The first notice of loss process: what AI should handle at intake, and what it shouldn't

The first notice of loss process does two jobs, and AI should own only one of them outright. The first job is building the claim record: receiving the notice wherever it lands, extracting the facts, matching the policy in force, creating the claim and dating every document. The second is making the first judgments about the claim, and that is where this guide draws the line: a severity flag with written reasons belongs at intake, while the coverage position, any late-notice position, the reserve and anything done with a fraud signal stay with a person.

A large carrier has already written that split down. In a May 12, 2026 newsroom post on its claims transformation, State Farm said it is piloting an AI-powered virtual assistant to streamline initial loss reporting and gather the information needed to help, and added: "Claims professionals are making coverage and payment decisions supported by tools that streamline intake, triage, and documentation." That is a sound design rule for any carrier, TPA or MGA, and the reasons are mostly legal. The acknowledgment deadline starts when the notice arrives, sometimes in an insurance agent's inbox, and the rules on late notice and fraud expect a specific reason for what the insurer does.

Below: the process step by step, the acknowledgment clocks, the claim record at creation, the evidence for flagging severity, the calls AI should leave alone, and what happens after intake. For the term itself, see the definition of first notice of loss. Intake is the first of eight stages in our guide to automating the claims lifecycle.

Most explanations describe one event: someone reports a loss. Inside a claims operation it is seven jobs, and they split cleanly: five build the record and carry deadlines, two are judgments. Treat them as one job and you either automate a judgment or key by hand a record a machine should build.

The notice arrives everywhere at once. J.D. Power's 2026 U.S. Property Claims Satisfaction Study, based on 5,093 homeowners who filed a claim, found 38% of customers reported first notice of loss digitally, which leaves 62% (derived) reporting by phone, through an agent or some other way. Paper has not left commercial lines either. Markel's small business claims forms page lists the ACORD 1 property loss notice, ACORD 2 automobile loss notice, ACORD 3 general liability notice of occurrence and ACORD 4 workers' compensation first report of injury or illness, and tells the insured: "When complete, fax them to us." Add broker emails, client-program spreadsheets, portal submissions and voice calls, and step one is noticing that a notice has arrived at all.

The regulatory definition is wider than most intake setups. NAIC Model Regulation 902 defines a notification of claim as any notification, in writing or by other means acceptable under the policy, to an insurer or its agent, "which reasonably apprises the insurer of the facts pertinent to a claim." A phone call that names the insured, the date and what happened is a claim notice whether or not anyone has opened a file. Guidewire's ClaimCenter documentation calls FNOL "the event in which the insurer is informed of a potentially covered loss," and the table follows that event to the point where a person takes over.

Steps one to four are where intake time goes, and much of it is re-keying. For a TPA claims operations leader, every client program sends notices in its own format to its own inbox, and the first job on many files is typing. That is the work Hesper's FNOL intake agents take on: capturing the notice from portals, email, phone and agent or broker submissions, extracting the facts from ACORD loss notices, emails, attachments and photos, and creating the claim in the core system with a record of where each fact came from.

Section 6.A of the NAIC model is three sentences long, and all three matter at intake. An insurer that receives notification of a claim must acknowledge it within 15 days unless it pays within that time, and Section 3.D defines days as calendar days. An acknowledgment made other than in writing must be noted in the claim file and dated. And "notification given to an agent of an insurer shall be notification to the insurer." Read the first and third sentences together and the clock starts when the insurer's agent receives the email, not when someone keys the claim.

California goes further than the model on the agent point. 10 CCR 2695.5(e) requires the insurer, immediately and in no event more than 15 calendar days after notice, to acknowledge the claim, provide forms, instructions and reasonable assistance, and begin any necessary investigation, and it imputes an agent's or claims agent's failure to pass a notice on promptly to the insurer. New York's 11 NYCRR 216.4(a) counts 15 business days and treats notice to an agent as notice to the insurer, unless the agent tells the claimant it is not authorized to receive notices of claims. Florida's rule is the shortest and the narrowest: Section 627.70131(1)(a) gives 7 calendar days to review and acknowledge a claim communication, and it applies to residential property insurers only.

Washington's rule changes its wording on October 18, 2026: under WSR 26-17-089, WAC 284-30-360 reads "business days" where it said "working days," and a new preamble defines violations as an unfair method of competition and an unfair or deceptive act. The counts stay at 10 and 15. Texas sets the most demanding standard for the record itself. When an acknowledgment is not made in writing, Texas Insurance Code Section 542.055(c) requires the insurer to "make a record of the date, manner, and content of the acknowledgment."

Two operational readings follow. For a TPA, each client program has its own intake channel and may carry a contractual service standard tighter than the statute, so the clock that decides the client audit can sit in the contract. For an MGA claims team working under delegated authority, a capacity partner's claims audit will read acknowledgment dates and file documentation, and a lean team cannot rebuild either after the fact. The fix is the same for both: stamp first receipt wherever it happens, and close the gap between that stamp and claim creation.

