Descriptions:
Varsha Shah, enterprise technical architect at Tata Consultancy Services working with Microsoft, presents research on a framework that moves financial compliance from document-level validation to cross-document intelligence. The central problem she identifies: sophisticated enterprise fraud rarely appears within a single payroll record, tax filing, or vendor invoice — it only surfaces when those records are correlated across systems. Traditional rule-based NLP validates documents independently and structurally cannot see these cross-document patterns.
Shah’s framework is built on three complementary components. An entity correlation engine constructs a graph of relationships across payroll, tax, procurement, and financial systems, creating a unified activity view rather than siloed records. An adaptive probabilistic risk model combines anomaly strength, source reliability, and historical audit outcomes into a confidence-weighted risk score that improves over time through feedback. A cross-jurisdictional normalization layer standardizes currencies, tax classification schemes, and reporting periods so that risk signals are interpreted consistently regardless of where a transaction originated.
The framework was evaluated on approximately 3 million financial records collected over five years across four regulatory jurisdictions, achieving roughly 91% precision in flagging genuine compliance anomalies. Shah frames the work as transforming compliance from a reactive, document-checking process into a proactive intelligence capability — one that can surface hidden fraud patterns that would otherwise require human analysts to manually correlate data across disparate enterprise systems.
📺 Source: AI Engineer · Published June 28, 2026
🏷️ Format: Deep Dive
