Agentic Classification and Response for Insurance Query Queues on Amazon Bedrock

Industry: Insurance and Health Benefits Administration. Offering: Agent and Automation Development, Generative AI.

About Customer

Customer is an Indian health-benefits administrator operating as a third-party administrator for insurers, corporate clients, and the insured employees covered by those policies. Member queries arrive all day by email and messaging: claim status, coverage questions, network hospitals, e-cards, and document requests. Query volume grows with each new corporate account and each enrolment cycle.

Customer wanted to use agentic AI on that query flow so its service team spends less time researching answers and more time reviewing them.

Business Objectives

1. Classify and route every query on arrival. Apply intent, product, urgency, and routing category automatically as a message enters the queue, so no one has to triage by hand.

2. Give agents a draft to review. Agents review and approve an answer instead of researching and writing it.

3. Ground every response in live, member-specific context. A coverage answer depends on which member is asking, which policy applies, what the rule engine determines about eligibility, and what the claims record shows. Customer wanted responses assembled from those live sources, with citations back to the policy or SOP they draw from.

4. Handle more queries without adding staff at the same rate. Query volume rises with account growth and enrolment cycles, and the service team needed to absorb that growth.

5. Extend the same capability to WhatsApp. Customer wanted WhatsApp to use the same reasoning and retrieval path as email, with no separate build to maintain.

6. Keep member data in India with a complete audit trail. Insured-employee personal data was to remain within the country, with every classification, retrieval, and generated response logged.

Solution Delivered

1. AspireNXT built an agentic layer in front of the existing service queues, orchestrated by Amazon Bedrock AgentCore and grounded in the customer's own knowledge base, rule engine, and insurance database. The existing queues stayed in place.

2. An auto-response agent authenticates the member before it acts. The agent establishes member identity first, so it retrieves and acts on that member's data only. The check runs on every request, not once per session.

3. Retrieval combines three sources. The agent calls the rule engine for eligibility and policy decisions, reads member, policy, and claim context from the insurance database, and retrieves member-relevant content from an Amazon Bedrock Knowledge Base backed by an Amazon OpenSearch Serverless vector store.

4. Agents review every draft. Drafts return to the queue for review. A controlled auto-send path exists for high-confidence, low-risk intents, with the intent list and confidence threshold agreed with the customer.

5. The agentic layer is read-only to systems of record. It reads from the rule engine and insurance database and writes only to the ticket.

6. WhatsApp uses the same reasoning path. The WhatsApp channel is integrated through AWS End User Messaging Social and runs through the same classification, authentication, retrieval, and drafting pipeline as email.

7. Everything is logged. Every classification, retrieval, and generated response is written to a tamper-evident audit trail, and member and policy data stays in India.

Anthropic Claude on Amazon Bedrock

Claude classifies each query, decides which sources to check, and writes the draft reply. It runs on Amazon Bedrock in India, so insured-employee data stays in the country.

1. Classification on arrival. Claude reads each inbound message and assigns intent, product, urgency, and routing category, with a confidence score written back to the ticket alongside the labels. A single message can ask about claim status, coverage, and a document at once, so keyword rules are not enough.

2. Agentic tool use through Bedrock AgentCore. Claude decides which sources a query needs and in what order: the rule engine for an eligibility determination, the insurance database for policy and claims context, the knowledge base for approved content, or a combination. Claude picks the order for each query.

3. Knowing when it has enough. The agent keeps retrieving until it can answer or has nothing more to check. Claude's confidence in the assembled context decides whether a draft goes to an agent for review, routes to a person unanswered, or qualifies for the controlled auto-send path.

4. Grounded drafting with citations. Claude composes the reply from retrieved context only, and cites the policy clause or SOP each statement rests on. The reviewing agent checks a cited draft instead of researching the answer again.

5. Member-scoped by construction. Claude operates inside an authenticated member context. Tool calls are scoped to that member, so the model cannot retrieve another member's data, whatever a message asks for.

6. Routed to the model the task needs. Classification runs at high volume on short messages. Drafting reasons across rule-engine output, claims context, and approved policy content at once. Each request goes to the Claude model suited to that work.

7. Governed on every call. Amazon Bedrock Guardrails apply PII and content-safety controls to every request and response, and every classification, retrieval, and generated reply is written to the audit trail.

Services Used

• Anthropic Claude on Amazon Bedrock

• Amazon Bedrock AgentCore

• Amazon Bedrock Knowledge Bases and Bedrock Guardrails

• Amazon OpenSearch Serverless

• AWS Lambda

• Amazon EventBridge and Amazon SQS

• Amazon API Gateway

• AWS End User Messaging Social

• Amazon CloudWatch and AWS CloudTrail

What Was Delivered

1. Classification applied on arrival, with intent, product, urgency, routing category, and confidence written back to the ticket.

2. Agents reviewing instead of composing, working from a cited draft.

3. Responses grounded in live sources: the rule engine, the insurance database, and approved knowledge-base content, with citations attached.

4. Capacity to absorb query growth from new corporate accounts and enrolment cycles.

5. One reasoning pipeline for email and WhatsApp.

6. A complete audit trail, with member data kept in India.

Applying agentic AI to a high-volume service operation? Talk to the AspireNXT AI team.