Financial fraud patterns often appear across payroll, tax, procurement, and transaction records rather than inside one document.
2
The framework combines graph-based entity correlation, probabilistic risk scoring, and cross-jurisdictional normalization to connect records and interpret risk consistently.
3
Across approximately three million records from four jurisdictions, the framework reached 91% precision, 87% recall, and reduced false positives by 76% and manual audit effort by about 40%.
Summary
Varsha Shah presents a framework for finding compliance risks that individual document checks miss. It connects employees, vendors, accounts, transactions, and regulatory records in a graph, then uses multiple signals to score the risk of each connected pattern. A normalization layer accounts for differences in currencies, tax rules, reporting periods, and classification schemes across jurisdictions. Shah says the framework was evaluated on approximately three million records collected over five years across four jurisdictions. It achieved about 91% precision, 87% recall, and an F1 score of 0.89. The evaluation also reported a 76% reduction in false positives and about a 40% reduction in manual audit effort. The model uses completed audits and investigator feedback to adjust future scoring. Shah also describes the integration, local configuration, audit workflow, and scalability requirements for enterprise deployment. Her argument is that compliance systems should connect evidence across systems and identify likely risks before they become audit findings.
Shah says enterprises now produce large volumes of payroll, tax, procurement, and transaction data, yet compliance teams still miss hidden risks because most systems analyze records independently. A payroll record, vendor invoice, and tax filing can each pass its own checks while revealing an inconsistency when connected. Traditional rule-based systems and document-level NLP validate individual records, but they do not understand relationships across documents. The framework therefore starts from the idea that the missing information is the connection between records, not more isolated validation of each record.
The proposed architecture combines an entity correlation engine, an adaptive probabilistic risk model, and a cross-jurisdictional normalization layer. The first connects information from payroll, tax, procurement, and financial systems. The second evaluates the connected data and prioritizes cases using several risk signals instead of a single rule. The third standardizes currencies, tax structures, reporting standards, and compliance rules so that transactions receive consistent treatment across jurisdictions. Shah presents the components as a pipeline from raw enterprise records to compliance intelligence.
A graph reveals relationships among enterprise entities
The graph-based entity correlation engine creates a unified network linking employees, vendors, accounts, transactions, and regulatory files. Shah describes its basic question as, "What is connected?" The graph can expose structural anomalies and hidden patterns that are hard to see in document-by-document review. These relationships form the basis for later risk assessment. The engine is intended to connect records across several enterprise systems rather than replace those systems with one new document format.
After the graph establishes relationships, the adaptive probabilistic risk model estimates which connections indicate meaningful compliance risk. Shah names anomaly strength, source reliability, and historic patterns as examples of the signals used to calculate a risk score. The system prioritizes cases that need attention and aims to reduce unnecessary investigations. It also learns from completed audits. Confirmed fraud cases strengthen future detection patterns, while false positives help refine scoring and reduce unwanted alerts.
Normalization gives cross-border transactions a common context
The normalization layer addresses differences between countries and regulatory environments. It harmonizes currencies, tax rules, reporting periods, and classification schemes. Without this step, Shah says the same transaction could be interpreted differently depending on where it originated. The layer answers how a risk should be understood within the relevant regulatory context. This allows the other parts of the framework to operate across multiple jurisdictions without treating local rules as interchangeable.
The evaluation reported strong detection and operational results
Shah says the framework was tested on approximately three million financial records collected over five years across four regulatory jurisdictions. It achieved approximately 91% precision, 87% recall, and an F1 score of 0.89. She also reports a 76% reduction in false positives and an approximately 40% reduction in manual audit effort. The stated operational effect is that investigators spend less time reviewing legitimate cases and can focus on higher-risk cases.
Continuous feedback moves compliance earlier in the process
The framework updates its behavior using completed audits and investigator feedback instead of relying only on static rules. Shah describes a cycle in which confirmed fraud improves future pattern detection and false positives improve risk scoring. This supports a shift from finding compliance issues after audits or investigations to identifying likely problems earlier. In her formulation, organizations can ask what is likely to go wrong next, rather than only asking what went wrong.
Enterprise deployment depends on integration and local configuration
Shah identifies four deployment considerations: integration with ERP, payroll, procurement, and tax platforms; jurisdiction-specific configuration; alignment with the audit framework; and scalability. Investigators should receive prioritized risk cases instead of manually reviewing every document. She points to the evaluation on millions of records as evidence that the design targets large enterprise environments. These requirements connect the research architecture to existing systems and audit processes.
"The information already exists. What missing is the ability to understand the relationship between these documents."04:30
Who should watch
You work on compliance, fraud detection, or financial governance and your data is spread across payroll, tax, procurement, and transaction systems.
Your current process flags too many legitimate cases and you want to prioritize investigations with connected evidence.
You are designing an enterprise AI system that must account for different jurisdictions and fit into existing audit workflows.
