The State of Claims Automation in 2026

Five findings, each with its number and its source.

1. 01

   Handling cost per dollar of loss sits near a decade low, and the fall is arithmetic rather than efficiency.

   US P&C insurers incurred 15.6 cents of loss adjustment expense per dollar of net losses incurred in 2025, against 19.1 cents in 2016. Over that decade net losses incurred rose 70.7 percent while loss adjustment expense rose 39.1 percent, so the ratio fell because the denominator grew, not because a file got cheaper to work. The ratio bottomed at 14.9 cents in 2023 and has risen in each of the two years since. Source: Hesper AI calculation from NAIC, U.S. Property & Casualty and Title Insurance Industries - 2025 Full Year Results (2026)

2. 02

   The people who handle claims are projected to decline, and the projection names automation.

   The Bureau of Labor Statistics counted 389,700 claims adjusters, appraisers, examiners and investigators in 2025 and projects a 6 percent decline, 21,800 positions, by 2035. Its stated reason includes that computer software can evaluate photographs of damaged property and calculate an estimated claim amount. Replacement demand still generates about 21,600 openings a year, so the constraint is not hiring volume but experience: 25.9 percent of the occupation was 55 or older in 2025 against 1.8 percent aged 20 to 24. Source: BLS, Occupational Outlook Handbook and Current Population Survey Table 11b (2026)

3. 03

   Cycle time improved for the second year and remains the largest single lever on customer outcome.

   Average time from first notice of loss to final payment on a US homeowners claim was 40.7 days in the 2026 J.D. Power study, down 3.4 days, with repairs at 29.6 days. In the prior year's edition, overall satisfaction averaged 762 on a 1,000-point scale where the claim finished within 10 days and 595 where repairs ran past 31 days - a 167-point gap. Source: J.D. Power, 2026 and 2025 U.S. Property Claims Satisfaction Studies

4. 04

   The industry cannot size its own leakage to within an order of magnitude.

   IRMI defines claims leakage and publishes no percentage. Conventional industry wisdom puts it at 2 to 4 percent; EY reports 7 to 14 percent of total spend from its own claims quality assessments; The Lab Consulting says it routinely documents 20 to 30 percent and more. Applied to the $13,349 average US homeowners claim closed with payment in 2024, that range is $267 to $4,005 per file - a factor of 15. Source: Hesper AI calculation from EY (2025), Insurance Thought Leadership (2020) and NAIC/CIPR homeowners market dynamics report (2026)

5. 05

   Automation is real at claim assignment and absent at the money decision.

   In the NAIC's private passenger auto AI/ML survey, 135 of 193 large auto insurers already used AI/ML in claims, more than in any other function. But of the classified models, 106 for claim assignment ran with no human intervention, while models that determine a settlement amount were 94 augmentation to 30 automation. Nine companies reported a claim approval model. Zero reported a claim denial model in any status - in use, prototype, proof of concept or research. Source: NAIC, Private Passenger Auto Artificial Intelligence/Machine Learning Survey Results (2022; responses collected October 2021)

Six published figures, each with its publisher, its vintage and the caveat that belongs with it.

Table 1. Sources in full in the methodology. The NAIC bulletin count is taken directly from the adopted-states reference list on the NAIC implementation map, which carries the header "Status as of April 1, 2026".

Loss adjustment expense is the only industry-wide, regulator-filed measure of what claims handling costs. In 2025 it was $86.003 billion, set against $551.755 billion of net losses incurred and $958.726 billion of net premiums earned.

The denominator trap, first

The NAIC prints a 66.5 percent net loss ratio and a 25.8 percent expense ratio for 2025. Loss adjustment expense sits inside the first, not the second: (551,755 + 86,003) / 958,726 = 66.5 percent, which reconciles exactly to the printed figure. Quoting "loss adjustment expense is about 9 percent of premium" alongside "the expense ratio is 25.8 percent" as though the two are additive double-counts the entire claims function. This is the most common arithmetic error in the category.

Table 2. First two columns as published by the NAIC. Last two columns are Hesper AI calculations from them. 2025 is the only year in this ten-year series in which net losses and loss adjustment expense fell as a combined total, by 1.4 percent. Losses did the falling: net losses incurred dropped 1.7 percent while loss adjustment expense still rose. That is a catastrophe-mix effect, with no US hurricane landfall, rather than a structural improvement in claims cost.

