The entire operative rule in New Jersey Assembly Bill A5494 is one sentence long. Nothing in the bill says what the human who makes the denial 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.
A5494 is a proposed AI claim denial law, not an enacted one. The New Jersey Legislature's bill record shows it introduced on 14 September 2026 and referred to committee the same day. No hearing, no floor vote, no identical Senate bill.
What follows is the bill clause by clause, the multi-state map around it, and what belongs in the file behind a denial. Where AI already sits across the lifecycle is in the state of claims automation report.
What New Jersey's A5494 would actually do
Answer
What is the operative text of New Jersey A5494?
A5494 would prohibit an insurer writing automobile, homeowners or flood coverage in New Jersey from using artificial intelligence to make the final decision to deny a claim. Violations carry a penalty of up to $5,000 each, collected in a summary proceeding. The bill is three sections long and takes effect immediately on enactment.
The introduced text takes about two minutes to read. No rulemaking is delegated and there is no implementation runway.
That is the entire prohibition. There is no "in whole or in part", the construction Texas used, and no "sole basis", the construction Florida and California used. It bites on one act, the final decision to deny, and on nothing adjacent to it.
The definition of artificial intelligence is narrow. Among its three limbs, the system must be "designed to simulate human communication". A fraud score or a rules flag is not obviously that. Whether the bill reaches the predictive models most carriers already run in claims is a question the text leaves open.
A5494 bites on one act, and its definition is drafted around generative AI
All three limbs of the Section 1 definition have to hold together; Section 2.a then reaches the final decision to deny and nothing adjacent to it.
- 1.Is trained on data
- 2.Is designed to simulate human communication: text, audio or visual
- 3.Generates non-scripted outputs with limited or no human oversight
The penalty is per violation, so the arithmetic scales with denial volume, not with the model. The bill names no enforcing officer. And on a mixed book this is a rule about part of the claims operation rather than all of it.
What the bill leaves alone, and why permits is the wrong word
Answer
Does A5494 allow insurers to use AI elsewhere in claims?
Yes, but by omission rather than by permission. The prohibition reaches only the final decision to deny. Triage, coverage analysis, estimating, fraud scoring, denial recommendations and approvals are untouched because the bill never mentions them, not because it carves them out the way Texas SB 815 does in statute.
Texas SB 815, signed 20 June 2025 and effective 1 September 2025, wrote its exception into the statute: the requirement does not apply to the use of an automated decision system for "administrative support or fraud-detection functions". An express carve-out is a defence a carrier can cite. Silence is a boundary that has to be argued. The table shows where each claim stage lands.
Two rows deserve a second look. Partial denials are unclear on the text, and most contested property and casualty outcomes are reductions, not flat denials. Approvals are untouched: an insurer could pay a claim on a fully automated decision with no human in the path. Leakage is where a claims executive spends most of her risk budget, and no AI-denial statute in the country addresses it.
The missing clause: what the reviewing human must see
Answer
What does A5494 require the reviewing human to do?
Nothing, on the face of the text. Section 2.a bars AI from making the final denial decision and stops. The bill specifies no evidence the reviewer must see, no definition of review, no disclosure to the policyholder, no recordkeeping duty and no retention period. The review language appears only in the sponsor's statement.
The sponsor's statement is not the statute
Trade coverage leads with the idea that A5494 requires a claims adjuster or investigator to review the denial. That phrase comes from the sponsor's explanatory STATEMENT, which describes the bill as "requiring a claims adjuster or investigator to review them." It does not appear in Section 2, and a statement is not enacted text.
If the only rule is that a machine cannot make the final decision, compliance is demonstrated by showing that a person did. The proof is a name, a role and a timestamp in the claim system. Every carrier already produces that.
The gap is one clause wide. A single sentence requiring the reviewer to have the claim file, the policy language relied on, and the AI output with the inputs that produced it would convert the prohibition into a record. Florida wrote roughly that sentence.
New Jersey wants a human to sign every claim denial. The bill never says what that human has to read first - and three instruments already in force do.
Why a review duty without an evidence standard becomes a signature
Answer
Why does human review of an AI recommendation drift toward approval?
Because review is a behaviour and a statute can only regulate a record. Where the machine's output arrives before the human judgement and the reviewer is volume-constrained, review drifts toward ratification. Radiology research finds even very experienced readers lose accuracy when the AI suggestion is wrong, and health-insurer reporting has described denial review measured in seconds.
Fields that ran the experiment show why human review needs a specification. Dratsch and colleagues, in Radiology in 2023, gave 27 radiologists 50 mammograms carrying suggestions attributed to an AI system. When the suggestion was wrong, accuracy fell at all three experience levels. Very experienced readers went from 82.3% to 45.5%.
This is radiology, and the percentages do not transfer to adjusters. 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.
ProPublica and The Capitol Forum reported in March 2023 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 on a health insurer, not a finding, and there is no verified property and casualty equivalent.
A statute can require a signature. It cannot require attention. What distinguishes a considered denial from a ratified one, in a market conduct exam or a bad-faith deposition, is the record of what was checked before the signature went on.
Florida already wrote the version with an evidence standard
Answer
What did Florida's AI claim denial bill require?
CS/CS/HB 527 let AI assist in claims handling and generate approve or deny recommendations, but barred it from being the sole basis for a denial. A qualified human professional had to analyse the claim independently, review the accuracy of the AI output, document the basis, and affirm in the denial letter that AI was not the sole basis.
The House staff analysis of 2 March 2026 adds that the limit also covered denying a portion of a claim or reducing a payment, and that the retained record had to carry a unique identifier for the individual who made the decision. The table sets the two bills side by side.
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. On 13 March 2026 it died in Senate Rules.
