New Jersey's AI claim denial bill regulates who signs, not what is in the file

The entire operative rule in New Jersey Assembly Bill A5494 is one sentence long: an insurer may not use artificial intelligence to make the final decision to deny an automobile, homeowners or flood claim. Three sections, about 140 words of enacted text, and a sponsor's statement that runs about as long as the act it explains. Nothing in the bill says what the human who makes that decision has to be handed, has to check, or has to write down. A statute that regulates the signature can be satisfied by a signature.

Start with what the bill is not. A5494 is not law. The New Jersey Legislature's bill record shows it introduced on 14 September 2026 and referred the same day to the Assembly Science, Innovation and Technology Committee. No hearing, no floor vote, no identical Senate bill. Trade coverage has been inconsistent about the referral, and the Legislature's own record controls.

The comparison that makes the bill legible is not New Jersey against no regulation. It is New Jersey against Florida. Six months earlier the Florida House passed CS/CS/HB 527 by 108 votes to nothing. That bill let AI assist in claims handling and even generate a recommendation to approve or deny, then required a qualified human professional to analyse the claim independently of the AI system, review the accuracy of the AI output, document the basis for the determination, and affirm in the denial letter that AI was not the sole basis for it. Florida wrote the evidence standard. New Jersey wrote the signature line. Florida's bill died in Senate Rules; New Jersey's sits in committee.

The second thing to know is that New Jersey already regulates AI in claims, and has since February 2025. Bulletin 25-03 from the Department of Banking and Insurance expects every authorized insurer to run a written AI Systems Program covering "claim administration and payment, and fraud detection", and it lists what the Department can demand in a market conduct action. A5494 never mentions it. What follows is the bill clause by clause, the multi-state map around it, and the part a carrier can act on either way: what belongs in the file behind a denial. The decision-rights question underneath all of it - which claims decisions stay with a person, and what that person is handed - is mapped in which claims decisions stay human, and where AI already sits across the lifecycle is in the state of claims automation report.

The whole of A5494 is a title, two working sections and an effective date. The title reads "An Act concerning the use of artificial intelligence in deciding certain insurance claims and supplementing Title 17 of the Revised Statutes." Section 1 defines two terms. Section 2 creates one prohibition and one penalty. Section 3 says the act takes effect immediately, with no rulemaking delegated and no implementation runway. Reading the introduced text end to end takes about two minutes, which is the first thing worth knowing about it.

The definition of artificial intelligence is narrower than the coverage suggests. Section 1 reaches a technology system that "(1) is trained on data; (2) is designed to simulate human communication through one or more of the following: (a) text; (b) audio; or (c) visual communication; and (3) generates non-scripted outputs with limited or no human oversight." All three limbs have to hold together. A gradient-boosted fraud model that returns a score, or a rules engine that returns a flag, is trained on data and produces output with little oversight, but it is not obviously designed to simulate human communication. On the face of the definition the bill is drafted around generative systems. Whether it reaches the predictive models most carriers already run in claims is a question the text leaves open, and the bill delegates rulemaking authority to nobody who could close it.

That is the entire prohibition. Read what is absent from it. There is no "in whole or in part", which is the construction Texas used. There is no "sole basis", which is the construction Florida and California used. There is no "materially relied upon". It bites on one act, the final decision to deny, and on nothing adjacent to it.

Section 2.b sets the consequence: "An insurer who violates this act shall be subject to a penalty of up to $5,000 for each violation", collected in a summary proceeding in accordance with the Penalty Enforcement Law of 1999, P.L.1999, c.274 (C.2A:58-10 et seq.). Per violation, so on a personal auto book the arithmetic scales with denial volume, not with the model. The bill names no enforcing officer. It does not vest the Commissioner of Banking and Insurance with anything, sets out no administrative penalty schedule, and creates no private right of action. A summary proceeding is a court process, so who brings it is left to be worked out later.