The spread by line is the real story

An all-lines average of 15.4 cents per dollar of loss conceals a range that runs from under 4 cents to over 50. The NAIC's by-line profitability report publishes loss adjustment expense and losses incurred each as a percent of net premiums earned. Dividing one by the other gives handling cost per dollar of loss, which the report does not print.

Table 3. First three columns as published by the NAIC on an Insurance Expense Exhibit basis, which the report itself describes as producing "imperfect approximations". Final column is a Hesper AI calculation. Ordered by handling cost per dollar of loss.

What the money is actually spent on

Roughly four fifths of loss adjustment expense is people. In 2019, the last year the NAIC published the industry-wide expense exhibit, $33.1 billion of $69.5 billion - the exhibit's own loss adjustment expense total, a little above the $69.242 billion in Table 2, which comes from a different NAIC tabulation - was net claim adjustment services bought from outside the company, and a further $22.8 billion was claims staff salaries and payroll taxes - 80.5 percent combined. Rent, computing equipment and software, travel, and legal and audit made up the remainder. The NAIC lists that publication series as discontinued, so 2019 is the most recent public industry-wide breakdown that exists. Treat the composition as structural rather than current.

Nearly half of US loss adjustment expense, on the last year it was published, was claims work bought from somebody else. That is the shape of a function that has already been outsourced once.

Litigation and duration are the two multipliers

Two state datasets let the abstraction be priced. Texas statutory Page 14 shows $2.700 billion of direct defense and cost containment expense against $57.763 billion of direct losses incurred in 2024 - 4.7 cents per dollar of loss in litigation and defense cost alone, before a single adjusting hour is counted, because Page 14 excludes adjusting and other expense by design. Within Texas the same measure runs from 0.43 percent in private passenger auto physical damage to 41.65 percent in products liability occurrence.

Florida measures the other multiplier. On claims closed in calendar year 2024, average loss adjustment expense paid was $12,701 on a litigated property claim against $1,778 on a non-litigated one - 7.1 times. And for non-litigated hurricane claims, average loss adjustment expense paid rose from $1,713 on claims closed in under 61 days to $11,652 on claims that took more than a year to close, a factor of 6.8. That is correlation rather than clean causation: complex and disputed claims both take longer and cost more to handle. But the direction is not in doubt, and duration is the variable a claims organization can actually act on.

No US regulator or statistics agency publishes a national average days-from-notice-to-closure across all property and casualty lines. Every credible figure is line-specific, survey-based or vendor-network-based, with a different denominator. Presenting a cross-line national average as though a regulator measured it is a fabrication, so this section presents the measures that do exist, with their denominators attached.

Table 4. Note the average-versus-median gap in the Florida rows. A 47-day average time to report against a 6-day median means the reporting distribution has a long tail, and any operating target built on the average will be built on the tail.

The link to satisfaction is a dose-response curve

J.D. Power's 2025 property study carries the clearest published relationship between cycle time and customer outcome in US claims: overall satisfaction averaged 762 on a 1,000-point scale where the claim was completed within 10 days and 595 where repairs ran beyond 31 days, a 167-point gap. It is correlational - severe and complex claims both take longer and score worse for reasons other than duration. Communication moved the score further still: 777 where customers called it very easy to communicate with their insurer against 337 where they called it difficult, with 82 percent reporting they often interacted through channels they did not prefer. The named failure points are unglamorous: repeated voicemails, repeated calls asking the same questions, missing follow-up emails and texts.

The link to cost is arithmetic

Duration is not only a satisfaction variable. The Florida duration curve in Part 1 prices it: non-litigated hurricane claims cost $1,713 in loss adjustment expense when closed inside 61 days and $11,652 when open past a year. A claims organization that shortens the tail of its duration distribution is reducing handling cost and raising satisfaction with the same action. That is unusual, and it is the reason cycle time keeps recurring as the target metric in this category.

