---
title: "Claims fraud prevention with data and AI"
description: "A model flags anomalies on claims with the reason and the agent prepares the file; the fraud unit investigates, the claims handler decides on the claim."
url: https://www.joinx.it/en/industries/use-cases/claims-fraud
---

# Claims anomalies reach the fraud unit with the file and the reasons

Many suspicions start with an experienced claims handler's intuition, and the file for the fraud unit contains what that handler had time to gather.

## Suspicion depends on who opens the claim

A notification made a few days after the policy was issued, the same witness in several claims, a loss adjuster's report that does not match the photos. Today the person who notices is usually the claims handler, who passes the claim to the fraud unit with the notes they have written. The unit then redoes the searches, looks in the claims system for earlier cases involving the same people, vehicles or repairers, and cross-checks archives that do not share an identifier. A suspicious claim opened by a less experienced colleague, on the other hand, can pass with no signal at all.

## The model brings the signals, the fraud unit decides on the investigation

In the example, a Data Intelligence model compares every new notification with the insurer's history and flags the anomalies, writing beside each one the reason and the related claims. For the cases flagged, the agent prepares the file with documents, earlier cases and links between the parties. The fraud unit decides whether to open an investigation, and a signal on its own never becomes the rejection of a claim, because the decision on the claim stays with the claims handler after the unit has given its opinion.

1.  The model compares the notification with the history and gives the reason for the signal
2.  The agent prepares the file with earlier cases and links between the parties
3.  The fraud unit decides whether to open an investigation
4.  The claims handler decides on the claim, never the signal alone

Illustrative example

## What it takes to flag anomalies on claims

| Item | In this case |
| --- | --- |
| Data | Notifications and loss adjusters' reports, the claims history with the parties involved, policy inception dates, the outcomes of investigations already closed |
| Modules | Data Foundation, Data Intelligence, Agents |
| Who decides | The fraud unit on the investigation; the claims handler on the claim; the head of the unit on the model's thresholds and on the signals to correct |
| What we measure | The share of flagged cases the unit judges well founded, the time taken to prepare a file and legitimate claims slowed down by a signal, against the starting point |
| AI Act risk class | The high-risk uses in Annex III that concern insurers are risk assessment and the setting of life and health premiums for individuals, and looking for anomalies in claims is not among them. The health data in claim files has separate access. We classify the case with risk and compliance before development |
| Where to start | With Deployment on a line of business with many notifications and with the outcomes of investigations on record; without those outcomes the model cannot be tested, and the work starts from the data |

## If the parties to claims do not have a single name

When the same repairer appears under three different spellings and the people involved have no common identifier across systems, the model does not see the links that matter, and we first build single master data for the parties with Data Foundation. On a line of business with few notifications, moreover, the fraud unit already knows the cases one by one, and the model adds little to what it knows.

The insurer's other processes are in [Insurance](https://www.joinx.it/en/industries/insurance), the claim that this check fits into is in the sheet on [claims from notification to settlement](https://www.joinx.it/en/industries/use-cases/claims-from-notice-to-settlement), and all the sheets are in the [use cases](https://www.joinx.it/en/industries/use-cases).

## Let's go back over an investigated case with the fraud unit

Starting from a claim on which the investigation has closed, we look with you at which signals were there and which systems they sat in.

[Contact us](https://www.joinx.it/en/contact?settore=assicurazioni)
