Case studies · Insurance

An insurer puts a named approver in front of its claims agent

A claims assistant whose tool calls are checked, with payments and outbound letters held for a named adjuster, policyholder data redacted before the model, and each step on the record.

TowerApprovalsAgent actions waiting for a person

claims-assistant The payment for claim 20931 is ready. Shall I issue it?

Issue a claim paymentClaim 20931 · above the auto-pay limit High risk
Tool
issue_payment
Amount
$18,400 to R. Proto
Rule
AGENT-PAY-02, an adjuster approves payments over the limit
Expires in 1:00
Illustrative product view. Sample data, not client data.

The problem

The insurer built an agent to help adjusters gather documents, draft letters and prepare payments. Risk and compliance would not let it act on policyholder data or send anything outside without a person in the loop and a record they could show a regulator.

  • Namedadjuster on each payment and outbound letter
  • Redactedpolicyholder data before the model
  • Eachtool call on the record, with the rule that applied

Before Raidu

  • Agent kept in a sandbox
  • No control over which tools it called
  • No record a regulator could follow

With Raidu

  • Agent working on live claims
  • Payments and letters wait for a person
  • Each step signed and on the record

The challenge

The NAIC model bulletin on insurers’ use of AI, adopted by many states, asks insurers to run a written AI program with governance, risk controls and documentation. The insurer’s claims agent worked well in testing, but nobody could say what it would be allowed to touch in production or show afterwards what it had done.

How they set it up

The agent’s model calls and tool calls go through Raidu. Reading a claim file is allowed. Policyholder names, policy numbers and bank details are redacted before any text reaches the model. Sending a letter outside the company or issuing a payment waits for a named adjuster, and the action runs only after its signed record is written.

What compliance gets

A record of each tool call, each hold and each approval, with who approved it and the rule that applied, signed with the insurer’s own key. Adjusters keep the time the agent saves, and a person still decides what leaves the building.

Rollout

  1. Step 1The agent's model calls and tool calls routed through Raidu
  2. Step 2Approval rules for payments and outbound letters
  3. ThenCompliance reviews held and denied actions each month

What they deployed

  • Agent Firewall: each tool call checked against the insurer's policy
  • Approvals: payments and outbound letters held for a named adjuster
  • AI Firewall: policyholder data redacted before the model
  • Evidence and Audit: each step signed with the insurer's own key

Talk to us about your data and AI.

Tell us what you need to prove and to whom. We will show you the record it produces.

Talk to us