Closing fast is not the same as closing well

Fourteen months after Hurricane Milton made landfall, Florida insurers had closed 92.0 percent of the 385,146 claims reported. But 134,177 of those, about 34.8 percent of all claims reported, closed without any payment. Of those, 51,625 were damage below deductible and 5,998 were flood exclusions, alongside 44,750 closed for administrative reasons. Closed without payment is not the same as denied, and the administrative category is where a claims organization should look first when it wants to know what its closure rate is actually measuring. Florida OIR states the data is compiled from insurer submissions and has not been audited.

What complexity is doing to the auto cycle

The front end of the auto cycle has compressed sharply: CCC puts last-estimate-sent to vehicle-in at about 2.2 weeks, down from almost 30 days at the Q4 2022 peak. The repair itself has moved the other way, on a separate measure from J.D. Power rather than CCC: in the 2025 auto claims study, vehicles from model year 2015 and older with no advanced driver assistance features averaged 17.9 days of total cycle time, against 21.5 days for model year 2019 and newer vehicles with three or more such features affected - a comparison that confounds vehicle age with feature count, so the 3.6-day spread is not a clean effect. Back on CCC data, in Q1 2025 86.9 percent of repairable direct-repair-program appraisals included a diagnostic scan and 32.2 percent included a calibration, up from 23.9 percent a year earlier. CCC is a claims software vendor publishing analysis of its own platform data, which makes it primary for its network and not an independent measure.

The Bureau of Labor Statistics projects the claims occupation to shrink. It is the only federal projection that bears directly on this question, and in its August 2026 vintage it is explicit about why.

Table 5. Denominators do not interchange. The Occupational Outlook Handbook's widely quoted 21,600 annual openings covers the combined group; the 20,900 above is SOC 13-1031 alone. The May 2025 Occupational Employment and Wage Statistics survey counts 335,790 in the combined group because it is an establishment survey and excludes the self-employed. All three counts are correct for their own frame.

BLS attributes the decline to automation in plain terms, writing that computer software can evaluate photographs of damaged property and calculate an estimated claim amount. Note the vintage: this is the 2025-35 round published in August 2026. The prior 2024-34 round showed -5 percent from a 365,300 base and is still circulating in trade coverage as though current.

The age distribution is the harder constraint

Headcount is replaceable; adjudication judgment developed over a career is not. In 2025, 19.0 percent of the occupation was aged 55 to 64 and a further 6.9 percent was 65 or older - 25.9 percent aged 55 or over - against 1.8 percent aged 20 to 24. Median age was 43.4 years. BLS notes that its 2025 annual estimates are 11-month averages excluding October because of the federal government shutdown.

Widely circulated claims that the industry will lose about 400,000 workers to attrition by 2026, usually attributed to BLS, could not be traced to any BLS publication and are excluded from this report. The well-defined occupational age split above is the defensible version of the same argument.

Caseload, and what frontline time is spent on

The best public caseload measurement is workers' compensation specific. In Rising Medical Solutions' 2023 benchmarking study of 1,388 frontline claims professionals, 53 percent reported average indemnity caseloads of 125 or fewer, meaning 47 percent carried more, and 26 percent carried 151 or more. Among third-party administrator respondents, 69 percent carried 126 or more, though that segment base is 137 respondents and its percentages are indicative only. The study cites a two-year case study setting a maximum of 111 claims per lost-time claims professional for best practices to be executed effectively. That is a citation of Kern (2019), not an original finding of the survey.

What those professionals spend time on is the more actionable finding. Forty-six percent reported spending 30 to 40 percent or more of their time on administrative tasks - letters, data collection, internal claims-system requirements - and 23 percent spent that share on regulatory compliance activities. Fifty-five percent said they use five or more separate systems in daily claims management, up from 42 percent in the same study's 2019 wave. Asked which technology or AI capability would help most, 35 percent chose automation of administrative tasks such as form filing or other regulatory requirements, more than double the 17 percent who chose automation of claims tasks such as indemnity payments or bill pay.

The frontline's own first choice for AI is not the adjudication. It is the paperwork around the adjudication - by a margin of more than two to one.

The labor market itself is stable rather than tight. Jacobson Group and Aon's Q3 2026 study put six-month voluntary turnover at 5.3 percent against a 12-month average of 7.6 percent, with 89 percent of respondents intending to increase or maintain staff. In the Q1 2026 edition, 73 percent said they were most likely to hire experienced talent against 25 percent prioritizing entry-level, and claims ranked second in entry-level demand at 33 percent. That is a market buying experience it will increasingly struggle to find, from an occupation whose 20-to-24 cohort is 1.8 percent of its headcount.

