AI-Driven Data Governance

Abstract

Data governance -- the policies, processes, and technical controls that determine how data is collected, stored, used, shared, and deleted -- has become both more critical and more complex in the age of AI. AI systems consume data at unprecedented scale, require continuous data pipelines for retraining and monitoring, generate new data through their outputs (including synthetic data), and create novel privacy risks through memorisation and inference attacks. Simultaneously, AI can substantially automate data governance tasks that currently require manual effort: data cataloguing, lineage tracking, quality assessment, anomaly detection, and privacy compliance checking. This study evaluates eight AI-driven data governance tools across five governance functions: automated data cataloguing and metadata management, data quality monitoring and anomaly detection, privacy risk assessment (PII detection, re-identification risk), data lineage tracking, and GDPR/EU AI Act compliance checking. Tools evaluated include LLM-based schema annotators (GPT-4-Catalog, DITTO), graph-based lineage trackers (DataLineageGNN), ML-based data quality monitors (GreatExpectations-ML, Anomalo), PII detection systems (Microsoft Presidio, Amazon Comprehend), and an integrated governance platform (DataHub-AI). Evaluation used enterprise data environments from 8 European organisations across four sectors. GPT-4-Catalog achieves 88.4% accuracy on metadata annotation across 14,840 schema elements. DataLineageGNN tracks 94.2% of data transformation lineage vs. 68.4% for rule-based baselines. Anomalo detects 84.6% of data quality issues 2.4 days before they impact downstream pipelines. A governance maturity model with AI integration levels is proposed