Scope is three lines. Section 1 defines an insurer by reference to the "homeowners, flood, or automobile insurance" business in New Jersey, and Section 2.a repeats the same three policy types. Commercial property, general liability, commercial auto, workers' compensation, life, health, disability, title, pet and specialty all sit outside it. Most carriers, TPAs and MGAs write more than three lines, so on a mixed book this is a rule about part of the claims operation rather than all of it.

The record is short. Synopsis: "Prohibits certain use of artificial intelligence in denying homeowner, automobile, and flood insurance claims." Primary sponsor: Assemblyman Chris Tully. Introduced: 14 September 2026. Status: introduced, referred to the Assembly Science, Innovation and Technology Committee, same date. Identical Senate bill: none listed.

One procedural fact is worth stating neutrally, because it bears on whether the bill gets a hearing at all. Tully chairs the Assembly Science, Innovation and Technology Committee, the committee his own bill was referred to. Sponsoring a bill into a committee you chair is ordinary legislative practice and is not evidence of anything improper. It does mean the usual first obstacle, persuading a chair to schedule a hearing, is not the obstacle here. The absence of a Senate companion is the more informative gap: a bill with no identical Senate bill needs a Senate sponsor before it can move, and none is listed.

Two trade reports conflict on the referral. Insurance Business reported that the bill had not yet been referred to committee. Repairer Driven News, on 18 September 2026, reported that it had. The Legislature's own record agrees with the second, so treat the referral as fact. The same Insurance Business piece carries the framing the next section takes apart: that insurers could still use automated tools for triage, assessment, fraud scoring and approvals. That is a sound inference about what the bill would leave alone. It is not a description of the text, because the text says nothing about any of them.

For a compliance calendar the practical read is this. Nothing in New Jersey changes today. If the bill did advance in its current form, Section 3 gives no transition period at all, so the version to plan against is one that becomes effective on signature. That argues for treating the exposure as a documentation question now instead of a project later, which happens to be the same conclusion the Department's own bulletin pointed at nineteen months ago.

Trade coverage reads A5494 as leaving triage, assessment, fraud scoring and approvals available, and on the outcome that reading is right. The mechanism is the part a compliance officer has to work with, and the mechanism is not permission. Section 2 contains no carve-out of any kind. It names one prohibited act and is silent on everything else. Those uses survive by omission.

Compare the two drafters who did it explicitly. Texas SB 815, signed 20 June 2025 and effective 1 September 2025, bars a utilization review agent from using an automated decision system to make an adverse determination "wholly or partly", then writes the exception into the statute: the requirement does not apply to the use of an automated decision system for "administrative support or fraud-detection functions". Florida's committee substitute went the other way and explicitly allowed AI to assist in claims handling, including the generation of recommendations to approve or deny a claim. Both legislatures decided what the machine may do. New Jersey's bill decides what it may not do, once.

The difference shows up in three places. An express carve-out is a defence a carrier can cite, while silence is a boundary that has to be argued. An express carve-out fixes legislative intent at enactment, while silence leaves intent to be reconstructed from a sponsor's statement. And an express carve-out survives amendment pressure better, because deleting a clause is a visible act while widening a prohibition reads as a drafting tweak. "Permitted by omission" is a weaker position than "permitted" even when the day-one effect is identical.

The asymmetry: an AI approval is untouched

Section 2.a prohibits using AI to make the final decision to deny. It says nothing about the decision to pay. Under A5494 an insurer could approve and pay a claim on a fully automated decision with no human in the path, and the bill would have nothing to say about it. The asymmetry is defensible in the sense that consumer-protection drafting follows consumer harm, and a wrong denial harms a policyholder in a way a wrong approval does not. It is worth naming anyway, because the leakage side of the ledger is where a claims executive spends most of her risk budget, and no AI-denial statute in the country addresses it. Speed and leakage are the same problem; this class of bill regulates one of them.

The open question: partial denials and payment reductions

Most contested property and casualty outcomes are reductions, not flat denials: a coverage sublimit applied, a depreciation position taken, a supplement partly allowed. If the prohibition reaches only a total denial, then the most common adverse outcome on an auto or homeowners claim sits outside it. If a court read deny a claim to include denying part of one, the bill would reach a large share of routine handling with no evidence standard attached. That gap is the single amendment that would change the bill's practical reach the most, in either direction.