Florida's evidence standard won every House vote, then died in Senate Rules
The clause A5494 leaves out has already been drafted and voted through a chamber without a single vote against it.
- Subcommittee18-0Insurance and Banking Subcommittee, 9 December 2025
- Committee26-0Commerce Committee, 26 February 2026
- House floor108-0Passed the House, 5 March 2026
- Senate RulesDied13 March 2026, before its proposed 1 July 2026 effective date
The reading that matters for New Jersey is that 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 and could be lifted almost verbatim.
The rest of the map: what other states actually regulate
Answer
Does any state require a human to decide property and casualty claim denials?
Not by statute in force as of September 2026. The enacted US rules sit in health lines and turn on medical necessity: California SB 1120, Texas SB 815 and Arizona's A.R.S. 20-3103. New York's DFS circular covers underwriting and pricing only, by its own terms. The two P&C decision-rights bills are New Jersey's, in committee, and Florida's, dead.
Washington is the closest thing in force to a property and casualty evidence standard. Under rulemaking R 2025-05, effective 18 October 2026, "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." Nothing in the rule asks who signed. The table lays out the rest.
New Jersey already regulates AI in claims, and has since February 2025
Answer
Does New Jersey regulate insurers' use of AI today?
Yes. Department of Banking and Insurance Bulletin 25-03, issued 11 February 2025, applies to all insurers authorized or admitted in New Jersey and expects a written AI Systems Program covering the insurance life cycle, including claim administration and payment and fraud detection. It also lists what the Department may demand in a market conduct action.
Bulletin 25-03 is New Jersey's adoption of the NAIC model bulletin, and it does the thing A5494 does not: it describes documentation.
The bulletin's central move is one sentence long: the Unfair Claims Settlement Practices Act standards for the investigation and disposition of claims apply regardless of the methods the insurer used. An AI-assisted denial is judged by the same standard as a manual one and carries the same exposure when the file is thin.
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 and documentation of compliance with it.
- Model inventories and descriptions of the AI systems in use.
- Data source, provenance, lineage, quality, bias analysis and currency.
- Validation, testing and auditing documentation, including model drift.
- Third-party due diligence records, vendor contracts and audit reports.
The bulletin also expects contract terms giving the insurer audit rights over AI vendors. Ask for the audit clause and the model documentation before the pilot, not during the exam. New Jersey's regulator asks what the insurer can show. The bill asks who signed.
Regulators do not need an AI statute to find a denial pattern either. Under the NAIC Market Conduct Annual Statement ratios for 2025, the bill's three lines are lines where states already collect closed-without-payment behaviour annually, carrier by carrier. When a ratio moves away from the peer group, the carrier gets asked for the files.
AI is already in those lines. The NAIC reports that 88% of 193 auto respondents and 70% of 194 homeowners respondents use, plan to use, or are exploring AI and machine learning.
What to put in the file, whatever happens to the bill
Answer
What belongs in the claim file behind an AI-assisted denial?
The record that makes the decision reconstructable: what was checked, what each source returned, what came back clean, when, which policy language the coverage position rests on, what the model contributed, and who decided. Every instrument in force asks for the record, not the signature, and so does every bad-faith deposition.
Four instruments, three in force and one dead, converge on the same requirement, and none is satisfied by a signature. Bulletin 25-03, Washington's amended rule, the Unfair Claims Settlement Practices Act and Florida's bill all point at a reconstructable record.
That record has two audiences. 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. That overlap 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 carries six things:
- The policy language the coverage position rests on, quoted and cited to the form and the endorsement.
- Every source checked, including the ones that came back clean, with the date and time of each check.
- The AI output that informed the decision, its inputs and its version.
- The reviewing human's identity, role and decision time, plus what they changed.
- The factual and legal basis for the denial in the words that went to the policyholder, matched to the evidence in the file.
- The retention posture for the file and the model record, set by the claim-file retention rule in each state you write.
Almost no claims-automation vendor publicly claims to deny claims autonomously, so 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.
Hesper sits on the second question. Agents take a claim from first notice to final recovery, and the evidence is a by-product of doing the work. The investigation runs its phases in parallel, in minutes rather than weeks, and writes down what each source returned, including the clean ones.
A signature satisfies A5494. The record is what an examiner asks for.
What the agent assembles before a denial, and the calls that stay with the adjuster in every configuration.
- 01Quote and cite the policy language the position rests on
- 02Log every source checked, clean ones included, with the time
- 03Keep the recommendation with the inputs it ran on
- 04Draft the factual basis, each element tied to file evidence
- The final decision to deny
- Identity, role and decision time on record
- Any change to the recommendation, and why
- The basis that goes to the policyholder
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. That is a design choice about decision rights, not a compliance claim. The decision-rights question is mapped in which claims decisions stay human.
The stage-level detail sits on coverage verification, claim file review and fraud investigation, with the whole path on the platform page.
A5494 is one clause short of doing what its sponsor's statement says it does. If it advances, the clause is available: Florida drafted it. If it does not, 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.
Key takeaways
- A5494's entire prohibition is one sentence barring AI from making the final decision to deny an automobile, homeowners or flood claim.
- The bill specifies nothing about what the reviewing human must see, disclose, document or retain, so on its face a name and a timestamp satisfy it.
- Triage, coverage analysis, estimating, fraud scoring, denial recommendations and approvals survive by omission, not through an express carve-out.
- Florida's CS/CS/HB 527 wrote the evidence standard New Jersey left out, and it passed the House 108-0 before dying in Senate Rules.
- New Jersey's Bulletin 25-03 has expected a documented AI Systems Program covering claim administration and fraud detection since 11 February 2025, so the work worth doing now is the record behind the denial.