Document Intelligence and Fraud Scoring

Insurance fraud is rarely invisible. It hides in plain sight inside the documents that every claim produces -- and most of the time it goes undetected because nobody reads every document closely enough, quickly enough, before the payment decision is made.

That is the problem Mercury's document intelligence capability is designed to solve. When a claim comes in, the platform ingests the attached documents through NLP-powered document imaging, extracts the structured data inside them, and runs an AI fraud detection pass that scores each document from 1 to 100. That score is available to the adjuster at the same moment the claim record is opened. It does not require a separate investigation step, a batch report run overnight, or a referral to a different system.

Let me be specific about what the scoring reflects. NLP document imaging can detect anomalies in billing code patterns -- procedures coded together that rarely appear in legitimate claims, or billing amounts that diverge from regional norms for the reported type of loss. It reads repair estimates and flags line items that are inconsistent with the vehicle type, the reported damage, or the prevailing rate. It checks document metadata against the reported facts: a timestamp that postdates the incident, a provider address that does not match a licensed entity, a signature block that appears duplicated across multiple claims.

None of these signals is definitive on its own. That is exactly why the output is a score, not a verdict. A score of 72 means the document has elevated anomaly signals worth reviewing. A score of 14 means the document looks consistent with the reported facts. The adjuster sees the number, sees the flagged elements that drove it, and makes the decision. The platform's job is to make that review systematic and fast, not to replace the professional judgment that has to accompany it.

The business case for early scoring is straightforward. Fraud that gets detected at document review is fraud that does not get paid. Fraud that gets detected post-payment requires recovery efforts that rarely return full value -- and the organizational cost of running those investigations is substantial. The earlier in the claims lifecycle that anomalies surface, the less expensive the outcome.

There is a second benefit that is less discussed but equally important: consistency. Human document review varies with caseload, fatigue, and reviewer experience. An adjuster handling forty claims on a heavy day does not give every document the same scrutiny as an adjuster handling fifteen. AI scoring does not vary with volume. Every document on every claim gets the same pass. That consistency matters for both loss control and for audit trail quality -- carriers that face regulatory scrutiny or litigation can show a documented, repeatable process rather than a discretionary one.

The NLP and fraud scoring capabilities in Mercury are built on top of the platform's document imaging infrastructure, which handles ingestion from fax, email, upload, and API feeds. The fraud scoring layer sits on top of the extracted text and structured fields, not on the raw image, which means it works on any document format that Mercury can ingest. Policy documents, claim forms, medical records, repair invoices, COI attachments -- the scoring applies broadly, not just to a specific document type.

I want to be clear about what this is not. It is not a black-box AI system that autonomously denies claims. Mercury does not take adverse action on the basis of a fraud score without human review. The score is an advisory signal routed to the human in the workflow, and the workflow itself is configurable: carriers can set the threshold at which a high score triggers an SIU referral, a supervisor hold, or simply a notation. That configurability matters because the right threshold is different for a small TPA handling workers' comp than for a large carrier writing commercial property, and the platform should not presume to know which threshold is correct.

What Mercury does is close the gap between what the documents contain and what the adjuster has time to read. In a claims operation running at volume, that gap is where a meaningful share of fraud leakage lives. Closing it systematically -- on every document, every claim, every time -- is what document intelligence is actually for.

-- Sean Pitcher, CEO, Quick Silver Systems, Inc.