The bill creates a decision-maker and says nothing about the record that decision-maker holds. A full read of the introduced text turns up no provision on any of the following: what information the reviewing human must be given; what "review" means; any duty to disclose to the policyholder that AI was involved; any recordkeeping, documentation or retention requirement; any appeal right; any attestation or certification; any rulemaking authority that could add those later. The bill is not vague about them. It is silent.

The consequence is mechanical. If the only rule is that a machine cannot make the final decision, then compliance is demonstrated by showing that a person did. The artefact that proves a person decided is a name, a role and a timestamp in the claim system. Every carrier already produces that artefact on every denial, which is why the fair reading of the bill is that most carriers would satisfy it on day one and notice nothing on the effective date.

That is not a criticism of the sponsor. The bill responds to a documented risk, and it recites that risk in the same terms the state's own insurance regulator uses. The gap is one clause wide. A single sentence requiring the reviewing individual to have before them the claim file, the policy language relied on, and the AI output together with the inputs that produced it would convert the prohibition into a record. Florida wrote roughly that sentence and 108 House members voted for it.

The evidence that human review needs a specification comes from fields that ran the experiment. Dratsch and colleagues, in Radiology in 2023, gave 27 radiologists 50 mammograms carrying BI-RADS suggestions attributed to an AI system, wrong on 12 of the 40 cases in the test set. When the suggestion was correct, inexperienced readers rated 79.7% of mammograms correctly; when it was wrong, 19.8%. Moderately experienced readers went from 81.3% to 24.8%. Very experienced readers went from 82.3% to 45.5%, a difference the authors report at P = .003. Their conclusion is that readers at all three experience levels "are prone to automation bias when being supported by an AI-based system."

Read the very-experienced row first. Expertise roughly halved the damage and came nowhere near preventing it. This is radiology and not claims, and the percentages do not transfer to adjusters: reading a mammogram is a perceptual task with a defined answer, and a claim denial is a judgement about coverage, causation and credibility. The transferable finding is structural. When a system's output arrives before the human judgement, the output moves the judgement, and seniority is a partial defence rather than a control.

The field version of the same problem has so far been documented only in health insurance. ProPublica and The Capitol Forum reported in March 2023, from internal documents, that Cigna doctors denied over 300,000 requests in a two-month period at an average of 1.2 seconds each, a characterisation Cigna disputed, saying the system was designed to accelerate payment of claims. That is reporting, not a finding, and it describes a health insurer's utilization review rather than a property and casualty claim. It is also the clearest published picture of what a review requirement looks like when nobody specified what review means.

The US Senate Permanent Subcommittee on Investigations majority staff report of 17 October 2024, built on more than 280,000 pages obtained from insurers, found that in 2022 UnitedHealthcare and CVS denied prior authorisation requests for post-acute care at three times the rate at which they denied requests for other types of care, and that Humana denied post-acute requests at sixteen times the rate at which it denied other requests, in the context of predictive technology use. Again: Medicare Advantage prior authorisation, not property and casualty claims.

The litigation is in the same place. In Estate of Gene B. Lokken v. UnitedHealth Group in the District of Minnesota, plaintiffs allege the nH Predict model was used to evaluate medical necessity for post-acute care in place of physician review; in February 2025 the court declined to dismiss certain state-law claims, allowing breach of contract and good faith claims to proceed, and discovery into how the model works has been compelled. In Kisting-Leung v. Cigna in the Eastern District of California, the court in March 2025 allowed parts of a proposed class action over the PxDx algorithm to proceed while dismissing others, and accepted the insurer's evidence that the three named plaintiffs did not have their claims denied through that review. These are allegations; nothing has been proven in either case. There is also no verified property and casualty equivalent. The fact pattern is being litigated in health insurance while P&C carriers watch it.

All of which argues for treating this as a file-design problem rather than a bill-tracking problem. A statute can require a signature. It cannot require attention. The only thing that distinguishes a considered denial from a ratified one, eleven months later in a market conduct exam or a bad-faith deposition, is the record of what was checked before the signature went on.

