Straight-through claims processing: machines may pay, only people deny

Straight-through claims processing in the United States is a one-way door. Nothing in US insurance law stops a carrier paying a claim without a person reading it, and almost every binding constraint lands on refusing one. Texas Insurance Code Chapter 542 pays a carrier for speed and fines it for delay: notice of acceptance or rejection no later than the 15th business day after receiving everything needed for proof of loss, payment no later than the fifth business day after acceptance, and 18% per annum in damages plus the insured's attorney's fees when the deadline slips. Washington, eight days before this was published, made it an unfair claims settlement practice to refuse a claim on a database check alone. That rule governs denial, not approval.

So the useful question is not what share of claims could close untouched. It is which claims have passed enough checks that nobody needs to read them. Our glossary already holds the definition of what straight-through processing means and the claim types that tend to qualify, so this piece goes past it: what is measured today, the regulator-filed proxy that tells you which lines can support no-touch closure, the ten gates a claim has to clear, the counter-metrics that catch a gamed rate, and why the queue left behind gets harder with every point of coverage you win.

Speed and leakage are the same problem. The organisation under cycle-time pressure and the organisation under leakage pressure are the same organisation, and straight-through processing is where the two meet, because every claim routed around a person is a claim nobody read. That makes gate design the whole subject. Straight-through processing is one routing decision inside the eight stages of the claims lifecycle, and it is only ever as good as the checks standing in front of it.

Start with the only straight-through number in US insurance that sits inside an audited filing. In its FY2025 Form 10-K, filed with the SEC on February 25, 2026, Lemonade states that "as of December 31, 2025, 96% of the time, it is AI Jim that will take the first notice of loss from a Lemonade customer without human intervention," and that "as of December 31, 2025, roughly 55% of our claims were automated, resulting in instant or near-instant processing from start to finish." Read the caveats with the number. Lemonade is a digital-native personal-lines carrier writing renters, homeowners, pet and car; automated is its own definition; and no regulator audits the denominator. It is still better evidence than anything else published on this subject, because it is a company statement inside a securities filing rather than a marketing page.

The only regulator-measured automation data in US property and casualty claims is five years old and counts models, not claims. The NAIC's private passenger auto AI/ML survey, a staff report dated December 8, 2022 built on responses collected in October 2021, covered 193 large auto insurers. 135 of them, 70%, already used AI or machine learning in claims, more than in any other insurance function. Of the classified claim-assignment models, 106 ran with no human intervention, against 81 classed as augmentation and 8 as support. Of the models that set a settlement amount, the split inverts: 94 augmentation against 30 automation. 9 companies reported a claim-approval model in use, and zero reported a claim-denial model in any status, including research and prototype. Of the fraud-referral models, 0 ran without human intervention, against 89 augmentation and 93 support. Two caveats matter: the survey predates generative AI entirely, and it excluded anything a company could have realistically used in the year 2000 or earlier.

Then there is the forecast nobody went back to score. In the 2021 edition of LexisNexis Risk Solutions' Future of Claims study, reported by Repairer Driven News, 6 of 25 leading auto insurers were using or piloting touchless claims, five of them top-10 carriers, where touchless meant technology handling the entire claims process including damage assessment, with automatic approval and payment. LexisNexis vice president Bill Brower said the industry expected 80% of claims to be virtual by 2025, with up to 50% touchless. In the same coverage, Ramon Lopez of USAA said touchless worked best for "low-complexity losses" and that the solution was "not full-turnkey just yet." Virtual claims handling itself went from under 15% before the pandemic to near 100% during lockdowns, then settled near 60% by May 2021. No regulator and no named study has published a measured industry touchless rate since, which means that 50% forecast for 2025 has never been checked against anything.

Vendor disclosure fills the gap with scale figures instead of qualifying shares. CCC's Estimate-STP announcement reports 75 or more auto insurers actively applying CCC advanced AI, around 16 million claims processed annually, 30,000 customers, and more than 5 million unique claims processed with that AI through 2020, with insurer adoption up 50% year over year in 2020. The line that travelled was that "with the rollout of CCC Estimate - STP, we expect the industry will be able to realize automated estimating in mere minutes." What the page never says is what share of claims qualify. It says qualified repairable claims and leaves the share undefined, and the vintage is 2021.

