Benefits and claims assistant

The challenge

A health-insurance provider had a single front door for member questions: the call center. Whether a member wanted to know if a procedure was covered, why a claim was denied, how much of their deductible remained, or which tier a medication fell into, the answer lived behind a phone queue. Most of these questions had answers that already existed in the member’s plan documents or claims history. The bottleneck was not missing information; it was that the only way to get it was to wait for a human agent to look it up.

That model is expensive for the insurer and frustrating for members. It also fails in a specific way: benefits questions are personal. “Is physical therapy covered?” has no single answer, because it depends on the member’s plan, their network, and where they sit against their accumulators. A generic FAQ cannot answer it, which is exactly why everything routed to the phone.

What we built

We built a conversational assistant that answers benefits and claims questions directly in the member’s authenticated web and mobile experience, grounded in that member’s own plan documents and claims data.

Conversation layer

A dialogue orchestrator handles intent classification and multi-turn state, so members can ask follow-up questions the way they would on a phone call. “What’s my deductible?” followed by “and how much have I used?” resolves correctly because the orchestrator carries context across turns rather than treating each message as a fresh query.

Identity-scoped retrieval

This is the part that makes insurance hard. Every retrieval is filtered by member ID and plan ID before any search runs. The plan-document index covers Summaries of Benefits and Coverage, Evidence of Coverage documents, and formularies, and the assistant only ever searches the documents that govern this member’s plan. Claims lookups go through the claims system API, scoped the same way, returning claim status, EOB details, and accumulator balances for that member alone. A member can never see another member’s data, and just as importantly, can never get an answer from another plan’s documents. A correct-sounding answer drawn from the wrong plan is a wrong answer, and in insurance it is a costly one.

Grounding is the second hard part. Plan language is precise, and a plausible paraphrase that drifts from it can mislead a member about coverage. Every answer cites the retrieved evidence it is based on, drawn from semantic search over the plan index or scoped lookups against the claims API. If the system cannot ground an answer in the member’s documents or claims data, it does not guess.

Escalation with context

Not every conversation should be self-service, and the design treats escalation as a feature rather than a failure. When a question falls outside what the assistant can answer with confidence, the conversation hands off to the call center with the full transcript and the retrieved evidence attached. The agent picks up mid-thread instead of starting from zero, so the member never repeats themselves. Roughly 10 percent of conversations take this path; the other 90 percent resolve without a human.

How it was delivered

We delivered the system end to end: the dialogue orchestration, the identity-scoping layer, the plan-document indexing pipeline, the claims API integration, and the escalation handoff into the existing call-center workflow. The work ran as a focused engagement with the insurer’s team, starting with the retrieval and scoping foundations, then layering the conversational experience on top once grounded answers were reliable, and finally wiring the escalation path so agents received context rather than cold transfers.

What shipped

• Member-facing conversational assistant on web and mobile, behind existing authentication
• Dialogue orchestrator with intent handling and multi-turn state
• Identity-scoping layer filtering every retrieval by member ID and plan ID
• Grounded retrieval over plan documents (SBC, EOC, formularies) with evidence cited per answer
• Claims system integration for status, EOB, and accumulator queries
• Call-center escalation with full transcript and retrieved evidence attached

The measured outcomes: 90 percent of member conversations resolve self-service, and inbound call-center volume dropped 80 percent. The call center now spends its time on the questions that genuinely need a human.