This is the section of the report with the weakest evidence base, and saying so is more useful than picking a number. There is no agreed measurement standard for claims leakage, no regulator publishes a figure, and the published estimates differ by a factor of ten.

Leakage: a contested range, not a number

Table 6. The spread is the honest finding. Note that the denominators are not the same: EY's "total spend" is not defined in the source and it is not clear whether it means indemnity alone or indemnity plus loss adjustment expense. Do not average these.

Recovery: the one place with real regulator-filed data

Salvage and subrogation is the exception. A peer-reviewed article in the NAIC's Journal of Insurance Regulation aggregated NAIC annual statement Schedule P data across 1996 to 2021 and produced measured ratios. The headline: in 2021, insurers recovered nearly $51.6 billion across auto physical damage, commercial auto liability and personal auto liability combined. The dollar figure is the authors' own aggregation, not an NAIC-published statistic, and it covers three lines rather than all of P&C.

Table 7. Source: Bisco and Fier, "How's the Recovery? Salvage and Subrogation in the Property Liability Insurance Industry", NAIC Journal of Insurance Regulation, 2023. Continuous variables winsorized at the 1st and 99th percentiles. These are 26-year firm-year averages, not current-year industry ratios.

For scale on the infrastructure: Arbitration Forums, Inc., the largest US intercompany arbitration and subrogation services provider, reports members filing 1.1 million arbitration disputes and 2.3 million subrogation demands a year, worth almost $27 billion in claims, across a membership of more than 5,000 companies. That is self-reported by the operator of the forum, with no audited annual report, and the $27 billion is the amount in dispute filed rather than the amount recovered.

Fraud that gets paid, with the denominators made explicit

The industry's headline fraud figure is $308.6 billion a year, from the Coalition Against Insurance Fraud's 2022 study conducted by the Colorado State University Global White Collar Crime Task Force. Three things about it matter before it is quoted at a property and casualty audience. First, it is a meta-analysis: the report states plainly that the task force "will not be collecting any 'new' data as part of this study". Second, only $45 billion of the $308.6 billion is property and casualty; roughly 85 percent is health, Medicare and Medicaid, life, disability, premium avoidance, workers' compensation and auto theft. Third, the report's own Section VII concedes that absent provable involvement of the insured, auto theft "is not insurance fraud but an insurance crime", and counts $7.4 billion of it anyway.

The $45 billion property and casualty component is a single multiplication: 2020 industry losses and loss adjustment expense of $450.8 billion, times an assumed 10 percent fraud rate. Where that 10 percent comes from is covered in the provenance section below, and it is the most important sentence in this report.

What is measured rather than estimated

Two datasets in this domain rest on real samples. The Insurance Research Council's fraud and buildup study analyzed more than 35,000 auto injury claims closed with payment, contributed by twelve insurers representing 52 percent of the US private passenger auto market, and found that 21 percent of bodily injury claims and 18 percent of personal injury protection claims showed the appearance of fraud or buildup, adding $5.6 billion to $7.7 billion in excess payments. That is the strongest sampling frame in the literature and the data are from 2012, published in 2015, with no successor edition. It must be presented as 2012 data.

The second is the California Department of Insurance Fraud Division, which publishes a referral-to-case funnel no other state matches. NICB provides the volume signal alongside it: 208,956 questionable claim submissions reviewed in 2025, up from 180,508 in 2024, an increase of about 15.8 percent. A questionable claim is one a member insurer referred because it displayed indicators of possible fraud. It is a referral count from member companies, not an adjudicated finding and not an industry census.

For completeness on the claims-environment cost that is not leakage and not fraud: a Triple-I study, the latest in a series run with the Casualty Actuarial Society, estimates that increasing inflation added between $231.6 billion and $281.2 billion to US liability losses and defense and cost containment expense over 2015 to 2024 across four lines - 14.4 to 17.5 percent of ten-year booked loss and defense cost. Note what it does not say: the study states that its loss-development method "can detect the presence of inflation, but it cannot detect its source", so this is not a measured social-inflation figure. And it measures what the wider environment added, which is a different thing from what better claims handling would have saved. The two should never be summed.

