AI Engineer World's Fair 2026
AI-Driven Multi-Document Correlation for Enterprise Financial Compliance and Fraud Detection
Loading slides…
Read the talk
AI-Driven Multi-Document Correlation for Enterprise Financial Compliance and Fraud Detection
Individually valid records can conceal collective risk. Graph correlation, probabilistic scoring, and jurisdictional normalization connect the evidence compliance teams need.
From a talk by Varsha Shah
What does a valid document fail to tell you?
Can every document pass its own compliance checks while the organization still misses fraud? Payroll, tax, procurement, and transaction systems each hold part of the financial picture. When those records are analyzed independently, inconsistencies between systems can remain invisible. Document validity does not establish consistency across documents. This is the problem Varsha Shah’s framework addresses by combining graph-based entity correlation, probabilistic risk modeling, and cross-jurisdictional normalization.
The difficulty grows along three dimensions. Operations across countries introduce different reporting standards and compliance requirements. Increasing volumes of financial records make manual review impractical. Meanwhile, fraud can exploit subtle inconsistencies across systems rather than obvious errors in individual records. Traditional rules and document-level NLP can validate a record without understanding its relationships to the rest of the enterprise’s activity.
Consider Shah’s conceptual example: a payroll register, a vendor invoice, and a tax filing each pass their respective checks.
| Record | Individual validation |
|---|---|
| Payroll register | Payroll rules |
| Vendor invoice | Procurement policies |
| Tax filing | Tax regulations |
The payroll record appears accurate, the invoice legitimate, and the tax filing correctly submitted. Connecting them can nevertheless reveal inconsistencies that none of those checks exposes alone. The example does not specify a particular fraudulent transaction; it isolates the missing capability: understanding relationships among records whose individual contents appear acceptable.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Connect the evidence, score the risk, supply the context
The proposed architecture divides that work among three complementary components. It is a framework for combining capabilities, rather than a single model or algorithm.
| Component | Responsibility |
|---|---|
| Entity correlation engine | Connect related information across enterprise systems |
| Adaptive probabilistic risk model | Prioritize connected patterns using multiple risk signals |
| Cross-jurisdictional normalization layer | Supply consistent financial and regulatory context |
The correlation engine creates a unified view of activity. The risk model evaluates that connected evidence to identify cases deserving attention. Normalization addresses differences in currencies, tax structures, reporting standards, and compliance rules so that the resulting assessment has the appropriate context.
The graph-based entity correlation engine establishes the relational foundation. It connects employees, vendors, accounts, transactions, and regulatory filings across payroll, tax, procurement, and financial systems. The analytical object becomes a network of entities and relationships rather than a collection of isolated documents. Hidden patterns and structural anomalies can then become visible in those connections.
Establishing a connection is only the first step. The graph identifies what is related; a subsequent assessment must determine whether the relationship represents meaningful compliance risk. Shah describes this division of responsibilities without specifying a graph schema, an entity-matching algorithm, or a traversal procedure. The architectural distinction matters: relationship discovery supplies evidence for risk assessment rather than making every connection an alert.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Prioritize connected patterns for investigation
The adaptive probabilistic risk model combines several indicators into a confidence-based risk score:
- Anomaly strength: the strength of the detected irregularity.
- Source reliability: the reliability of the evidence contributing to the assessment.
- Historical patterns: information from previously observed activity.
This moves prioritization beyond an alert triggered by a single static rule. Multiple signals contribute to deciding which connected patterns are most likely to represent genuine compliance risk.
The score is intended to direct immediate attention toward higher-risk cases and reduce unnecessary investigations. Completed audits then supply feedback that can improve later assessments. The presentation describes the inputs, purpose, and learning relationship, but does not give a scoring equation, decision threshold, or calibration method. Its central mechanism is the combination of evidence for prioritization, followed by learning from investigation outcomes.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make records comparable without losing regulatory context
Cross-border comparison introduces a further problem: the same transaction can be interpreted differently under different jurisdictions. A risk assessment therefore needs more than connected entities and a score. It also needs an account of which financial conventions and regulatory requirements apply.