Florida's bill was titled "Mandatory Human Reviews of Insurance Claim Denials" and it started from the opposite end of the problem. AI was allowed to assist, including generating a recommendation to approve or deny. What it could not be was the sole basis for denying a claim, denying a portion of a claim, or reducing a payment. Around that limit, the House staff analysis of 2 March 2026 describes a set of affirmative duties on the human: a qualified professional had to analyse the claim independently of the AI system, review the accuracy of the AI output, determine whether the claim was payable, and record the basis for that determination including the information the AI provided. The denial letter had to affirm that AI was not the sole basis. The retained record had to carry a unique identifier for the individual who made the decision.

It was not a marginal bill. The Insurance and Banking Subcommittee passed it 18-0 on 9 December 2025, the Commerce Committee 26-0 on 26 February 2026, and the House 108-0 on 5 March 2026. Proposed effective date: 1 July 2026. Proposed fines for workers' compensation carriers: up to $2,500 per nonwillful violation, not to exceed $10,000 for all nonwillful violations arising from the same action, and up to $20,000 per willful violation, not to exceed $100,000, with market conduct examination authority sitting at two separate state agencies. On 13 March 2026 it died in Senate Rules.

Two honest readings of that. The first is that Florida's bill was much broader in scope - workers' compensation carriers, most authorized insurers and health maintenance organizations, with express exclusions for reinsurance, wet marine and transportation, title, and credit life or credit disability coverage - and a broader bill carries more to negotiate in a short session. The second reading is the one that matters for New Jersey: the drafting work has already been done. If A5494 gets a hearing, the amendment that would give it an evidence standard exists in another state's legislative record, passed a chamber unanimously, and could be lifted almost verbatim.

Nearly every enacted AI-denial rule in the United States is a health-lines utilization review rule. That is not an accident of drafting. Prior authorisation is where the volume, the automation and the consumer complaints concentrated first, so it is where legislatures wrote first.

Arizona is the reason to read the chaptered law and not the bill. The House-engrossed version of HB 2175 would have provided that "ARTIFICIAL INTELLIGENCE MAY NOT BE USED TO DENY A CLAIM OR A PRIOR AUTHORIZATION FOR MEDICAL NECESSITY, EXPERIMENTAL STATUS OR ANY OTHER REASON THAT INVOLVES THE USE OF MEDICAL JUDGMENT", and would have made a denial without individual review an act of unprofessional conduct. The Senate struck all of it. What the Governor signed on 12 May 2025 as Chapter 165, effective 1 July 2026, does not mention artificial intelligence anywhere. New A.R.S. 20-3103 provides that before a health care insurer may deny a provider-submitted claim on the basis of medical necessity, "THE MEDICAL DIRECTOR SHALL INDIVIDUALLY REVIEW THE DENIAL", and that during each review the medical director "SHALL EXERCISE INDEPENDENT MEDICAL JUDGMENT AND MAY NOT RELY SOLELY ON RECOMMENDATIONS FROM ANY OTHER SOURCE". A companion section, 20-3407, does the same for a direct denial of prior authorisation. Arizona ended up regulating the quality of the review instead of the identity of the tool, which is the specification A5494 leaves out.

California's SB 1120, chaptered on 28 September 2024, provides that AI "shall not deny, delay, or modify health care services based, in whole or in part, on medical necessity", and requires that determination to be made by a licensed physician or a competent professional, with non-discrimination, disclosure, regulator inspection and periodic review duties attached. Texas SB 815 is the "wholly or partly" construction with the fraud-detection carve-out, and it lets the commissioner audit the automated decision system at any time. Note what an audit power really is: a documentation requirement in disguise, because a system that can be audited at any time has to be documented all the time. Illinois HB 35 would create an Artificial Intelligence Systems Use in Health Insurance Act barring adverse consumer outcomes resulting solely from an AI system; it passed the House 79-35 on 9 April 2025, and the last action on its record is a re-referral to Senate Assignments on 2 June 2025.

