An AI agent automates first notice of loss by taking the initial report on whatever channel the policyholder reaches for, capturing every field an adjuster needs, verifying the policy against your admin system, collecting evidence and routing the file with the full record attached, in one conversation. The stakes are measurable. J.D. Power's 2025 U.S. Property Claims Satisfaction Study scores satisfaction at 762 out of 1,000 for claims completed within 10 days and 595 when repairs run past 31 days, a 167-point gap. Much of that gap is decided in the first minutes, when the record is either complete or full of holes that become callbacks.
Key takeaways
A complete FNOL record has around 20 fields in five groups: reporter, policy, event, parties and property, evidence. If a required field is empty when the file reaches the adjuster, the intake failed.
Voice comes first. J.D. Power's 2026 property study found only 38% of homeowners reported first notice of loss digitally, so the phone still carries most FNOL volume, and chat and email must inherit the same record.
Speed compounds: Deloitte reports digital FNOL with photo evidence cut time to payment by up to 5.5 days, and the average property claim still takes 40.7 days from notice to final payment.
Decision rule: the agent captures facts and runs your rules; a licensed adjuster makes every coverage, liability and payment call. Injury, fraud signals and any coverage question route to a human on hard rules.
Regulators expect the record: 24 states and the District of Columbia have adopted the NAIC AI model bulletin as of August 2026, and the NAIC claims-practices model regulation expects acknowledgement within 15 days.
What does a complete FNOL record contain?
A complete first notice of loss record holds enough verified facts for an adjuster to open the claim, set an initial reserve band and assign it without calling the policyholder back. That is the test. Every missing field becomes a callback, and callbacks are where cycle time hides.
The five field groups
Reporter and contact: who is reporting (insured, third party, agent, tow operator), relationship to the policy, callback number, preferred channel and language.
Policy: policy number or a lookup by name plus date of birth or address; status on the date of loss; named insureds; the vehicle or risk address on the schedule.
Event: date, time and location of loss, cause in the policyholder's words and in your loss coding, whether the loss is ongoing (water still running), police or fire report number.
Parties and property: other drivers and vehicles, witnesses, injuries reported, damaged items or rooms, whether the vehicle is drivable or the home habitable.
Evidence and next steps: photos or video received, documents promised, the claim number issued, and what the policyholder was told will happen next and by when.
Define "complete" per line of business in configuration, not in the model's head. Where the policyholder does not know a value, the record marks it unknown with a follow-up task rather than leaving it blank, because blank and unknown look identical in a queue and only one is actionable. The timestamped transcript, with every system lookup and every declined answer, belongs in the record too; it is what you produce when an examiner asks how a claim was opened.
How does AI take first notice of loss across voice, chat and email?
The agent takes FNOL on the phone first, because that is where a distressed policyholder goes, then continues the same claim on chat, SMS or email without asking for the story again. In J.D. Power's 2026 property claims study, 38% of customers reported first notice of loss through digital tools, 49% submitted photos digitally and 45% received updates digitally: most first reports still arrive by voice while the follow-up moves to text. The 2025 study found 82% of customers interacting through non-preferred channels, so the claim will cross channels whatever you optimise.
Voice: the first report
The caller says what happened in their own words. The agent checks for injury or an active emergency first, then works through the five field groups conversationally. It finds the policy with a live lookup and confirms administrative facts only: the policy was active on the date of loss and the caller is a named insured. Latency matters, because a stressed caller fills silence; sub-second responses and clean handling of interruptions are the bar for the voice channel. Mid-call, the agent texts a secure upload link and confirms receipt of the photos before hanging up.
Chat, SMS and email: the same claim, more complete
Two hours later the policyholder emails the police report number. The agent recognises the claim, attaches the number and confirms. Next morning they open chat to ask what happens next and get an answer from the same record. The workflow was written once, for voice, and deployed to the other channels without a rebuild, which makes the promise literal: the customer never repeats the accident, and the claim file just gets more complete. Because the agent holds the record, it can also send a consented SMS or email when an adjuster is assigned or a document is still missing, which matters: J.D. Power's 2025 Claims Digital Experience Study found insurers deliver adequate digital updates just 22% of the time.