This is why our own state of claims automation research carries a box titled "Excluded: straight-through processing rates." Figures such as 70% to 90% straight-through processing at leading carriers, or 65% industry-wide by 2026, circulate widely, and no regulator, government statistics agency or named published study behind them could be located, so they were excluded outright rather than carried with a caveat. What is measured is the clock. The J.D. Power 2026 U.S. Property Claims Satisfaction Study puts the average US homeowners claim at 40.7 days from first notice of loss to final payment, down 3.4 days year over year, with repairs completing in 29.6 days, from 5,093 claimants surveyed between December 2024 and December 2025. Those are survey-reported, homeowners-only figures, and the two durations are not additive.

Since nobody publishes straight-through rates by line, substitute a number every carrier already files and every regulator already holds. Loss adjustment expense divided by losses incurred gives handling cost per dollar of loss, and it reads as a measure of how much judgment a line demands. The NAIC does not print the ratio; it prints the two rows it is built from, each as a percent of net premiums earned, in its Report on Profitability by Line by State. Dividing one by the other is arithmetic, and it is available for every line, every year, for the whole market. That makes it the most defensible substitute available for the rate the industry will not publish.

Read the ordering and the qualification question starts to answer itself. The lines at the top are the ones where liability has to be argued: products liability at 51.1 cents, medical professional liability at 38.4 cents, workers' compensation at 31.0 cents, where compensability, duration and medical necessity are each a judgment. The lines at the bottom are the ones where a damaged object has a price: private passenger auto physical damage at 11.1 cents, pet insurance at 6.1 cents, where the claim is usually an invoice from a licensed provider with a number on it. Homeowners sits between them at 12.8 cents, pulled up by catastrophe and roof files. None of these is a straight-through rate and none should be presented as one. They are a filed, peer-benchmarked proxy for how much of a line a fixed set of checks can safely clear.

The all-lines trend needs one sentence of caution before anybody reads it as progress. US property and casualty insurers incurred 15.6 cents of loss adjustment expense per dollar of net losses incurred in 2025, against 19.1 cents in 2016, and the fall is denominator-driven: across that decade net losses incurred rose 70.7% while loss adjustment expense rose 39.1%. A falling ratio there is not evidence that a claim got cheaper to work. It is evidence that claims got more expensive to pay, faster than they got more expensive to handle.

Even the most automatable line is less uniform than the straight-through narrative implies. CCC's Crash Course 2026 reports total loss frequency at 23.1% of claims, an industry high, and 28.3% of repairable estimates including calibrations, leaving a mix of 55.1% straightforward repairs, 21.8% with calibrations and 23.1% total losses. That is vendor-published analysis of CCC's own network rather than an independent industry measure, and it is the best available shape argument: in the single biggest candidate line for no-touch closure, roughly a quarter of claims are total losses and more than a quarter of repairable estimates carry a calibration step. Neither is a clean straight-through shape.

The honest column in that table is the last one. Three rows are primary evidence, one is a named study that is now five years old, four are a derived regulatory proxy, and one is an admission that the number does not exist. Every other page that ranks for this subject fills the same table with its own customer results. A filed proxy with a stated derivation beats a vendor average with an undisclosed denominator, even when the proxy measures something adjacent to the question.

Write the gates down and programme design stops being a debate about claim types. A claim is eligible for no-touch closure when it has passed every check a person would have performed, and the list is longer than any vendor page admits. Five of the gates sit in front of the payment decision and five sit behind it, and the asymmetry at the top of this piece sets the shape: a claim that clears every gate can be paid, while a claim that fails one has to reach a human, because refusal is the side of the decision the rules actually constrain.

Gate one is coverage on the face of the policy, and it is where money leaks quietly. EY's work on indemnity in litigated claims attributes more than 85% of the leakage it found to three causes, one of which is coverage determination. The volume of below-deductible and excluded losses is not theoretical either: Florida insurers closed 51,625 Hurricane Milton claims as damage below the deductible alone. Gate two is the duplicate and prior-claim match, and it is the cheapest gate to run because the infrastructure already exists. Verisk's ISO ClaimSearch holds 1.8 billion records covering roughly 95% of the US property and casualty market, ingests 175,000 records a day and serves 200,000 daily users, and most carriers are already contributing to it.

