Filing a first notice of loss is rarely straightforward. A homeowner is dealing with water damage, fire, storm, or theft while trying to recall dates, describe affected rooms, and explain what they have already done. The claims team needs to collect that information, compare it with the policy, identify urgent follow-up, and route it to the right person for review. In this walkthrough, we build Harbor Home, a voice-first FNOL demo that combines Twilio Agent Connect, OpenAI, and Box AI to create a real-time voice experience for homeowners to report a loss and a web-based claims dashboard for adjusters to process it. Here is how the five-stage workflow runs: Stage 1: Twilio Agent Connect collects the claim by voice. A homeowner calls a Twilio number and speaks with Harbor, an AI intake assistant. Harbor confirms safety first, then gathers one topic per turn: caller name, callback number, property address, loss date and time, what happened, affected areas, immediate hazards, mitigation already performed, and temporary lodging needs. No Social Security numbers, bank details, or payment card information are ever requested. Stage 2: OpenAI extracts reported facts. When the caller hangs up, OpenAI extracts the transcript into a typed FNOL object using a Zod schema. OpenAI does not receive the policy and is explicitly instructed not to assess coverage. Missing information stays missing. The model is never asked to invent facts. Stage 3: The application creates an intake report and metadata in Box. The FNOL report is uploaded to Box as a Markdown file with policy analysis marked as pending. Box metadata captures claim number, claimant details, loss type, severity, claim status, and analysis status. Box becomes both the content store and the queue state for the dashboard. Stage 4: Box AI compares the report with the homeowner's policy. Both the FNOL report and the synchronized homeowners policy already live in Box. The application passes both file IDs to the Box AI multiple item Q&A endpoint. Box AI returns a conservative preliminary assessment: likely covered, partially covered, needs review, or likely excluded, along with a rationale, specific policy references, the potential deductible, and recommended next steps. No separate policy retrieval pipeline is needed. Stage 5: The application versions the report, updates metadata, and creates a human review task. A new version of the Markdown file is uploaded with the enriched analysis. Metadata is patched with the final assessment. A Box review task is created and assigned to the adjuster. The AI organizes and surfaces the right information. The adjuster owns the coverage decision. The Next.js claims dashboard reads metadata directly from Box, providing a searchable queue, claim counts, a focused overview, and a link to open each file in Box. The transcript is only downloaded when an adjuster opens it, keeping the queue response small and sensitive claim detail protected until it is needed. No application database is required. The Markdown report is the durable claim record, Box metadata powers the queue, and Box file versioning preserves the full progression from raw intake to enriched report. A live no-auth demo is available at https://box-twilio-agent-demo.vercel.app The source code is available on GitHub: https://github.com/box-community/box-twilio-agent-demo FAQs: Q: What is a First Notice of Loss (FNOL)? A: FNOL is the first report made by a policyholder to their insurance company after a loss event such as water damage, fire, storm, or theft. It initiates the claims process and typically requires collecting contact information, property details, loss description, and any immediate risks or mitigation steps already taken. Q: What does Twilio Agent Connect do in this demo? A: Twilio Agent Connect handles the live phone conversation. It receives the inbound call and streams each caller utterance to the application, where OpenAI generates Harbor's spoken replies and manages the conversation flow. Q: What does OpenAI do in this demo? A: OpenAI has two jobs: producing Harbor's spoken replies during the call, and extracting the final transcript into a typed FNOL object using the Responses API with a Zod schema. OpenAI does not receive the policy and is explicitly instructed not to assess coverage. Q: What does Box AI do in this demo? A: Box AI compares the FNOL intake report with the homeowners policy using the multiple item Q&A endpoint. Because both files already live in Box, no separate policy retrieval pipeline is needed. Box AI returns a preliminary coverage assessment with a rationale, policy references, potential deductible, and recommended next steps.

 2219      8