A policyholder rear-ends another car at 11:14 p.m. on a rainy Tuesday. By 11:23 she is sitting on the shoulder, hazards on, calling her carrier. Her hands are shaking. The claims line answers on the second ring. There is no hold music, because the carrier's after-hours FNOL queue is staffed by an AI voice agent.
In the next four minutes the agent confirms her policy number, captures the date and time of loss, geolocates the intersection, walks her through which other vehicles were involved, asks whether anyone is hurt, asks whether police are on scene, captures the responding officer's name and report number, asks whether the airbags deployed, and asks her to take six photos when she feels safe to do so. It opens a claim file in the carrier's claim management system, assigns a claim number, and texts her the number along with a tow-vendor link.
What the agent does not do: tell her whether the claim is covered, decide whether to dispatch a rental, decide whether the totals look suspicious, or offer a settlement. That waits for a human adjuster who will see the full file in the morning queue, sorted by severity.
This is what FNOL automation looks like when it is drawn correctly. Most of it is not.
FNOL is the most predictive call in the claim lifecycle
First Notice of Loss is the first contact a claimant has with the carrier or TPA after a covered event. Auto, property, general liability, workers' comp. It is also the call that disproportionately drives total claim cost and customer retention.
The first 24 to 48 hours after a loss are the window where evidence is freshest, where rental and lodging decisions get made, where preferred-vendor networks are mobilized, and where the claimant decides whether they trust the carrier enough to stop shopping for a lawyer. Carrier CSAT studies routinely find that the FNOL experience accounts for an outsized share of overall claim NPS, often more than the settlement itself. JD Power's annual claims studies have made this point in different forms for two decades.
Speed matters, but completeness matters more. An FNOL call that captures the right structured fields the first time prevents three callbacks, two duplicate claim files, and one adjuster reopening the file because the loss-location ZIP is wrong.
The FNOL schema is well-defined and that is exactly why voice agents fit
Every claim management system, whether Guidewire ClaimCenter, Duck Creek Claims, or Insurity ClaimsXPress, is organized around a similar core schema. The fields differ in name and granularity, but the shape is consistent.
A typical auto FNOL captures:
- Policy number and named insured verification
- Date, time, and time zone of loss
- Loss location: address or lat/long, not just a city
- Description of loss in claimant's own words
- Vehicles involved: year, make, model, VIN if available, drivability
- Other parties involved: name, phone, address, insurance carrier and policy number
- Injuries: yes or no, who, severity at the time of the call, ambulance dispatched
- Property damage to non-vehicle property
- Police report: agency, officer, report number, citation issued to whom
- Witnesses: name and phone
- Photos: whether available and how the carrier should receive them
- Tow status and current vehicle location
- Rental need
- Preferred contact method and best times
Property and liability FNOLs swap vehicle fields for premises and damage-extent fields, but the structure is identical: who, what, when, where, how, what's the evidence, who else is involved.
The schema being structured is the reason AI voice agents fit FNOL better than they fit, say, complex coverage disputes. The agent is filling a known form. It is not deciding anything. The benefit compounds at 11 p.m. on a Tuesday when losses do not respect business hours.
Why the human handoff still matters
A well-built FNOL agent finishes the structured intake in three to six minutes. The handoff to a human starts when the file leaves capture mode and enters assessment mode.
Four moments where the human takes over and should not be automated away:
- Severe-injury or fatality losses. When the caller mentions someone unconscious, severe bleeding, or a fatality, the agent's job is to keep the call short, capture the absolute minimum required to dispatch resources, and warm-transfer to a human or 911 depending on the carrier's protocol. This is not the moment to read a fifteen-field schema.
- Coverage adjudication. Whether a loss is covered is an adjuster decision. The agent captures facts. It does not say "yes, that's covered" or "no, your deductible is X." Carriers that let agents make coverage statements are setting up E&O exposure.
- Fraud-suspicion judgment. The agent flags. The human decides. More on this below.
- Immediate-pay or fast-pay decisions on small claims. Some carriers have express-pay paths for small auto-glass or property claims. The decision to pay is human, even when the path looks automatable. The reputational cost of wrongly auto-denying is higher than the labor savings.
The shape of the right handoff is not "AI does the easy ones, humans do the hard ones." The shape is "AI captures every call cleanly, and humans do the assessment work on every claim that needs assessment." Volume of files goes up. Time per file for the adjuster goes down. Rework goes down. That is the math that justifies the spend.
For carriers running both inbound FNOL and outbound follow-up calls, the warm vs blind transfer question decides whether the adjuster gets context on the live transfer or has to start from scratch.