Gates three and four get harder every year, which is the part most qualification lists have not caught up with. The National Insurance Crime Bureau said on September 2, 2025 that it had analysed questionable claims submitted from 2022 through June 30, 2025 and projected a 49% year-over-year rise in insurance fraud linked to identity theft for 2025, with nearly a quarter of those cases involving synthetic identities built from a mix of real and fabricated data. NICB president and chief executive David J. Glawe put it plainly: "Identity theft and the use of synthetic identities are the foundation for life insurance, medical-related fraud and cargo theft." On documents, Allianz UK disclosed on June 5, 2026 that it had handled over 34,200 insurance fraud cases in the prior year worth nearly £174m, including a van with an AI-added cracked bumper supported by a false repair invoice of about £1,000, a Mercedes where AI had altered a numberplate on a different vehicle, and an £80,000 escape-of-water claim with fake invoices, caught on inconsistent ripple effects and distorted reflections. Its own survey found 52% of UK drivers aged 18 to 34 have posted car pictures online and 73% were unaware that could expose them to insurance fraud. Matt Crabtree, Allianz UK's head of financial crime intelligence and investigation strategy, said: "Insurance fraud constantly evolves and we're seeing this through the rise of AI and photo editing, which is a current challenge for the industry."

Two gates on that list deserve more attention than they get. Gate five is where the tuning argument lives, because rules-based screens run false positive rates in the 60% to 85% range on a Hesper internal benchmark, and a screen tuned to protect the straight-through rate will quietly pass claims a person would have stopped. The screen itself can be somebody else's: FRISS, Shift Technology and Verisk all sell one, Hesper has built-in detection and runs standalone or beside them, and the gap is never the screen. It is what happens to the claim the screen rejects. Gate nine is the one no competing page mentions at all. California's Department of Insurance Bulletin 2022-5, issued June 30, 2022, covers claims processing and the investigation of suspected fraud, requires that the specific reasons be given to the consumer where a complex algorithm contributes to an adverse action, and names as a live Department concern allegations that insurers unfairly flag claims from certain inner-city ZIP codes and refer them to SIU. A straight-through gate is also a drop-out machine, and the drop-out population is testable.

The common thread across all ten is that each gate is an evidence question rather than a rule. That is why the evidence depth built for investigation work belongs behind the coverage read and the payment check too, and not only behind the fraud flag. A coverage position that cites the endorsement it rests on, a payee check that reconciles against the claimant and the lienholder, and a document check that reads the metadata rather than the layout are the same kind of artifact: a conclusion with its sources attached. Our write-up of how automated claims investigation runs goes through the mechanics of producing that record on a file that fails a gate.

A rate whose numerator is claims that closed without a person rewards closing, and closing is not resolving. The cleanest public demonstration of the difference is a catastrophe book. Florida's Office of Insurance Regulation, in its Hurricane Milton catastrophe claims data as of December 9, 2025, 14 months after landfall, reported that insurers had closed 92.0% of the 385,146 claims reported, and that 134,177 of them, about 34.8% of all claims reported, closed without any payment: 51,625 because damage was below the deductible, 5,998 on flood exclusions, and 44,750 for administrative reasons. FLOIR states that data is unaudited insurer submissions. Its separate Property Insurance Stability Report of January 1, 2026 prices duration. Non-litigated hurricane claims cost $1,713 in loss adjustment expense when closed inside 61 days, against $11,652 when left open past a year, a factor of 6.8, and average loss adjustment expense ran $1,778 on a non-litigated property claim against $12,701 on a litigated one, a factor of 7.1.

The counter-metrics are not a new instrumentation project, because every US carrier already files them. The NAIC's Market Conduct Annual Statement homeowners instructions for 2026 require claims open at the start of the period, claims received during it, claims closed with payment, claims closed without payment, claims reopened during the period, and aging buckets by days to closure. The instructions are specific about the edges: a claim reopened for a subsequent supplemental payment counts as a new claim with separate aging, and a claim reopened solely to refund a deductible is not counted as a paid claim. MCAS Ratio 1 for personal auto and homeowners is claims closed without payment divided by total claims closed. If you are reporting a straight-through rate to a board, those series already exist inside your own regulatory filings, and they are the ones that catch a rate moved by auto-closing rather than by resolution.

Two honest negative findings belong here. No published industry reopen-rate benchmark exists, so there is nothing to compare your own against, and no public US enforcement action against automated claim approval could be found. The absence is the point: the counter-metric is collectable and uncollected, which means the first carrier to publish its own closed-file audit methodology, sample and denominator sets the standard rather than arguing about somebody else's. The leakage band everyone reaches for instead is weak evidence. The published aggregate is 7% to 14% of total claims spend per EY, sitting inside a wider published spread that runs from 2% to 4% at the conventional end to 20% to 30% at the vendor end. We have already set out why no published source gives a leakage rate by line of business, and that caution applies with full force to anyone pricing the risk of a straight-through programme.

This is the part no page on this subject prices. A gate chain does not fail randomly, and it is not supposed to. It refuses exactly the files with contested coverage, disputed liability, inconsistent documents, an identity that will not verify or a fraud signal. So every point of straight-through coverage you win removes an easy claim from the human queue, and the queue left behind is denser rather than lighter. Capacity planning that treats the two as substitutes has it backwards: the file count falls while the work per file rises.