New York is the instructive near miss for anyone in P&C. DFS Circular Letter No. 7 (2024), issued 11 July 2024, sets governance, testing and transparency expectations for external consumer data and AI systems across all lines, and then draws its own boundary: "This Circular Letter also is not intended to address phases of the insurance product lifecycle other than underwriting and pricing." A carrier that built its AI governance to CL7 built it for the front of the policy, not for the claim.

Washington is the closest thing in force to a property and casualty evidence standard, and it regulates the investigation rather than the signature. Under rulemaking R 2025-05, adopted 18 August 2026 and effective 18 October 2026, the adopted filing amends 15 claim-handling sections. WAC 284-30-330(4) as amended makes it an unfair claims settlement practice to deny or refuse to pay a claim "in part or in full without conducting a reasonable investigation", and adds that "A reasonable investigation may not rely solely on the use of a database, which includes, but is not limited to, estimating software and benchmarks gathered from one or multiple databases."

The amended definition of investigation in 284-30-320 reaches "a reasonable assessment of whether each claim or loss is covered under the policy, the scope and value of loss or damage caused by an event, and the reasonableness of the costs or expenses". Two provisions in the same filing are worth a compliance officer's time: 284-30-380(7) makes insurers responsible for the accuracy of their evaluations "including, but not limited to, evaluations made on their behalf", and 284-30-380(8) requires an insurer that uses a database, survey, estimating software or benchmarks for material pricing or labour rates to tell the claimant, on request, when that data was collected, where it came from and which businesses supplied it.

Read the adopted text and not the proposal: an earlier draft of 284-30-330(4) would have required an investigation to "include an individual assessment of loss or damage", and that phrase did not survive adoption. Nothing in the rule asks who signed. It asks what the investigation contained and where the numbers came from.

Over the top of all of it sits the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted 4 December 2023. The NAIC implementation map, status as of 1 April 2026, shows 25 jurisdictions adopted, with four more - California, Colorado, New York and Texas - carrying their own insurance-specific AI regulation or guidance. Guidance is not statute. It is also the instrument most likely to be quoted back at a carrier in an examination, because the examiners wrote it.

Three conclusions hold across that table. No US state has a statute in force requiring a human to make the final denial decision on a property and casualty claim; the two bills that would have done it are in committee and dead respectively. The closest in-force P&C requirement is Washington's, and it is an investigation standard and not a decision-rights rule. And New Jersey's own regulator already reaches AI in claims, which is the next section.

Bulletin 25-03 was signed by Commissioner Justin Zimmerman on 11 February 2025 and addressed to all insurers authorized or admitted in New Jersey. It is New Jersey's adoption of the NAIC model bulletin, and it does the thing A5494 does not: it describes documentation.

Start with the overlap, because it shows where the bill came from. The bulletin says AI "can present unique risks to consumers, including the potential for inaccuracy, unfair discrimination, data vulnerability, and lack of transparency and explainability." A5494's STATEMENT recites "inaccuracy, unfair discrimination, data vulnerability, and lack of transparency". Same four risks, same order. The bill's rationale is the bulletin's rationale. What the bill adds is a decision-rights boundary on top of a governance regime that already exists, and it adds it without an evidence standard.

The bulletin's legal hooks are the Unfair Trade Practices Act (N.J.S.A. 17:29B-1 et seq.) and the Unfair Claims Settlement Practices Act (N.J.S.A. 17B:30-13.1 et seq.), which sets standards for the investigation and disposition of claims. Its central move is one sentence long: those standards apply regardless of the methods the insurer used to determine or support its actions. Method-neutral, which means an AI-assisted denial is judged by the same standard as a manual one and carries the same exposure when the file is thin.

Two of the bulletin's proportionality factors read differently next to the bill. Section 3 tells insurers to calibrate controls by, among other things, "the extent to which humans are involved in the final decision-making process" and "the transparency and explainability of outcomes to the impacted consumer". The regulator already treats human involvement as a variable on a dial, weighed against explainability. The bill treats it as a switch, and drops the explainability half.