The fraud-flag question is where most vendors overclaim
Voice AI vendors love to talk about fraud detection. The honest version is narrower.
What the agent can do during FNOL:
- Capture inconsistencies inside the call. The claimant says "no police were called" at minute two and at minute six mentions "when the officer asked for my license." That contradiction is captured verbatim and surfaced to the SIU queue.
- Capture timing red flags. Loss reported on day 27 of a 30-day-old policy. Loss reported within hours of a coverage increase. Loss reported in a state the policy doesn't typically operate in.
- Capture script-coaching signals. Long pauses before factual questions, requests to repeat questions verbatim, background voices feeding answers, identical phrasing to known fraud-ring scripts.
- Capture network signals. Same phone number, same VIN, same intersection, same body shop appearing across unrelated claims.
What the agent cannot do, and should not be sold as doing:
- Decide that a claim is fraudulent. SIU does that.
- Confront the claimant. Confrontation tips off the bad actor and burns the investigation.
- Deny the claim. That is an adjuster decision with a paper trail.
The agent's role in fraud is identical to a smoke detector. It detects. It alerts. It does not put out fires. Vendors who pitch "AI fraud detection" as a closed-loop product are either misrepresenting the workflow or building in legal exposure the carrier will inherit.
Carrier reality check on AI in FNOL
The major US carriers have been investing in FNOL automation since roughly 2020, with significant acceleration in 2022 and 2023 as LLM-based intake became viable. None of them run fully automated FNOL. All of them run AI-assisted FNOL in some form.
- Geico has invested heavily in voice and chatbot front-ends for auto FNOL, with structured handoff to adjusters.
- Progressive's photo-estimate path and digital FNOL flows have been multi-year priorities.
- State Farm has piloted AI-assisted intake across regions, with measured rollout.
- Liberty Mutual, Allstate, and Farmers each run varying degrees of AI in FNOL, often heavier on the digital side than voice.
- Lemonade is the cleanest end-to-end reference for AI-first claims in personal lines. It is also a smaller and simpler book of business than the legacy carriers, which is exactly why their architecture works.
The pattern across all of them: AI handles intake, structured capture, and routing. Humans handle assessment, payment, and disputes. None of them have removed the adjuster.
Integration depth is the question carrier ops teams actually ask
A voice agent that captures perfect FNOL data and emails a PDF to the claims inbox is worse than useless. It creates rekey work and a second source of truth.
The integration that matters writes structured fields directly into the claim management system. For a Guidewire ClaimCenter shop, that means Cloud API integration into the FNOL endpoint with the carrier's required validation rules. For a Duck Creek shop, it means OnDemand integration. For Insurity, it means their published API surface. For older mainframe-backed admin systems, it means a middleware layer or, in the worst case, RPA.
Carriers running on a major modern platform should ask the vendor: which version, what objects, which endpoints, what's the conflict-resolution behavior when the agent and a human update the same file. Carriers on legacy systems should ask: are you proposing a real integration or a glorified email-to-PDF pipeline. The CRM integration models post lays out the four common patterns and where each one breaks.
Bonus question: what happens to the call recording, the transcript, and the structured fields when the claim closes seven months later. State retention rules vary, and the answer should be in the contract, not in support email.
Five vendor-evaluation questions for FNOL automation
- Show me a live FNOL call where the claimant mentions an injury at minute three. Watch what the agent does. Does it abandon the schema and prioritize dispatch, or does it keep reading the form?
- Which fields write directly into our claim system, and which arrive as attachments? Anything that arrives as an attachment will get rekeyed. Get the answer in writing.
- What is your SIU-flag taxonomy and how does it surface to our investigators? If the answer is "we email a CSV," that is not an SIU integration.
- What is the latency between call end and claim file creation? Three minutes is normal. Three hours is broken. Three seconds is a red flag, because it means no human review on the routing decision.
- What's the failure mode when the agent doesn't understand the caller? Does it loop, escalate, drop, or transfer to a human? Watch this happen on a real call before signing.
The wrong reason to buy FNOL automation is to reduce headcount. The right reason is to make sure no claim ever sits in voicemail while a claimant decides whether to call a plaintiff's firm. Carriers that frame the spend as "we capture every claim, every hour, complete the first time" come out ahead. Carriers that frame it as "we replaced the night shift" find out in twelve months that the night shift was doing more than they thought.
Related: the failure modes that show up in the first 30 days, and open-enrollment capacity planning for carriers that face seasonal call spikes.
If you are evaluating FNOL voice automation for a carrier, TPA, or agency book, book a call with our team, start free trial, or check the pricing page.