The first notice of loss, or FNOL, is an emotional conversation and a data-acquisition event. A good interaction reassures the claimant, captures the facts needed to begin, identifies urgency, and creates the right workflow. A poor one forces repetition, misses risk, and creates downstream rework.
A domain-specific voice agent can handle routine intake, but only if it is built around claims logic, evidence, and escalation rather than call containment.
What the market is doing
In its 2026 survey of 347 insurers, EIOPA found that claims management was already a generative-AI use area for 32% of respondents and a planned area for 59%. Respondents specifically reported using GenAI to extract insight from FNOL. Voicebots were among the described customer-service applications.
Commercial deployments provide more concrete, though vendor-reported, signals. Branch and Liberate report 43% customer adoption of voice or digital FNOL and a fall in average call duration from 12:23 to 7:10, a 42% reduction. The case projects 70% lower handling cost and says a homeowners workflow was ready eight weeks after the auto launch.
A separate Liberate property-carrier case reports 40% autonomous FNOL resolution, 50% shorter call duration, one-second average answer speed, 0% abandonment, and about 72,000 human handling minutes saved per month. The carrier is described as writing more than $800 million in direct premium but is not named on the page. These are supplier-published metrics and should be validated against call mix, transfer definitions, customer outcomes, and corrections.
Heritage Insurance and Hi Marley describe a conversational workflow that moves between voice and text and lets claimants attach photos and other media. This highlights an important design choice: FNOL is often multimodal, not voice-only.
The job of the voice agent
The agent should:
- identify the caller and policy without exposing information
- check immediate safety and injury
- identify loss type, time, and location
- collect involved parties and property
- capture a neutral account in the claimant's own words
- ask loss-specific follow-up questions
- accept photos, documents, and contact details
- read required notices and record acknowledgment
- explain next steps without promising coverage
- create and route the claim with a reviewable transcript
It should not decide coverage, liability, fraud, or settlement during the call.
The specialized model stack
The specialized model stack
Claims ASR
The speech model needs policy numbers, vehicle details, addresses, provider names, damage terms, weather, injuries, and local accents.
Dialogue policy
A deterministic state machine defines mandatory questions, branching by claim type, disclosures, and escalation. A language model interprets natural answers inside that policy.
Entity extraction
An SLM maps the conversation to the claims schema and returns field-level confidence. Critical fields are repeated back. It preserves uncertainty instead of guessing.
Knowledge retrieval
RAG retrieves approved procedural and service information for the exact product and jurisdiction. It should not generate a coverage position from general language.
TTS and empathy design
The voice must be clear, interruptible, and paced. Empathy is not simulated emotion.
Multimodal evidence
Send a secure link for images, video, receipts, or documents. A vision or document model can classify and extract them for adjuster review.
Claims ASR
The speech model needs policy numbers, vehicle details, addresses, provider names, damage terms, weather, injuries, and local accents. It should handle distressed speech, background noise, and poor mobile connections.
Dialogue policy
A deterministic state machine defines mandatory questions, branching by claim type, disclosures, and escalation. A language model interprets natural answers inside that policy.
Entity extraction
An SLM maps the conversation to the claims schema and returns field-level confidence. Critical fields are repeated back. It preserves uncertainty instead of guessing.
Knowledge retrieval
RAG retrieves approved procedural and service information for the exact product and jurisdiction. It should not generate a coverage position from general language.
TTS and empathy design
The voice must be clear, interruptible, and paced. Empathy is not simulated emotion. It is good service behavior: acknowledge, explain, avoid repetition, and connect the claimant to a person when needed.
Multimodal evidence
Send a secure link for images, video, receipts, or documents. A vision or document model can classify and extract them for adjuster review.
The mandatory escalation set
Transfer or create priority review for:
- injury, medical need, or immediate danger
- stranded or vulnerable claimant
- suspected total loss or major property loss
- fatality or serious third-party involvement
- coverage dispute or cancellation issue
- fraud or coercion indicator
- legal representation or litigation
- low identity confidence
- low confidence on critical facts
- customer request for a person
Test these with exact phrases, euphemisms, accents, and indirect descriptions. Escalation recall should be close to 100% on the defined high-risk test set.
The metric hierarchy
| Priority | Metric | Why |
|---|---|---|
| Safety | high-risk escalation recall | prevents missed urgent cases |
| Accuracy | critical-field accuracy | protects downstream decisions |
| Completeness | required-field completion | reduces adjuster callbacks |
| Experience | abandon, repeat, transfer satisfaction | reflects claimant effort |
| Workflow | claim ready for assignment | measures usable intake |
| Efficiency | human minutes and cost per completed FNOL | captures economic value |
| Quality | post-intake correction and reopen | finds hidden rework |
Do not optimize containment at the expense of these higher-order outcomes.
A pilot design
Start with a lower-severity, high-volume claim type and limited hours. Build a test set from 500 to 2,000 consented calls with policy-number variation, distressed speech, multiple languages, interruptions, and complex edge cases. Define the branch logic with claims, legal, compliance, and contact-center teams.
Run in agent-assist mode, then automate only the clean branches. Sample completed claims for field accuracy and downstream correction. Compare repeat-contact and complaint rates with the human baseline.
An example gate might require at least 98% accuracy on identity-linked policy selection after confirmation, 95% required-field completion, 100% recall on defined emergency tests, and at least 25% lower human minutes per completed FNOL.
The economics
For 100,000 annual FNOL calls at 12 minutes each, direct handle time is 20,000 hours. Reducing average human involvement by five minutes creates an 8,333-hour gross capacity pool. Add after-call work, fewer callbacks, faster assignment, and better evidence. Subtract telephony, inference, integration, review, and exception staffing.
Keep expected and realized savings separate. A projected handling-cost reduction is not booked value until staffing, overtime, vendor expense, or throughput changes.
The conclusion
The first five minutes set the quality of the claim. A voice agent earns its place by capturing accurate, complete evidence and accelerating human help, not by keeping the caller away from a person.
Task-specific ASR, constrained dialogue, claims extraction, multimodal intake, and explicit escalation make FNOL a strong and measurable bespoke-AI use case.
Research note
Research is current through September 5, 2026. Deployment outcomes are vendor-published unless stated otherwise. Local legal, regulatory, claims, language, and vulnerable-customer requirements must govern the design.
Continue the research
- Insurance AI solutions across the policy lifecycle
- Insurance policy language AI
- RAG vs fine-tuning vs bespoke AI models
- Why an enterprise evaluation set becomes an AI moat
- Banking voice AI agents
Building a Production-Ready System
Conscious Engines builds AI first notice of loss workflows that capture the claimant's words, extract critical facts, preserve media evidence, and route the case under policy and severity controls. The system is evaluated on completeness, triage quality, cycle time, and downstream rework rather than call containment alone.