How LandingAI made insurance documents conversational using Vocal Bridge

Ask a travel-insurance question out loud, hear an answer grounded in the policy, and watch the exact supporting lines light up on the page.

The problem: policy answers are buried in policy language

Insurance questions rarely arrive in the vocabulary a policy uses. A traveler says their bag is missing. The document says "baggage delay," "maximum benefit," and "proof of loss." Search only works when the customer already knows the phrase to look for, and reading a long policy from top to bottom is a poor fit for the moment something has gone wrong.

The harder problem is trust. A fluent answer is not enough when coverage limits and claim requirements matter. The customer needs to know which clause supports it, and an insurer needs a way to keep a conversational system from filling gaps with general knowledge. FinePrint treats the source document as the boundary. If the answer cannot resolve to policy text, the interface has nothing to highlight.

What LandingAI built and deployed with FinePrint

FinePrint turns a policy into a voice-accessible knowledge base built and deployed on LandingAI. LandingAI ADE first parses the uploaded document into structured Markdown, then extracts benefits and constraints into separate views. That gives the question-answering layer both readable policy text and typed fields it can reason over.

The user asks a natural-language question through Vocal Bridge. FinePrint returns a grounded answer with coverage limits and next steps drawn from the policy, automatically scrolls the document, and highlights the precise blocks that support the response. The result is more useful than a spoken summary because the evidence stays visible beside the answer. The public sample project includes a travel policy so builders can try the complete loop.

How FinePrint uses Vocal Bridge

FinePrint uses Vocal Bridge's Voice for your app surface. Its public React implementation wraps the experience in VocalBridgeProvider, receives delegated policy questions through useAIAgent, and watches the live conversation with useTranscript. The public /app/api/voice-token route reads VOCAL_BRIDGE_API_KEY on the server and exchanges it for a room token, so the browser never receives the long-lived credential.

When a question arrives, the app sends the policy ID and conversation history to its grounded hotline endpoint. That endpoint returns a spoken reply plus an assessment containing citations. FinePrint releases the reply through Vocal Bridge and reveals the matching line boxes at the same moment, so the page does not appear to answer before the voice does.

import { useEffect } from "react";
import { useAIAgent } from "@vocalbridgeai/react";

export function PolicyVoice({ policyId }: { policyId: string }) {
  const { pendingQuery, respond } = useAIAgent();

  useEffect(() => {
    if (!pendingQuery) return;
    void fetch("/api/hotline/turn", {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({
        policyId,
        messages: [{ role: "user", content: pendingQuery.query }],
      }),
    })
      .then((res) => res.json())
      .then(({ reply }) => respond(pendingQuery.turnId, reply))
      .catch(() =>
        respond(pendingQuery.turnId, "Sorry, I couldn't reach the policy just now.")
      );
  }, [pendingQuery, policyId, respond]);

  return null;
}

Illustrative. FinePrint's public implementation also holds slow answers for a clean conversational turn and synchronizes citation boxes with the spoken response.

The architectural split is explicit. LandingAI DPT-3 Parse and Extract read the document and preserve line-level grounding. FinePrint's span bridge maps extracted fields back to page coordinates, and Claude produces the clause-level reply from a fixed set of references. Vocal Bridge owns the real-time WebRTC voice channel, delegated AI-agent turn, transcript stream, and spoken delivery.

What's hard about voice here

• The answer and the evidence have to arrive together. If the app highlights one clause while the voice is still speaking about another, the grounding promise breaks. FinePrint ties citation display to the same delegated turn it releases through Vocal Bridge.
• Document reasoning can outlast a conversational pause. The public implementation gives a completed lookup a short grace period when a follow-up may still be in progress, then releases it on a clean turn. Vocal Bridge's transcript stream and AI-agent delegation provide the state needed to avoid talking over the user.
• A fallback cannot improvise coverage. FinePrint configures its Vocal Bridge agent to speak delegated replies verbatim and limits the base prompt to concierge behavior. If document reasoning fails, the app asks the caller to repeat instead of inventing a policy answer.