Start with the vintage problem. The only regulator-published measurements of AI adoption in US property and casualty claims are the NAIC's private passenger auto survey, which collected responses in October 2021, and its homeowners survey, fielded across 2022 and 2023. Both predate the generative wave entirely and both explicitly excluded generalized linear models and anything "a company could have realistically utilized in the year 2000 or prior" from their definition of AI/ML. No newer NAIC property and casualty AI survey exists as of September 2026. Every regulator adoption figure available is three to five years old and describes predictive models, not large language models.

Where claims sat in the stack

Among 193 large private passenger auto insurers, 169 reported that they currently use, plan to use, or plan to explore AI/ML somewhere in insurance operations. That 88.6 percent is the NAIC's own printed percentage, though 169 of 193 is 87.6 percent. It is an all-operations figure and is routinely misquoted as "88 percent of auto insurers use AI to evaluate claims", which the report does not say. The claims-specific in-use figure is 135 of 193, or 70 percent - higher than marketing (50 percent), fraud detection (49 percent), rating (27 percent) or underwriting (18 percent). Among 194 large homeowners insurers, 136 used, planned or explored AI/ML, with claims again leading by function at 54 percent.

Stage by stage

The NAIC asked insurers to classify each model by how much of the decision it makes. Its three levels are the most useful vocabulary in this category: automation is no human intervention on execution, augmentation is a model that suggests an answer and advises the human deciding, and support is a model that provides information without suggesting a decision. The table below maps the eight lifecycle stages against what the surveys measured. A stage marked "vendor claim, unmeasured" is one where products are sold and no regulator has counted adoption: the cell says so rather than being filled.

Table 8. Sources: NAIC Private Passenger Auto AI/ML Survey Results (2022, Tables 3, 4, 5, 10 and 11) and NAIC 2022-23 Home AI/ML Survey Analysis memorandum (2023). Company counts and model counts come from different tables with different denominators and are labeled accordingly. Both surveys predate generative AI.

Across 193 of the largest auto insurers in the United States, the number reporting an AI or machine learning model for claim denial - in use, in prototype, in proof of concept, or in research - was zero.

Who builds the models

The homeowners survey counted 2,413 models: 1,407 built internally and 1,006 by third parties, a 58/42 split that the NAIC notes was nearly identical in the auto survey. Ninety-five unique third-party vendors appeared, of which 37 supplied claims models; the auto survey listed 76 unique third-party providers. Within claims specifically the split is directional: models for claim approval, claim assignment, adjuster information and other claim functions tend to be built in-house, while models that determine settlement amounts and evaluate images of the loss tend to be bought. Some models are counted more than once because the same model has separate uses.

Cross-line context

Health insurers show the highest adoption the NAIC has surveyed. Its May 2025 report puts 84 percent of 93 responding companies - 78 of them - as currently using AI/ML, and the NAIC's AI topic page cites 92 percent for the same survey. Both are right: 92 percent is 86 of 93 on the broader "use, plan to use, or plan to explore" question, which is the measure the auto, homeowners and life surveys report, so 92 percent is the figure that compares across lines and 84 percent is the narrower in-use figure. Even there, claims adjudication is one of the least penetrated functions: 31 companies reported it already in production against 43 reporting not applicable. Life insurers report the lowest adoption at 58 percent of 161 respondents, on a broader model definition that included generalized linear models, which makes the gap to the property and casualty figures wider than it looks.

What carriers themselves disclose

The pattern in property and casualty 10-K filings is that AI in claims is described and never measured. Travelers states in its fiscal 2025 Form 10-K that it has invested significant additional resources in many of its claims handling operations, including digital, analytics, artificial intelligence and automation capabilities. No adoption rate, spend, cycle time or claim volume is disclosed. Progressive tells investors it has used machine learning and other forms of AI for many years and flags generative and agentic AI as a distinct and newer risk category - a useful marker that carriers themselves separate legacy predictive models from current generative systems. Neither filing gives a claims-specific AI metric.

The NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers on 4 December 2023. As of the NAIC's implementation map dated 1 April 2026, 25 jurisdictions - 24 states plus the District of Columbia - had adopted it. Four more states run their own insurance-specific AI instrument instead and are counted separately, which is why adoption counts in circulation disagree.

