AI-ASSISTED DATA CLASSIFICATION FOR ENTERPRISE GOVERNANCE: BALANCING ACCURACY, PRIVACY, AND OPERATIONAL EFFICIENCY

Abstract

Data classification is the foundation of enterprise data governance: retention, access control, encryption, sharing and incident response all depend on knowing how sensitive a document is. Manual classification does not scale, and rule-based tools are brittle, while machine learning promises scale but raises new concerns, because training a classifier on sensitive content can itself create privacy exposure and because errors in classification have very different costs in different directions. This paper proposes the Governed AI Classification Framework (GACF), which treats accuracy, privacy and operational efficiency as three objectives to be balanced explicitly. GACF combines a governed sensitivity taxonomy, a privacy-preserving training architecture based on federated learning with gradient clipping and noise injection, confidence-based triage that routes uncertain documents to human reviewers, and a three-level audit structure covering governance, model and application. We examine the framework in a simulation of five business units with heterogeneous documents and four sensitivity levels. A federated model reached 94.4% accuracy without raw data leaving the units, compared with 96.4% for a centralized model trained on pooled data and 74.6% for rule-based keywords. Adding privacy noise reduced accuracy gradually at moderate levels (88.5% at a noise multiplier of 1) and sharply at high levels (62.4% at 4), and the rate of under-classification, the most dangerous error, rose from 4.8% to 21.6%. Confidence-based triage raised overall accuracy to 96.8% while automating 76% of documents, which would cut reviewer effort for 10,000 documents from 250 hours to about 59. These results are synthetic and show the logic of the design, not performance on real documents. The paper closes with governance guidance and a research agenda.