance challenge from external oversight bodies.
The method proved even more dramatically effective when applied to a second agency: the Internal Control Unit of the Federal District Department of Economic Development, Labor and Income, throughout 2025. Processing time fell by 50%. The unit's technical-report production surged 92%, staff issued 288 formal recommendations to public managers, and the office analyzed cases totaling US$104.3 million in financial volume. Again, internal control mechanisms flagged zero security incidents or data breaches.
What made this approach distinctive was its portability and its compliance-first design. Gomes built legal governance into every layer of the methodology, ensuring the work aligned with international and national data-protection law and the principles of Brazil's public administration. Critically, he used only free AI models, making the system accessible to government agencies operating under severe budget constraints—a reality that applies to most public sectors globally.
The implication reframes how governments think about digital transformation. Executives have long assumed that AI adoption delays reflected immature technology or insufficient computational resources. This Brazilian evidence suggests the actual bottleneck is far more tractable: structured training that teaches operators how to use AI safely, methodically, and in ways their institutions can audit and verify. Generic AI workshops had failed to close the gap. What worked was a custom methodology built specifically for the institutional context and legal requirements of public service.
Both datasets come from the Federal District's official information system and are verifiable by third parties, giving them the kind of institutional weight that public-sector decisions demand. The case studies suggest the method can be replicated across agencies with entirely different missions. If that holds true, the barrier to transforming public-sector productivity may finally be something governments can actually remove.