The bulletin reaches claims, explicitly

A bulletin is guidance, not enforceable law in its own right. Its force comes from the statutes it cites: the Unfair Trade Practices Act (Model #880) and the Unfair Claims Settlement Practices Act (Model #900). Section 3 says the insurer's written AI Systems Program should address the use of AI systems across the insurance life cycle "including areas such as product development and design, marketing, use, underwriting, rating and pricing, case management, claim administration and payment, and fraud detection". Claims is named. The bulletin does not prohibit any specific AI claims use and does not mandate human review of an AI-informed claim decision.

What a claims organization has to be able to show

The operative risk concept is the "Adverse Consumer Outcome", defined as a decision by an insurer, subject to regulatory standards enforced by the department, that adversely impacts the consumer in a manner violating those standards. Controls are sized to that risk and to the degree of potential harm. Section 4 then lists what a department may request on examination. That list is the practical specification for any claims organization deploying AI:

• Documentation evidencing the AI Systems Program itself.
• Pre-acquisition and ongoing diligence over third-party AI systems and the data behind them.
• Contractual audit rights over those third parties.
• Documentation of validation, testing and auditing, including evaluation of model drift.

Section 4 closes by saying the bulletin does not prescribe specific documentation requirements; it lists what a department may request. Insurers remain responsible for third-party AI and data under Section 3. Given that 42 percent of the models in the NAIC's homeowners count were third-party built, the diligence and audit-rights expectations are not a corner case.

The instrument that will actually be used on exam

The NAIC AI Systems Evaluation Tool was being piloted by 12 participating states as of March 2026, with adoption anticipated at the 2026 Fall National Meeting. The tool's project background enumerates claims functions in scope: claim assignment, triage and fast-tracking, individual and bulk claim reserving including loss estimation, imaging and video analysis, fraud detection, litigation, estimation of closure rates, and salvage and subrogation. It covers AI used by third-party claim administrators and managing general agents, not only by carriers. Pilot states apply it across a mix of financial and market conduct exams and choose which companies receive tool-related requests. The NAIC has no authority to compel participation.

The timing law the system has to satisfy anyway

Whatever the AI governance position, the deadlines that bind a claims system are the unfair claims settlement practices rules, and they vary on both the number and the unit. The parent act, Model #900 adopted in June 1990, works on a reasonableness standard and carries a single numeric deadline - 15 calendar days to provide the forms needed to present a claim. Every other number lives in the model regulation and in individual states.

Table 9. Model #902 is a 1997 template and states adopt it with variation, so its numbers are not the operative deadline anywhere. California is notably stricter than the model on the decision clock: 40 calendar days from proof of claim against 21 days from proofs of loss. New York runs on business days, which is a different unit entirely. Texas figures are quoted from the Texas Department of Insurance's own consumer page rather than the statute, because the statutory text could not be fetched to resolve business days against calendar days. Florida deadlines are tolled during mediation and where the policyholder does not supply requested material information within 10 days.

One detail in Model #902 is worth flagging for anyone building fraud-referral automation: the regulation relieves the insurer of the 21-day and further-time requirements where there is a reasonable basis, supported by specific information available for regulator review, for suspecting the first-party claimant fraudulently caused or contributed to the loss. The relief is conditioned on documented evidence available for review, not on suspicion. A system that flags a claim without producing the specific information behind the flag does not buy the extension.

What has not happened

Three widespread misconceptions are worth correcting, because each one is used to sell something.

1. New York's AI circular does not reach claims. Insurance Circular Letter No. 7 (2024), issued 11 July 2024, states that it "is not intended to address phases of the insurance product lifecycle other than underwriting and pricing". Claims handling in New York remains governed by Regulation 64.
2. California SB 1120 is a health law. The Physicians Make Decisions Act, Chapter 879 of the Statutes of 2024, requires that an AI tool used in utilization review does not supplant health care provider decision-making. It amends the Health and Safety Code and the Insurance Code provisions covering health care service plans and disability insurers. It does not apply to property and casualty insurers. This is the single most misquoted instrument in AI-claims marketing.
3. No state has enacted a property and casualty human-review requirement. The closest attempt, Florida CS/CS/HB 527 (2026), would have barred AI from being the sole basis for denying or reducing a claim and required qualified human professionals to make denial decisions - but it amended ss. 440.131, 627.4263 and 641.31091, so it reached workers' compensation carriers, health insurers and HMOs, not property and casualty lines generally. It passed the Florida House 108-0 on 5 March 2026 and died in Senate Rules on 13 March 2026. The entire 2025 wave of enacted AI-in-claims legislation - Arizona, Illinois, Maryland, Nebraska and Texas - was health utilization review. Texas SB 815 bars a utilization review agent from using an automated decision system to make an adverse determination while expressly permitting AI for administrative support and fraud detection.

