# Special investigations unit assistant

> Expand the network around a suspicious claim, gather evidence across claims and documents, and draft the referral with citations. Investigators spend most of their time collecting facts.

Source: https://datanature.ai/use-cases/siu-fraud-investigation-assistant-insurance
Publisher: Dlytica Inc. (DataNature is a Dlytica company)
Updated: 2026-11-05

Industry: Insurance. Function: Risk and fraud. Built with: Voir, DataNature, DocuContext.

## What does the special investigations unit assistant do?

Investigators spend most of their time collecting facts. A Voir assistant for the special investigations unit expands the claim network through the knowledge graph, pulls prior claims, policies and documents, and drafts the case summary for the investigator to review.

## How is the agent designed?

- **Pattern**: Multi-agent: a graph researcher over claims and parties, a document researcher over claim files, an analyst over certified claims cubes, and a writer that uses only verified findings.
- **Data scope**: Claims Gold tables, the claims entity graph and claim document partitions, limited to the investigator's lines of business.
- **Human approval**: Referring a case, contacting third parties or recommending denial require investigator approval.
- **Platform components**: Voir, Knowledge graph, DocuContext, Semantic Layer, StreamHouse, Atlas

## How does it work on DataNature?

1. **Ground it in governed data.** Certified Gold tables and semantic-layer cubes give the agent consistent numbers; DocuContext partitions give it documents with links to source; the knowledge graph gives it relationships between entities where the use case needs them.
2. **Scope the twin.** The Voir twin's data scope, document partitions and SQL allowlist are enforced by the Agent Meta-Harness inside every tool call, and Atlas applies masks and row filters for the user it acts for.
3. **Route the model.** The AI gateway sends requests to an approved model endpoint and logs tokens, cost and latency.
4. **Keep a human in the loop.** The agent recommends; people approve actions that change systems.
5. **Trace and evaluate.** Every turn is traced, and answer quality is evaluated against real questions before and after each change.

## What data does it use?

- Claims and payments
- Claimants, providers, repairers and addresses
- Claim documents and notes
- Prior investigation outcomes

## Why run it on a sovereign platform?

The data, prompts, retrieved passages, model inference and logs all stay inside your jurisdiction, so the agent can work on regulated data without sending it to a foreign service.

## How do we get started?

Dlytica Consulting can scope and build it, NeuralShore engineers can join your team, and Dlytica Academy-trained talent is ready to support it. Book a demo with Dlytica to see it on your own data.
