Senthil Babu Malli Jayaraj
State and federal agencies modernizing legacy IT systems increasingly consider introducing machine learning capability into those modernization programs, including document classification, fraud detection triage, eligibility screening support, and caseworker workload balancing. Unlike deterministic, rule-based system components, machine learning introduces probabilistic behavior, model drift over time, and, in higher-impact use cases, direct consequences for individuals and families that traditional IT governance frameworks were not originally designed to address. This research article presents a responsible AI adoption framework for public sector IT governance, structured around a risk classification approach positioning AI use cases by autonomy, decision impact, and reversibility, a tiered governance gate model spanning intake screening through pre-production validation, human oversight design principles, and a five stage AI governance maturity model. Findings are illustrated with a distinct set of three-dimensional visualizations, including a 3D scatter plot positioning illustrative AI use cases by risk dimension, a 3D wireframe plot of model drift over time, a 3D contour-surface plot of governance maturity, and a 3D stem plot of governance gate review findings, alongside extensive grid-style data tables. Drawing on federal AI governance guidance, NIST AI risk management standards, and public sector algorithmic accountability research published through July 2025, this article finds that agencies applying a tiered governance gate model calibrated to AI use case risk classification report substantially fewer post-deployment AI-related incidents and materially faster, more defensible adoption decisions than agencies applying uniform governance rigor regardless of use case risk, or no formal AI-specific governance at all.
Keywords: responsible AI, machine learning governance, public sector IT governance, algorithmic accountability, AI risk management, human oversight, legacy modernization, model drift