The Main Barrier to AI Adoption in the Public Sector is Lack of Training

Synthesis: Through Brazilian government case studies, demonstrates that a four-layer pedagogical methodology (Literacy, Protocol, Prompt Engineering, Audit) is the key to productivity gains (up to 50%), rather than premium models. This work emphasizes that AI Literacy is a developmental capacity requiring structured Scaffolding and Prompt Engineering discipline. It connects to the need for Curriculum Design that targets Metacognition and Agentic AI rather than just syntax mastery.

Key Findings

• Across two Brazilian federal-district government units with distinct mandates, official indicators recorded by the SEI-GDF system showed substantial productivity gains after the training method was applied: average case processing time fell by 18.2% at the Sectoral Internal Control Office of the Federal District Department of Health (SES/CONT) during 2024, and by 50% at the Internal Control Unit of the Federal District Department of Economic Development, Labor and Income (UCI/SEDET) during 2025.
• UCI/SEDET's gains were not only temporal: technical-report production rose 85%, the unit issued 286 formal recommendations to public managers, and it analyzed cases totaling US$94.8 million in financial volume — spread across the 122 technical reports consolidated for 2025, indicating increased analytical intensity rather than mere output volume.
• In neither unit did internal control mechanisms identify any information-security incident, sensitive-data leakage, or formal compliance challenge from external oversight bodies during the period examined.
• The determining barrier to adoption observed in these units was not technological but training-related: free, browser-accessible AI models were available to every public servant, yet were not being used productively until a structured pedagogical method was introduced.
• The four-layer method — literacy, protocol, prompt engineering, and audit — was designed to comply with international and national data-protection law and with the principles of public administration, supporting its portability across agencies with distinct mandates.
• A note-by-note review of the 286 recommendations (with classification by nature and materiality) identified material implications for US$43.7 million in payments and contracts analyzed; applying a probability matrix calibrated to public-audit literature, potential mitigation is estimated between US$1.1 million (conservative) and US$5.2 million (optimistic), with a central estimate of US$2.8 million.

Study Design & Method

The paper reports two auditable, third-party-verifiable cases rather than a controlled experiment. The method was applied throughout 2024 at SES/CONT — a multidisciplinary team with no homogeneous legal background that was manually processing a caseload whose average processing time the SEI-GDF recorded at 17 days, 22 hours, and 11 minutes — and throughout 2025 at UCI/SEDET. Outcomes were drawn from the official indicators of the Federal District Government's Electronic Information System (SEI-GDF). The two units differed in profile: at SES/CONT the method reduced processing time while keeping documentary output stable, whereas at UCI/SEDET the time reduction was accompanied by simultaneous growth in document volume and analytical depth.

What this means for practice

• Instructors. Sequence training as the paper's four layers — literacy, protocol, prompt engineering, and audit — so tool competence is paired with a disciplined, verifiable workflow instead of generic model access.
• Instructors. Keep critical review by a qualified professional as a functioning condition of use, with human review before any output is issued; the method accelerates output but does not replace technical knowledge.
• Administrators. Build the protection layer before deployment — a documented AI governance framework, de-identified inputs, and conversation training disabled on the tools used — since neither unit recorded data leakage, voided decisions, or external compliance challenges.
• Administrators. Invest in structured training rather than premium models: free, browser-accessible models were already available to every public servant but produced no productivity gain until the pedagogical method was introduced.

Limitations

• The evidence is two observational case studies rather than a controlled experiment: the author assumed direction of each unit at the start of its cycle, so a change of leadership and a reorganization of document workflows coincide with the method and cannot be separated from it.
• Outcomes rest on administrative indicators from the SEI-GDF system with no control unit or counterfactual, so the authors present the training-based account as the most parsimonious reading of the two cases rather than a demonstrated causal effect.
• The financial results are modeled, not realized: the US$1.1–5.2 million mitigation range (central estimate US$2.8 million) depends on a probability matrix drawn from international public-audit literature rather than local data, and the study does not track whether managers complied with the recommendations.
• Both cases come from Brazilian federal-district internal-control units with distinct mandates and baseline constraints, and the time gains diverged sharply between them (18.2% vs. 50%), so portability to other agencies or to educational institutions is untested.

Citation

Vinicius Santana Gomes (2026). The Main Barrier to AI Adoption in the Public Sector is Lack of Training.