The handling-cost spread is both the qualification answer and the proof. If the claims that clear your gates are the 6.1-cent and 11.1-cent files and the claims that drop out are the 31.0-cent and 51.1-cent files, then the average cost to handle what is left rises by construction, with no change in anybody's productivity. Florida's duration multipliers show what happens if that queue then sits: $1,713 of loss adjustment expense on a non-litigated hurricane claim closed inside 61 days against $11,652 on one left open past a year. The residual queue is the expensive queue and it is also the slow one, which is the combination that compounds.

The queue the gates feed is also the one nobody automated. In the same NAIC auto survey, 0 of the fraud-referral models ran without human intervention, against 89 augmentation and 93 support. Capacity on that side has not moved either. In Rising Medical Solutions' 2023 workers' compensation benchmarking study of 1,388 frontline claims professionals, 53% reported average indemnity caseloads of 125 or fewer, which means 47% carried more and 26% carried 151 or more, while 46% reported spending 30% to 40% or more of their time on administrative tasks. The Bureau of Labor Statistics counted 389,700 claims adjusters, appraisers, examiners and investigators in 2025 and projects a 6% decline, 21,800 positions, by 2035, with 25.9% of the occupation aged 55 or over. A straight-through programme hands a harder book to a bench that is not growing.

Count the days on the drop-out side and the gap explains itself. Manual investigation runs 14+ days per case against 200+ cases per investigator, both Hesper internal benchmarks, which is why only about 25% of flagged claims get a real workup and the rest get a disposition instead. Raising straight-through coverage without raising investigation capacity does not remove the bottleneck. It moves it, and it moves it to the files where the money is. The upstream half of this decision, including why most triage scorecards measure four of their five lanes, is in our piece on the claims triage decision framework.

The regulatory record and the survey record agree, which is unusual enough to be worth stating plainly. No US rule bars automated approval, and the industry behaves accordingly: of the 193 large auto insurers in the NAIC's survey, 106 classified claim-assignment models ran with no human intervention and 9 companies reported a claim-approval model in use, while the number reporting a claim-denial model in any status, including research and proof of concept, was zero. Nobody built the denial model. The constraints that do bite land on refusal, on documentation, and on who the gates reject.

The statutory anchor is the NAIC's Unfair Claims Settlement Practices Act, Model #900, adopted in June 1990, which prohibits failing to adopt and implement reasonable standards for the prompt investigation and settlement of claims, refusing to pay a claim without a reasonable investigation, and failing to affirm or deny coverage within a reasonable time, with enforcement turning on conduct committed flagrantly and in conscious disregard, or with such frequency as to indicate a general business practice. The codified version to read closely is Washington's, because it is the newest and the most operationally specific. WAC 284-30-330(4), as amended by WSR 26-17-089 and effective October 18, 2026, defines as a specific unfair claims settlement practice: "Denying or refusing to pay claims in part or in full without conducting a reasonable investigation. 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." Read it as a specification rather than a prohibition. It tells you what a refusal has to be able to show, and it says nothing at all about an approval.

The governance instrument everyone cites covers claims and mandates nothing about human review. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted December 4, 2023, states that decisions or actions made or supported by AI must comply with all applicable insurance laws and regulations, and sets forth expectations as to how insurers will govern the use of AI. Its Section 3 names the lifecycle areas a written AI Systems Program should cover, including case management, claim administration and payment, and fraud detection, so claims is in scope explicitly. It neither prohibits a specific AI claims use nor requires a human to review an AI-informed claim decision. Section 4 lists what a department may request on examination, including the programme itself, third-party diligence, contractual audit rights, and documentation of validation, testing and drift evaluation. 29 jurisdictions had adopted it as of the NAIC implementation map dated October 8, 2026, with four more states carrying their own insurance-specific AI instrument. The instrument most likely to arrive first in practice is the NAIC AI Systems Evaluation Tool, piloted by 12 participating states as of March 2026, whose in-scope claims list names claim assignment, triage and fast-tracking, reserving, imaging analysis, fraud detection, litigation, closure-rate estimation and subrogation, and which reaches third-party claim administrators and managing general agents as well as carriers.