Extraction is the visible part of intake. The record is the hard part, because it is what the regulator, the adjuster and every later stage of the claim read. Guidewire's documentation is exact about the state a claim starts in: "All open claims initially start as draft claims, regardless of how they are created." A draft claim is saved but lacks enough information for adjudication, draft claims are not assigned to any user, and open claim numbers are assigned only when the claim moves from draft to open. A notice that never leaves draft is a notice nobody is working.

Promotion depends on exposures. An exposure tracks a potential payment, and in ClaimCenter every exposure is linked to one coverage and one claimant. Guidewire's own example is a two-car accident in which one driver hits another and the second driver suffers a neck injury: one claim, three exposures, for each car's repairs and the injured driver's medical bills. To be assignable, an exposure needs a coverage and coverage subtype, a claimant and an incident. So one notice usually means one claim with several exposures, and intake has to split the facts into them correctly or some potential payments go untracked.

The regulator reads the same record. Section 4 of the NAIC model requires an insurer to be able to produce the claim number, line of coverage, date of loss and date of payment, denial or closure for the current year and the two before it. It requires claim files detailed enough "to permit reconstruction of the insurer's activities relative to each claim," and it requires each relevant document to be "noted as to date received, date processed or date mailed." Receipt dates are the evidence that the acknowledgment clock was met, which is the subject of what the claim-file rules asked for before AI.

One practice separates a usable intake record from a merely fast one: missing facts get flagged and requested, never filled in. A notice that says police attended but carries no report number should produce a dated request to the claimant, not a guessed number or a blank that looks complete. Under the Texas rule the request belongs to the same 15-day window, so a request logged with its date is also the proof that the deadline was met.

This is how Hesper builds the record. Its intake agents extract the facts from ACORD loss notices, emails, attachments and photos, create the claim in the core system through the API (Guidewire, Duck Creek, Majesco and others), and record where each fact came from. Every agent action is logged with its source, its reasoning and its timestamp, so the receipt date, the request and the reason behind each flag sit in the claim file rather than in someone's memory.

Severity is worth flagging at intake because cost is concentrated. NCCI's analysis of large losses by Jim Davis and Ryan Voll, published March 31, 2020, found that workers' compensation claims above $1 million account for 10% of losses but less than 0.2% of lost-time claim counts, fewer than 1 claim in 500 (derived), with similar shares across accident years 2002 to 2015. Report lag is one signal that travels with cost. NCCI's 2015 research brief on accident report lag, by Thomas Sheppard, covered lost-time claims from report years 2010-2011 valued at 18 months in 44 states, and found median case-incurred costs of $13,844 for claims reported in the first week, $13,210 in the second, $17,785 in the third, $19,936 in the fourth and $19,251 after that. The week-three median is about 35% above week two, $4,575 more (derived), and week four adds another 12%.

Two caveats travel with those numbers. NCCI says plainly that "while we are able to identify correlations in the data, we are not able to determine cause and effect relationships," and the data cover workers' compensation lost-time claims only. Same-day reports are left out of the comparison because NCCI notes their median, $17,298, likely includes very severe injuries that require immediate medical attention. Report lag is mostly the claimant's or the employer's delay, which intake cannot shorten. What intake can do is stop adding days of its own, and treat a late report as a severity signal rather than a routine file.

Attorney involvement moves with lag, and further. In the same NCCI data, 12.8% of claims reported with no lag involved an attorney, against 31.7% of claims reported after the fourth week, about 2.5 times the rate (derived). Most claims arrive quickly, which is what makes the late ones stand out: 19.7% were reported with no lag, another 52.2% during the first week, and more than 80% within two weeks.

The signals worth flagging are the ones a person would act on, each with its reason: an injury, an attorney letter, a possible liable third party, a late report, contradictory facts and fraud indicators. Scoring at first notice is an established category. Gradient AI launched ClaimVoyant in March 2026 to flag workers' compensation claims "likely to be expensive, complex or involve delayed recovery periods" from data captured at FNOL, and CLARA Analytics runs early triage of high-cost casualty claims.

The flag is where intake hands off. In Hesper, severity, fraud and litigation scoring run together at triage, clean claims route straight through, and flagged claims are assigned by license, skill and workload with the reasons attached. Clean claims resolve straight through; suspicious claims get an investigation-grade workup. The routing logic is covered in our claims triage automation playbook.

The case for keeping these with a person is in the claims-practices rules, not in a general unease about AI. Section 5.D of the NAIC model bars an insurer from denying a claim because a first-party claimant failed to give written notice within a time limit, unless written notice is a written policy condition or the claimant's failure, after being asked, is so unreasonable that it breaches the duty to cooperate. That is a judgment about the policy wording and the claimant's conduct, and an intake system that declines a late report has made it without looking at either. Nor can intake decide that an informal notice is not a claim, given how broadly the model defines a notification of claim.

Fraud signals carry a similar constraint. The model's relief for suspected fraud in Section 7.A applies only "where there is a reasonable basis supported by specific information available for review by the insurance regulatory authority." A score with no file behind it does not obviously meet that bar. The right action at intake is to route the signal, with its reasons, to the people who will build the file, not to hold, slow or deny the claim on the strength of a number.