How does the agent coordinate third parties while the policyholder is still on the call?
A capable FNOL agent dispatches sub-agents to book the tow, alert the mitigation vendor or page the on-call adjuster while the caller is still on the line, then reports the result inside the same conversation. The primary agent stays with the policyholder. A second agent contacts the tow provider with the location, vehicle and drivability status, gets an arrival window back, and the primary agent relays it. For a burst pipe, the same pattern reaches the mitigation vendor.
Three constraints keep this safe. Least privilege: the sub-agent shares only the fields the third party needs. Authority limits: it can book services inside pre-approved limits your workflow sets and must ask a human before committing beyond them. Logging: every outbound call or message lands in the claim file with a timestamp, so the adjuster sees the tow was booked at 9:14 pm and by whom. The team-of-agents pattern describes the architecture; FNOL is one of its clearest uses.
How much does FNOL speed change claims cost and satisfaction?
FNOL speed shows up in three documented numbers: satisfaction, retention and time to payment. Public cost figures specific to FNOL are thin, so treat days between notice and payment as the cost proxy: rental cars, alternative accommodation and follow-up calls.
Measure | Figure | Source |
|---|---|---|
Satisfaction, claim completed within 10 days vs repairs over 31 days | 762 vs 595 on a 1,000-point scale | |
Average time from first notice of loss to final payment | More than 44 days (2025), 40.7 days (2026) | |
Average repair cycle time, property | 29.6 days, down 2.8 days year on year | J.D. Power, 2026 property claims |
Average repairable cycle time, auto | 19.3 days, down from 22.3 days | |
Time to payment reduction from digital FNOL with photo evidence | Up to 5.5 days | |
Attrition risk, poor or just OK digital claims experience vs excellent | 52% vs 4% |
Between the 2025 and 2026 property studies, time to final payment fell 3.4 days and overall satisfaction rose 20 points to 702, with J.D. Power crediting faster repair and payment cycles. None of these studies isolates the FNOL step, so do not quote them as AI results; what they establish is that days matter, and FNOL is the step where the insurer controls the clock. The longer view is from McKinsey's Insurance 2030 analysis: claims processing remains a primary carrier function in 2030, but more than half of claims activities have been replaced by automation. First notice is the most automatable of them because it precedes any judgment.
Which rules govern an AI agent taking FNOL in the US?
Two NAIC instruments do most of the work: the unfair claims settlement practices model that states have built their timing rules on, and the AI model bulletin now adopted by roughly half the states. Neither prohibits automated intake. Both assume you can show the record.
The NAIC Unfair Property/Casualty Claims Settlement Practices Model Regulation sets the timing floor. Section 6 requires an insurer to acknowledge receipt of a claim notification within 15 days unless payment is made in that period, and adds that a non-written acknowledgement needs a dated notation in the claim file. An agent that issues the claim number on the call and sends a written confirmation before hanging up puts a dated acknowledgement in the file on day one.
The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted 4 December 2023, asks for a written AI systems program covering governance, risk controls, testing and third-party oversight. The NAIC's adoption map dated 6 August 2026 lists 24 states plus the District of Columbia as adopters, with California, Colorado, New York and Texas carrying their own AI guidance. Section 4.1 is the clause that matters for a vendor-built FNOL agent: it expects due diligence on third-party AI so that decisions it makes or supports "will meet the legal standards imposed on the Insurer itself." Section 4.2 asks for audit rights and regulator cooperation in vendor contracts. Be ready to hand an examiner pre-launch test results, transcripts and QA scoring for every FNOL conversation, not a sample.
