We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    [Case Study] Zero-Hold Claims Intake: Branch's Voice AI FNOL Deployment

    How an insurer launched 24/7 homeowners claim intake in eight weeks, reached 43% adoption, and cut average call duration by 42%.

    Conscious Engines

    Industry: Insurance
    Organization: Branch
    Use case: Voice and digital first notice of loss for homeowners claims
    Evidence basis: Liberate vendor case using Branch-reported results
    Disclosure: This is an independent analysis by Conscious Engines. The source is vendor published, and the cost-reduction figure described below was an expectation, not a realized result.

    1. Outcome at a Glance

    Branch deployed an AI-enabled homeowners claim-intake flow in eight weeks. The service operated 24 hours a day with zero hold time and reached 43% adoption among relevant users.

    Key Outcomes

    8 weeks

    Time to launch

    Implementation speed reported in the cited case.

    43%

    AI channel adoption

    Reported use.

    7 minutes 10 seconds

    Average AI call

    Observed call duration.

    12 minutes 23 seconds

    Prior outsourced-human call

    Comparison baseline reported in the cited case.

    42%

    Call-duration reduction

    Derived from reported averages.

    MeasureReported resultEvidence note
    Time to launch8 weeksImplementation speed
    AI channel adoption43%Reported use
    Availability24/7, zero holdService design
    Average AI call7 minutes 10 secondsObserved call duration
    Prior outsourced-human call12 minutes 23 secondsComparison baseline
    Call-duration reduction42%Derived from reported averages
    Cost reduction70%Expected future reduction, not achieved result

    The call-duration comparison represents a reduction of 5 minutes 13 seconds. It is meaningful because first notice of loss is a structured but emotionally sensitive workflow. The claimant needs to describe the event, provide policy and property details, understand next steps, and receive confirmation.

    2. The Operational Problem

    Claims do not arrive only during business hours. A homeowner may call immediately after water, fire, weather, or theft damage. Long hold times increase anxiety and delay mitigation. An outsourced intake service adds capacity, but the handoff can produce incomplete data, inconsistent scripts, and higher cost.

    FNOL is well suited to voice automation because the process has required fields and known branches. It is also risky because callers may be distressed, emergency situations require escalation, and a mistaken policy or loss detail can affect the claim journey.

    The voice agent must do more than transcribe. It needs to identify the policy, collect incident facts, detect urgent safety or mitigation needs, ask follow-up questions, preserve the caller's language, and create a structured claim record. It must transfer to a person when confidence or policy requires it.

    3. What Was Built

    The solution combined voice and digital intake with integrations into policy, claims, fraud, and vendor systems.

    System at a Glance

    Telephony and ASR

    Capture the caller reliably and in real time.

    Dialogue engine

    Follow the correct FNOL path and recover from interruptions.

    Structured extraction

    Populate incident, property, damage, and contact fields.

    System connectors

    Verify policy and create the claim.

    Risk checks

    Trigger fraud, emergency, or specialist pathways.

    Customer confirmation

    Restate critical facts and explain next steps.

    LayerFunction
    Telephony and ASRCapture the caller reliably and in real time
    Dialogue engineFollow the correct FNOL path and recover from interruptions
    Structured extractionPopulate incident, property, damage, and contact fields
    System connectorsVerify policy and create the claim
    Risk checksTrigger fraud, emergency, or specialist pathways
    Customer confirmationRestate critical facts and explain next steps
    Human handoffTransfer with transcript and collected context

    The narrow domain makes task-specific models attractive. A specialized intent and entity model can capture claim facts at lower cost. A larger model can handle unusual descriptions. Deterministic validators can enforce date, address, policy, and required-field logic.

    Every critical fact should be confirmed. A fluent conversation is not proof that the structured claim record is correct.

    4. How It Reached Production

    The eight-week timeline suggests a deliberately bounded first release focused on homeowners FNOL.

    Start with one product and journey. Policy fields, scripts, emergency rules, and integrations become testable.

    Design for emotion and interruption. Claimants pause, correct themselves, and provide information out of order. The agent must preserve state and allow a person to take over.

    Test critical entities. Addresses, dates, phone numbers, policy identifiers, damage categories, and monetary values need field-level accuracy measures.

    Measure completion after the call. A shorter call is valuable only if adjusters receive complete, usable information. Track missing-field callbacks, claim reclassification, transfer, abandonment, and reopen rates.

    Keep forecast and outcome separate. The vendor case anticipated 70% cost reduction. Until supported by actual operating data, it should remain a forecast in the business case.

    5. What Insurance Leaders Should Take Away

    Branch's deployment shows that voice AI can become the front door to a real insurance workflow, not merely answer FAQs. The combination of 43% adoption, zero hold, and a 42% shorter average call supports further evaluation.

    A production-ready FNOL agent combines insurance-tuned speech recognition, multilingual dialogue, policy-grounded retrieval, structured claims integration, emergency escalation, and continuous call evaluation. The enterprise should own the schema, call policy, and test corpus.

    The primary metric is cost per complete and correctly routed FNOL, with customer satisfaction, critical-field accuracy, transfer, and downstream rework as controls. This keeps the focus on accepted claims data instead of conversation volume.

    Sources