Colorado is the one state whose statute reaches claims on its face. C.R.S. 10-3-1104.9 defines an insurance practice to include "claims management in the transaction of insurance". Its implementing regulation, 3 CCR 702-10, has so far been extended only to life, private passenger automobile and health benefit plan insurers, effective 15 October 2025, with auto and health compliance reports due 1 July 2026 and annually thereafter. The statutory reach into claims is not yet matched by a claims-specific rule.

Finally, the whole state-level structure sits under an unresolved federal preemption question. The Executive Order "Ensuring a National Policy Framework for Artificial Intelligence", signed 11 December 2025, directs the Attorney General to establish an AI Litigation Task Force to challenge state AI laws in federal court and directs a Commerce Department assessment of state AI laws. The order makes no mention of insurance or of the McCarran-Ferguson Act. Silence is not a carve-out in either direction, and as of September 2026 no court has struck any state insurance AI instrument.

Seven metrics a claims leader can actually run, each built from data a carrier already has, each with a public benchmark from this report to compare against.

Table 10. The common property of these seven is that none requires a survey, a vendor, or a number the industry has not measured. Six come straight out of a claims data warehouse; the seventh comes out of a model inventory.

Several of the figures that appear most often in claims automation material are not measurements. Tracing them is not a rhetorical exercise: a business case built on a number with no methodology behind it will not survive the first serious question from a finance function.

The one that matters most: the 10 percent fraud rate

The statement that fraud accounts for about 10 percent of property and casualty claims, or of incurred losses and loss adjustment expense, is the load-bearing assumption under the industry's entire published fraud loss estimate. Its origin is stated inside the Coalition Against Insurance Fraud's 2022 study, which quotes the Insurance Information Institute verbatim:

"In the late 1980s, the Insurance Information Institute interviewed claims adjusters and concluded that fraud accounted for about 10 percent of the property/casualty (PC) insurance industry's incurred losses and loss adjustment expenses each year."Insurance Information Institute, as quoted in The Impact of Insurance Fraud on the U.S. Economy, Coalition Against Insurance Fraud and Colorado State University Global White Collar Crime Task Force, 2022, Section V(b).

Three observations. First, the vintage: this traces to interviews conducted in the late 1980s, roughly thirty-five years before the study that carries it forward. Second, the method: the Coalition's own report states in the immediately following footnote that "The III did not specify exactly how they derived at this figure", and then reverse-engineers what it believes the III did. Third, the recursion: having flagged that the derivation is unknown, the study adopts the same 10 percent as its own multiplier, applying it to 2020 industry losses and loss adjustment expense of $450.8 billion to produce the $45 billion property and casualty figure that is then quoted as current.

The Insurance Information Institute page that carried that sentence now returns a 404, and its replacement article no longer contains it, so the verbatim quotation inside the Coalition's PDF is currently the most durable citable instance of it. The figure is also not nothing: it has been carried by the industry for decades and the Coalition defends it as consistent with its authors' unpublished consulting work. But it is a thirty-five-year-old qualitative estimate from adjuster interviews, not a measurement, and it should be written that way every time it is used.

The rest of the chain

Table 11. Every entry in this table was traced by following the citation chain to its terminus. Where the chain terminates at a trade article, a removed page or a firm selling the remedy, that is stated.

What this report is. A synthesis and analysis of publicly available data. Hesper AI has run no survey, commissioned no panel, and collected no data of its own. Nothing in this report is Hesper's own research data. Its originality lies in assembling figures that live across more than twenty separate publications into one frame, computing measures those publishers do not print, and stating which widely quoted numbers do and do not have methodology behind them.