Scope keeps this simple. In this design, the agent never forms an opinion about whether the claim is covered. It gathers what the coverage rules need and runs them, and a licensed adjuster owns the coverage, liability and payment decisions. The issues that should never be automated covers the general principle; for FNOL it reduces to that one sentence.
What this looks like in practice: a worked example
A policyholder calls at 9 pm after a parking-lot collision. On Lorikeet's voice channel the agent confirms nobody is hurt, then captures the record: time, location, the other vehicle and plate, the other driver's insurer (unknown, flagged for follow-up), and the police report number. It verifies the caller against the policy system and confirms the policy is active and the caller is a named insured. It texts an upload link; four photos arrive while the caller is still talking. The car is not drivable, so a sub-agent contacts the network tow provider and returns a 40-minute window. The caller asks whether the damage is covered. The agent notes the question, explains an adjuster will review coverage and the deductible, and issues the claim number. Before the call ends it sends an SMS with the claim number, the tow window and next steps. At 7 am the caller emails the police report number, and the email agent attaches it to the same claim. Change one fact, a mention of neck pain, and the injury rule fires: the agent stops the intake, captures only what routing needs and pages the on-call adjuster with the record so far.
Under the hood this is a deterministic structured workflow for the field groups, identity check and escalation triggers, with natural-language handling for the loss narrative, combined in one interaction. Before launch it runs through adversarial simulations in which test callers push for a coverage answer or describe injuries mid-story, and every live conversation is scored by Coach for whether a boundary was approached. Lorikeet charges per resolution; an FNOL that escalates to an adjuster is not charged. The trust page covers SOC 2, PII redaction, role-based access and US, UK or Australian data residency, and the insurance page covers the claims lifecycle beyond FNOL.
The published proof is from adjacent deployments, not a carrier. Wonderschool uses Lorikeet voice to answer 100% of parent calls, up from around 10%, and absorbed the roughly half of inbound calls that were scam noise: evidence the voice layer holds under real volume. Carmoola, an FCA-regulated UK car finance provider, resolves 60% of inbound conversations end to end: evidence the guardrails hold in a regulated setting. Neither is an FNOL deployment. The voice agents for FNOL intake roundup compares vendors on this problem, and you can bring your hardest scenarios to a demo and watch the injury and coverage rules fire.
The honest limitation is integration depth. The one-conversation flow assumes your policy administration and claims systems expose lookups and claim creation through an API. Where policy lookup only exists as a batch file or a terminal session, the agent collects identity details for a human to verify later, and the record arrives complete but unverified. Catastrophes expose a second limit: Lorikeet scales to the call surge, but tow and mitigation vendors do not, so mid-call coordination during a hailstorm often returns no availability, and the agent has to say so.
What still needs a human
The list is short and should be enforced with hard rules rather than model judgment, because these are the cases where over-eager automation does the most damage.
Coverage, liability, reserve and payment decisions. A licensed adjuster owns these; the agent records the question and routes it.
Injury, fatality or an active emergency. Emergency-services guidance where relevant, the minimum needed to route, immediate handoff.
Fraud signals. Inconsistent timelines, a policy change shortly before the loss or a duplicate claim route to special investigations. The agent flags; it does not accuse.
Distress, vulnerability, complaints and requests for a person. Hand off without friction, record attached.
Disputed facts, litigation or regulator language. Route with the file flagged.
Catastrophe triage. Humans set priorities during a surge; the agent applies them at volume.
McKinsey expects the most complex claims in 2030 to still be handled by humans bringing empathy and expert judgment. FNOL automation does not change that. It changes what the human receives: a complete, verified record with the callbacks already made, instead of a voicemail and a half-filled form. The claims intake tools comparison covers what happens to that record after first notice.
Start with one line of business and the low-severity, no-injury band, write the field list and escalation rules before the happy path, simulate the callers who push for a coverage answer, and widen scope as record quality holds. The goal is a faster, cleaner first notice, not an AI adjuster.