The bulletin also carries the disclosure provision the bill lacks. AI Systems Program guideline 1.9 expects processes for "providing notice to impacted consumers that AI Systems are in use and provide access to appropriate levels of information based on the phase of the insurance life cycle in which the AI Systems are being used." A policyholder whose claim was scored by a model has an expectation of notice under the bulletin today. A5494 would add nothing to it.

Section 4 is the part to read with a market conduct examiner in mind. On investigation or in a market conduct action, the Department may ask for:

• The written AI Systems Program itself, plus documentation of compliance with all applicable AI Program policies, protocols and procedures.
• Model inventories and descriptions of the AI systems in use.
• Information about the data source, provenance, data lineage, quality, integrity, bias analysis and minimization, and data currency.
• Documentation related to validation, testing and auditing, including evaluation of model drift.
• Third-party due diligence records, vendor contracts, and audit reports.

Guidelines 4.1 to 4.3, on third-party AI systems and data, turn the vendor half of that into a procurement checklist. The bulletin expects due diligence on AI vendors, contract terms providing "audit rights and/or entitle the insurer to receive audit reports by qualified auditing entities", and terms requiring the vendor to cooperate with the insurer on regulatory inquiries and investigations. That is actionable this quarter regardless of what happens to A5494, and it is a better test of a claims AI vendor than any question about whether the product denies claims. Ask for the audit clause and the model documentation before the pilot, not during the exam.

The bulletin is careful about its own reach: "The goal of the bulletin is not to prescribe specific practices or to prescribe specific documentation requirements." That limit cuts both ways. A carrier cannot point at a checklist and call itself finished, and the Department is not bound by one either when it asks what a decision rested on. New Jersey's regulator asks what the insurer can show. The bill asks who signed. A carrier that can satisfy Section 4 has already built the thing A5494 forgot to ask for.

Regulators do not need an AI statute to find a denial pattern. They need a ratio and a data call, and they have both. The NAIC Market Conduct Annual Statement ratios for 2025, version 2025.0.4, define for Property and Casualty (Private Passenger Auto and Homeowner), among others: the number of claims closed without payment compared to the total number of claims closed; the percentage of claims unprocessed at the end of the period; the percentage of claims paid beyond 60 days; and suits opened during the period compared to claims closed without payment. Private Flood is a separate MCAS line with its own ratios, including the same suits-to-claims-closed-without-payment construction.

Line those up against A5494 and the overlap is exact. The three lines the bill would regulate - private passenger auto, homeowners and private flood - are three lines where states already collect closed-without-payment behaviour annually, carrier by carrier, and where the suits ratio measures how often those non-payments turn into litigation. The detection infrastructure for the behaviour the bill worries about is already installed. What a carrier gets asked for when its ratio moves away from the peer group is not an attestation that a person signed. It is the files.

The examination question is also about to acquire a standard form. NAIC reports its AI Systems Evaluation Tool being piloted by 12 states as of March 2026, with adoption anticipated at the 2026 Fall National Meeting. The separate AI Risk Evaluation Supplement reached version 5.0 on 31 August 2026 and is out for a comment period that closes 29 September 2026. What that supplement does to a claims or fraud model in practice is walked through in the NAIC AI risk evaluation piece.

The same NAIC source sizes the exposure. Among insurers surveyed, 88% of 193 auto respondents and 70% of 194 homeowners respondents use, plan to use, or are exploring AI and machine learning. Florida's House staff analysis recites the same figures and adds the detail that lands the point: the property and casualty insurers surveyed reported using AI in claims handling to analyse accident images, to estimate claim settlement values, and to detect fraud. None of those three is a denial, and where each of them sits in a claims operation is mapped in the guide to automating the claims lifecycle. The prohibition A5494 would create sits above where the technology is actually deployed, which is why the bill would change so little on the day it took effect.