The AI rules point the same way. The NAIC model bulletin on insurers' use of AI systems, adopted December 4, 2023, covers claim management and fraud detection and says an insurer's actions must not violate the unfair trade practices or unfair claims settlement practices laws "regardless of the methods the Insurer used to determine or support its actions." It asks for controls commensurate with factors that include "the extent to which humans are involved in the final decision-making process." By the NAIC's implementation map as of October 8, 2026, 29 jurisdictions had adopted the bulletin and 4 more (California, Colorado, New York and Texas) had insurance-specific regulation or guidance. Some legislators now want the human sign-off written into statute, as in New Jersey's AI claim denial bill.

Reserves stay with a person for a simpler reason: reserve authority is a delegated limit, and an intake step does not hold one. Five Sigma's FNOL guide lists "Reserves are populated" within its Orchestrate stage, and a populated figure is a useful starting point for the adjuster who holds that authority. As for what a good intake guardrail sounds like, Liberate's demo agent tells the caller it "can't promise coverage I can't confirm" before handing off to a licensed representative. That is the right line.

Hesper draws the same line at intake. Its agents capture notices from every channel, extract the facts, create the claim and flag severity and early signals. The claims team owns the customer and agent relationship, handles the exceptions the agents route to a person, sets the fields and rules its system requires, and decides reserves and next steps. Adjusters keep decision authority on every claim.

Front-door vendors have made reporting a loss fast, and some now publish measured results. Hippo's September 24, 2026 release says its AI claims agent, Clara, "successfully completes 81% of attempted FNOL calls with no human intervention," based on Hippo's internal August 2026 data, which leaves 19% (derived) that needed a person. Hippo's Clara is unrelated to CLARA Analytics. Hi Marley reports 130+ customers, including 13 of the top 30 P&C carriers, and a 35-50% reduction in inbound calls, noting that its statistics come from certain carriers and results vary. Liberate reports that one customer's agent resolves 85% of FNOL calls end to end, escalating when a request falls outside the guardrails.

None of that is in dispute; the argument is about the handoff. The public materials of the front-door and intake-to-file vendors describe capture, identity and policy lookup, claim filing, indexing, routing and status. They do not describe carrying one evidence file from the intake record through the coverage position, a completed investigation on a flagged claim and the subrogation demand. Bevaya describes identifying subrogation opportunities early, and Five Sigma's Clive Risk agent "flags suspicious activity and recommends further investigation", which is the right instinct and a different job from carrying the file. When the intake record shrinks to a row of fields, the evidence behind those fields is gathered again at each later stage. That shows up in cycle time. J.D. Power's 2025 study put the span from first notice of loss to final payment for homeowners at more than 44 days, the longest since 2008, and the 2026 study put the average time to final payment at 40.7 days, down 3.4.

Hesper is built around that handoff. Every claim, first notice to final recovery. The intake record, with each fact's source, each document's receipt date and each flag's reason, is the first page of the evidence file that the coverage position, the investigation and the recovery all reuse, so nothing is gathered twice. A liable third party named on the first call starts the subrogation file instead of being rediscovered at closure, the argument made in missed subrogation opportunities are lost at intake. A fraud signal raised at intake usually lands in a manual queue where, by Hesper internal benchmark, an investigation takes 14+ days per case, one investigator carries 200+ cases, and only about 25% of flagged claims get a real workup. Hesper runs 15+ investigation phases in parallel on a suspicious claim, in minutes, not weeks, starting from the same intake record, which takes coverage of flagged claims to 100%. Evidence behind every decision. Adjusters review evidence-backed files instead of building them.

Hesper does not replace the front door or the system of record. It captures notices directly or works behind the portal, conversational tool or call center a carrier already runs, creates the claim in Guidewire, Duck Creek, Majesco or another core system through the API, and runs its built-in fraud detection standalone or alongside FRISS, Shift or Verisk where they are already in place.

Ten checks, in the order a notice meets them. Each can be tested against a sample of last quarter's files.

1. Start the clock at first receipt anywhere: the insurer's agent, a broker, the portal, a fax line or a shared inbox. Whether a broker counts as the insurer's agent depends on the appointment and the state, so stamp it either way.
2. Date every document on receipt, as Section 4.C of the NAIC model requires.
3. Log every non-written acknowledgment with its date, manner and content. That is the Texas standard and the strictest of the six rules above.
4. Attach the source document and the page, line or timestamp to every extracted field.
5. Flag missing facts and request them, with the request dated. Never fill them in.
6. Create exposures with a coverage, a claimant and an incident, so the claim is assignable when it lands.
7. Give every severity flag a written reason: injury, attorney, liable third party, late report, inconsistency or fraud indicator.
8. Keep the intake agent from promising coverage, denying for late notice or acting on a fraud score.
9. TPAs: load each client program's service standard wherever it is tighter than the statute, and measure against both.
10. MGAs: keep acknowledgment dates and file documentation in the form a capacity partner's claims audit reads.

The test of an intake process is whether someone can reconstruct, a year later and from the file alone, when the insurer first knew, what it asked for, what it flagged and why. Build for that and the speed follows, because nothing has to be gathered twice.