Source selection. Preference order was: US regulator filings and federal statistics agencies first (NAIC, BLS, state departments of insurance); then peer-reviewed or named studies with a disclosed sample; then named industry studies and vendor platform data, labeled as such. Every figure in the body carries its publisher. Where only a vendor or consultancy figure exists, the text says so at the point of use.

Derived measures. Seven calculations in this report are ours, marked D1 through D7 with their arithmetic and inputs shown: loss adjustment expense per dollar of loss over ten years (D1); the same ratio by line for 2024 (D2); Florida handling cost per dollar paid and its litigation and duration multipliers (D3); loss adjustment expense per employed adjuster and the 2035 workforce arithmetic (D4); the leakage range priced per homeowners file (D5); the industry fraud multiplier rebuilt on 2025 data (D6); and California's referral-to-case funnel across three programs (D7). None is published by the sources they draw on.

Limitations. Statutory loss adjustment expense excludes the portion of claims-handling cost carriers classify as other underwriting expense, so every cost figure here is a floor. Calendar-year ratios mix accident years, so a single line in a single year can swing on reserve movements; cross-line ordering is the durable signal. State datasets - Florida, Texas, California - are used where they are the best available measurement and are never presented as nationally representative. The NAIC AI surveys are three to five years old and describe pre-generative predictive models. Where a figure could not be verified against a primary source in this research, it was excluded rather than carried.

Figures deliberately excluded for weak sourcing

• "FBI estimates non-health insurance fraud exceeds $40 billion a year" and the associated "$400 to $700 per family". The fbi.gov page carrying these is no longer served at its address; the archived version carried no date, study or methodology.
• "The FBI estimates $4,000 to $7,000 per family over ten years", as carried on the NAIC's insurance fraud topic page. Its "FBI estimates" hyperlink resolves to an NICB blog post rather than any FBI publication.
• "$15 billion in missed subrogation annually" and "15 percent of files close with a missed subrogation opportunity". Traced to a 2021 trade article; no study, sample or methodology published.
• "$29 billion in annual auto premium leakage" with its component breakdown. A 2017 vendor estimate reachable only through republication, and it measures underwriting premium leakage rather than claims leakage.
• "About 75 percent of flagged claims are never fully investigated". No primary source located in this research.
• "70 to 90 percent straight-through processing at leading carriers" and "65 percent industry-wide by 2026". No regulator, agency or named study located.
• "The US insurance industry will lose about 400,000 workers to attrition by 2026", usually attributed to BLS. No BLS publication carries it, and its horizon has already passed.
• "1.37 million insurance professionals are age 55 or older" and the related "6-to-1 ratio of retirement-age employees to young entrants". No traceable primary source; the BLS CPS occupational age split is used instead.
• A single national average days from first notice of loss to closure across all US P&C lines. No regulator or statistics agency publishes one.
• A current national split of loss adjustment expense into defense and cost containment versus adjusting and other. The NAIC series carrying it ends at 2019 data and is listed as discontinued.
• A national average dollar cost to handle one P&C claim. Industry-wide claim counts are not published on a comparable basis, so the denominator does not exist.
• A caseload-per-SIU-investigator figure. No primary source located. The workers' compensation caseload data in Part 3 is a different and narrower measure and is labeled as such.
• A leakage percentage attributed to Deloitte or McKinsey. Neither firm publishes one in any fetchable document; the attributions in circulation are unsupported.
• "6 percent of claim payments, about $67 billion a year" in leakage. The article it is usually attributed to does not contain it.

Sources

Publisher, title, year, URL. All URLs were fetched during research for this report in September 2026.

Hesper AI is an AI claims resolution platform for property and casualty insurers, third-party administrators and managing general agents. It works a claim from first notice through to final recovery, with investigation-grade evidence behind every decision, and it is built so that the evidence record behind a decision can be produced on examination. Fraud detection is built in rather than bolted on downstream.

We compiled this report because the public evidence base for claims automation is scattered across regulator filings, federal statistics and a set of trade figures whose provenance is rarely checked. Redistribution is permitted with attribution.

© 2026 Hesper AI. Published 23 September 2026. A synthesis of public data; no proprietary survey. All sources cited with URLs. Redistribution permitted with attribution.