Three instruments get misread in this debate, so it is worth correcting the record precisely. New York DFS Insurance Circular Letter No. 7 (2024), issued July 11, 2024, is the most-cited AI-in-insurance instrument in the country, and it states that it "is not intended to address phases of the insurance product lifecycle other than underwriting and pricing," so its pre-deployment and recurring unfair-discrimination testing does not reach claims. Colorado SB21-169, which became law on July 6, 2021, does reach claims by statute, because the codified definition at C.R.S. 10-3-1104.9 includes claims management in the transaction of insurance, but its implementing regulation 3 CCR 702-10 has so far been extended only to life, private passenger auto and health benefit plan insurers. And Annex III of the EU AI Act covers, at point 5(c), "AI systems intended to be used for risk assessment and pricing in relation to natural persons in the case of life and health insurance." No point in Annex III covers insurance claims handling, so motor and home claims management sits outside the high-risk list. The instrument-by-instrument table, with a status date for every US bill and bulletin, is in our piece on New Jersey A5494 and the AI claim-denial rules, whose finding is the one that matters here: no US state has a statute in force requiring a human to make the final denial decision on a property and casualty claim.

Consumers draw the line in the same place the regulators do, which is convenient and should still be read for what it is. A consumer survey conducted by Propeller Insights on behalf of Liberate, a complementary front-door vendor, released on October 20, 2026, put the question as a trade: a 60-second AI payout with no human contact, or a human on the phone in 10 minutes with payment in three days. 49.5% chose the AI. On denials the same respondents flipped: 61% said a person must review every one, and only 33% would accept a no from AI alone, while one in four wanted both, an instant payout on yes and a person on no. 86% wanted AI use disclosed. It is a vendor-commissioned survey, and it discloses its field house, its dates of September 2 to 22, 2026, its sample of 1,001 US adults with recent property or auto claim experience, and a margin of error of plus or minus 3.1 points at 95%, which puts it well above most of what circulates in this category.

A straight-through programme designed around a target rate will hit the rate and lose the argument later. A programme designed around gates produces a rate as a by-product, and the rate it produces is one that survives an examination. Six operating rules follow directly from everything above.

1. Tune on drop-out quality, not on the rate. The question is never how many claims cleared, it is whether the claims that dropped out should have, and whether the ones that cleared would survive a file review. A screen running false positive rates in the 60% to 85% range can be made to look efficient simply by loosening it.
2. Audit a sample of straight-through-paid claims exactly the way you audit a denial. The denial is the decision the rules constrain, so it already attracts the scrutiny. The automated payment is the decision nobody read, which is precisely why it needs the same closed-file review.
3. Publish your own audit methodology, sample and denominator rather than quoting an industry leakage band. The published aggregate is 7% to 14% of total claims spend per EY, inside a far wider spread, and no board can act on a range that wide. Your own closed-file numbers, with their method stated, are worth more than anybody else's average.
4. Track the counter-metrics beside the rate and not instead of it: reopen rate, claims closed without payment over total claims closed, post-payment recoveries, and the cycle time of the claims that did not qualify. All four already exist in your Market Conduct Annual Statement filings.
5. Test the drop-out population for disparate impact before anybody asks you to. California's Bulletin 2022-5 already names ZIP-code-driven SIU referral as a Department concern, and a gate chain is a referral machine as well as a payment machine.
6. Keep the record reconstructable at the gate level, not the claim level. Section 4 of the NAIC Model Bulletin is a list of what a department can ask for, and the model approved it is not an answer to any of it. Each gate's inputs, output, reasoning and timestamp is.

That sentence is the whole design claim, and it is where Hesper AI differs from a rules engine with a good qualification list. Clean claims resolve straight through; suspicious claims get an investigation-grade workup, and the same engine does both. Hesper runs 15+ investigation phases in parallel rather than one at a time, in minutes, not weeks, which takes coverage of flagged claims from about 25% to 100% on the same internal benchmark that puts manual investigation at 14+ days per case. Every agent action is logged with its sources, reasoning and timestamps, so the straight-through payment and the drop-out referral both leave a record an examiner, a reinsurer or a capacity partner can reconstruct. Adjusters review evidence-backed files instead of building them, and the licensed human still makes every decision that is a decision, including every denial. For the decision-rights version of that split, see our piece on which claims decisions stay human and what evidence they get handed.

The asymmetry at the start of this piece is not going away, because it is a feature of how US claims law is built. Payment is rewarded by the prompt-payment statutes and constrained by almost nothing. Refusal is constrained by the unfair-claims rules, by the reasonable-investigation standard, and by discrimination testing on whoever gets refused. A straight-through programme that reads that correctly puts its gates in front of the payment and its evidence behind the refusal. One that reads it as permission to raise a number will find out its real coverage level the first time somebody opens fifty paid files and recalculates them from the documents inside.