The cross-jurisdictional normalization layer harmonizes currencies, tax rules, reporting periods, and classification schemes. These transformations support consistent evaluation across financial systems, while the regulatory context determines how a risk should be interpreted locally. Comparable records and appropriate local interpretation are complementary requirements; standardization does not mean treating every jurisdiction as though it had identical rules.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Detection performance at enterprise record volumes
Shah reports evaluating the framework on approximately three million financial records collected over five years across four regulatory jurisdictions. The setting was intended to capture the complexity of large volumes of interconnected enterprise data under varying regulatory requirements.
The reported detection results address two different questions: how much of the flagged activity is genuine, and how much of the genuine fraud the framework finds.
| Metric | Reported result | Interpretation |
|---|---|---|
| Precision | Approximately 91% | Flagged cases confirmed as genuine anomalies |
| Recall | 87% | True fraud cases identified |
| F1 score | 0.89 | Balance between precision and recall |
These are aggregate results for the stated evaluation. Shah characterizes them as evidence of performance across complex enterprise environments; the presentation does not provide a per-jurisdiction breakdown.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Reduce the work generated by low-value alerts
Detection quality matters operationally because every alert can create work for an investigator. Shah reports a 76% reduction in false positives and approximately 40% less manual audit effort. Fewer legitimate cases entering investigation would let compliance teams spend more attention on high-risk cases instead of routine document review. Shah also describes faster investigations, better resource utilization, and greater confidence in decisions as resulting benefits.
The comparison with conventional rule-based approaches turns on the unit of analysis:
| Approach | Main capability | Treatment of relationships |
|---|---|---|
| Traditional rule-based validation | Check individual records | Struggles with patterns spanning documents |
| Proposed framework | Assess connected enterprise evidence | Combines correlation, scoring, and regulatory context |
Shah reports that the framework outperforms the baseline on key metrics. The presentation does not specify the baseline implementation, evaluation split, or denominators for the reported reductions, so the operational figures describe the presented evaluation rather than a forecast of staff-hours or cost savings for another deployment.
The proposed advantage also extends beyond the initial assessment. Completed audits and investigator feedback feed later decisions, allowing the framework to adjust instead of relying on static rules alone.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Use completed audits to improve the next assessment
Audit outcomes provide two complementary forms of feedback:
- Confirmed fraud: strengthens patterns used for future detection.
- False positives: helps refine risk scoring and reduce unnecessary alerts.
Together, these outcomes create a learning cycle between detection and investigation. As fraud patterns and business conditions change, the framework is designed to adapt through feedback rather than depend solely on manual rule updates.
That cycle supports a broader change in how compliance operates. In a reactive process, an audit, investigation, or regulatory review identifies an issue after it has occurred. Continuous monitoring combines cross-document correlation with adaptive learning to surface risks earlier and prioritize investigation. Compliance becomes an ongoing assessment of emerging risk rather than only a periodic review of past activity. Shah presents this as a path toward predictive governance; prospective prevention is an intended direction, not a separately measured result in the evaluation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Bring the framework into enterprise workflows
Turning the research framework into an enterprise capability requires four pieces to fit together:
- System integration: connect existing ERP, payroll, procurement, and tax platforms so that the framework can work across their records.
- Jurisdiction-specific configuration: account for local regulations and reporting standards.
- Audit workflow alignment: deliver prioritized, risk-scored cases that investigators can act on, instead of merely adding another layer of document review.
- Scalability: support the financial-record volumes large organizations generate.
Shah points to the evaluation’s record volume as evidence for the scalability consideration. The presentation describes deployment requirements rather than a production topology, throughput measurement, or rollout procedure.
The resulting design keeps the relationship between detection and operations explicit. Risks that exist between documents require connected evidence; graph correlation supplies that evidence, probabilistic scoring prioritizes it, and normalization gives it regulatory meaning. The value of that combination is realized when investigators receive more useful cases and their completed audits improve subsequent assessments. Shah’s closing ambition is to extend this cycle from detecting risk toward anticipating and preventing it as part of everyday financial governance.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Further reading
Shah's paper presents the three-component framework and reports detection and review-workload results.
A companion presentation explaining LLM-assisted interpretation, entity linking and an employee-relocation compliance example.
Full transcript
Loading transcript…