Four instruments, three in force and one dead, converge on the same requirement, and none of them is satisfied by a signature. Bulletin 25-03 asks what the insurer can produce about the model and the decision. Washington's amended WAC 284-30-330 asks whether the investigation rested on something more than a database lookup. The Unfair Claims Settlement Practices Act asks whether the investigation and disposition met the standard, regardless of method. Florida's bill asked the human to record the basis, including what the AI contributed. The common object is a reconstructable record.

That record has two audiences and they want the same thing. A market conduct examiner wants to know what was checked before the decision. Defence counsel in a bad-faith action wants the identical list with dates attached. The overlap is close to total, which is the argument in what survives a courtroom and the specification in the defensibility standard for investigation AI.

A denial file that answers both audiences, whatever happens to A5494, carries six things:

1. The policy language the coverage position rests on, quoted and cited to the form and the endorsement rather than summarised.
2. Every source checked during the investigation, including the ones that came back clean, with the date and time each was checked.
3. The AI output that informed the decision, the inputs it ran on and its version, so the recommendation can be reproduced rather than described.
4. The reviewing human's identity, role and decision time, plus what they changed if they changed anything.
5. The factual and legal basis for the denial in the words that went to the policyholder, matched to the evidence in the file supporting each element.
6. The retention posture for both the file and the model record, set by the claim-file retention rule in each state you write rather than by the AI policy.

One more thing changes the vendor conversation. Almost no claims-automation vendor publicly claims to deny claims autonomously. Five Sigma's Clive Risk agent flags suspicious activity and recommends further investigation. CCC surfaces high-risk claims before payment and prioritises investigator effort. Shift Technology describes an agentic system built to assess, prioritise, advise and act on every claim, human-in-the-loop by design, preparing the steps for human action or automation. FRISS scores every policy request, renewal and claim in real time and routes the risky ones to investigation, and Verisk returns claim scores and reason codes for triage. A score is not a decision. Owl.co describes itself as AI that reads the entire claim file and answers with citations. On the substance the market already sits on the permitted side of the line A5494 draws, so the bill would ratify existing vendor behaviour more than it would constrain it.

Which means the useful vendor question is not the one the bill asks. Everybody answers no to "does your AI deny claims". The question that separates vendors is what the adjuster can produce when the adjuster denies: whether the tool that shortened the clock also thickened the record, and whether that record exists on the ordinary file or only on the one somebody flagged. If a fraud score materially drives a denial and nobody can explain the score, the problem is not A5494. It is Bulletin 25-03 Section 4 and the Unfair Claims Settlement Practices Act, both already in force.

Hesper sits on the second question, and the claim is narrow enough to check. Hesper AI is an AI claims resolution platform: agents take a claim from first notice to final recovery, and the evidence is a by-product of doing the work, not a layer that reads the file after somebody else built it. Coverage analysis produces a cited coverage position on the way to the decision. The investigation runs its phases in parallel and writes down what each source returned, including the clean ones. Settlement and recovery inherit that record instead of rebuilding it. Evidence behind every decision. The numbers underneath that are Hesper internal benchmarks and not industry figures: 15+ investigation phases running in parallel, against a manual SIU baseline of 14+ days per case and caseloads of 200+ cases per investigator, which is why roughly 25% of flagged claims get a full manual workup; with the phases running in parallel that goes to 100% of flagged claims, in hours rather than weeks.

What Hesper does not do is the part that matters for this bill. Hesper takes no denial, no coverage position, no fraud-bureau filing and no examination-under-oath initiation autonomously; those stay with the adjuster, the investigator or counsel in every configuration. Adjusters review evidence-backed files instead of building them. That is a design choice about decision rights, not a compliance claim, and it is worth being explicit that nothing here is an assertion of compliance with A5494, which is not law and may never be. The stage-level detail sits on coverage verification, claim file review and fraud investigation, with the whole path on the platform page.

A5494 is a reasonable response to a documented risk and it is one clause short of doing what its sponsor's statement says it does. If it advances, the clause is available: Florida drafted it and 108 House members voted for it. If it does not advance, nothing in a New Jersey carrier's obligations changes, because Bulletin 25-03 asked the harder question in February 2025, and the answer to that question is the file.