Validated on 3M real records: compliance risk lives between the documents
AI-Driven Multi-Document Correlation for Financial Compliance - Varsha Shah, Independent · Varsha Shah
19 min total·Actually worth watching closely: ~3 min·3 must-watch clips
- 0:04 – 3:20Listen
The real risk hides in the gaps between documents
The opening poses the core proposition: organizations generate more financial data than ever across payroll, tax, procurement, and transaction systems, yet compliance teams still struggle with hidden fraud patterns, because existing solutions analyze documents independently while the most critical risks only become visible when information is connected across multiple systems.
Compliance risk sits not inside a single document but in the relationships between documents, where document-level detection is inherently blind.
This stretch is the speaker setting up the problem verbally, with the screen essentially parked on the title slide, so listening while commuting loses nothing.▶ Jump to 0:04Speaker · Varsha Shah - 3:20 – 7:13Listen
Why neither rule engines nor document-level NLP is enough
Breaks down how most compliance systems operate today: a payroll register is validated against payroll rules, vendor invoices are checked against procurement policy, a tax filing is reviewed under tax regulation. If each document passes its own validation the transaction is considered compliant, yet sophisticated fraud patterns emerge only when multiple documents are analyzed together.
This is not a model-accuracy problem but a wrong unit of analysis. The information already exists; what is missing is the ability to understand the relationships between documents.
The argument runs on reasoning rather than diagrams, with the speaker drawing conceptual contrasts throughout, so listening alone keeps you on track.▶ Jump to 3:20Speaker · Varsha Shah - 7:13 – 9:50Listen
Three components: correlation, scoring, cross-jurisdiction alignment
The heart of the framework: the entity correlation engine answers what is connected and provides the relational foundation for everything else; the adaptive probabilistic risk model decides which of those relationships represent genuine compliance risk; and the cross-jurisdictional normalization layer standardizes currencies, tax structures, reporting periods, and classification schemes.
The three divide the work rather than stack on top of each other. Without the normalization layer, the same transaction reads differently depending on the jurisdiction and the risk scores lose comparability.
The architecture is delivered as a spoken division of responsibilities, with bullet text rather than a structural diagram on screen; catching what each component owns is enough.▶ Jump to 7:13Speaker · Varsha Shah - 9:50 – 11:40Skim
Evaluation: 91% precision, 87% recall
Results on approximately three million real financial records collected over five years across four regulatory jurisdictions: about 91% precision, 87% recall, an F1 score of 0.89, and consistent performance across all four jurisdictions.
What matters is not just the level of the metrics but that they hold across four jurisdictions, evidence the effect comes from the method itself rather than overfitting one set of local rules.
The screen carries a dense results table with metrics and evaluation conditions crammed onto one page; pausing to scan it is faster than following at speaking pace.▶ Jump to 9:50Speaker · Varsha Shah - 11:40 – 12:58Skim
False positives down 76%, audit effort down 40%
Translates the technical metrics into an operational ledger: false positives down 76%, manual audit effort down roughly 40%, so investigators can put their time into genuinely high-risk cases.
The return here comes mainly from reviewing fewer legitimate cases rather than catching more fraud. The investigation hours saved are the quantifiable source of ROI.
Both percentages and their business interpretation are printed on the same page; note the numbers at a glance, since the narration only expands on them.▶ Jump to 11:40Speaker · Varsha Shah - 12:58 – 15:50Skim
Baseline comparison and the continuous learning loop
First an item-by-item comparison against traditional rule-based systems, confirming a consistent lead across key metrics; then how the system improves itself, with confirmed fraud cases strengthening future detection patterns, false positives feeding back to refine risk scoring, and no manual rule updates required as fraud patterns evolve.
Static rules get bypassed by new fraud patterns, while the loop formed by audit and investigator feedback lets detection evolve along with the adversary. That is the most fundamental difference from a rule engine.
The first half has a side-by-side comparison worth reading; the second half on the learning loop is purely spoken, so you can scan the chart and then switch to listening.▶ Jump to 12:58Speaker · Varsha Shah - 15:50 – 18:58Listen
Four deployment considerations and the paradigm shift
Closes on the conditions for deployment: integration with ERP, payroll, procurement, and tax platforms; jurisdiction-specific configuration for local regulations and reporting standards; alignment with existing audit frameworks; and scalability to millions of financial records. It ends by arguing compliance should move from after-the-fact validation to predictive governance, then delivers the key takeaways.
Success on the ground often turns not on the model but on whether it plugs into existing ERP and audit workflows, and organizations need to move from asking what went wrong to asking what is likely to go wrong next.
The four recommendations and the closing outlook are all delivered verbally; only the final takeaways slide is worth a glance.▶ Jump to 15:50Speaker · Varsha Shah