Claims leakage by line of business: nobody publishes the rate. Six bureaus publish the drivers.

There is no published claims leakage rate by line of business. Not for auto, not for property, not for workers' compensation, not for general liability. The single aggregate band that is tied to a named firm's documented file-review practice is EY's, at 7% to 14% of a carrier's total claims spend, and it describes a whole book. IRMI, which supplies the definition the industry uses, publishes no percentage at all.

That absence is not a research gap to be filled with an estimate. It is the reason a single blended leakage number sends the first fix to whichever queue happens to be loudest. What actually differs by line is the driver mix: which mechanism is doing the leaking, how heavily, and under what conditions. Severity, cycle time, attorney involvement, litigation rate, reserve development, catastrophe share, medical share, recovery ratio. All of those are line-specific, all of them are published annually, and none of them is a leakage rate.

So this post does the per-line work the honest way. It covers what the one aggregate band actually measures, the four mechanisms leakage runs through and how their weights change between auto physical damage, auto bodily injury, property and catastrophe, workers' compensation and commercial liability, the one per-line dataset a US regulator files, what the regulatory measurement apparatus collects line by line and what it leaves out, and how to run the measurement on your own book without borrowing anyone's percentage. The cross-line view of where leakage and fraud overlap sits in our pillar guide to claims leakage; this is the deep expansion of its one by-line section.

Start with the line that has the deepest per-line data in US property and casualty. Workers' compensation has a rating bureau that publishes severity, frequency, combined ratios and reserve adequacy every single year. Presenting NCCI's 2026 State of the Line in May 2026, the bureau's chief actuary said this about her own line.

If the bureau with the deepest per-line dataset in the country will not reduce its own line to one number, a vendor quoting a leakage rate for your auto book is not citing anything. What EY does publish is worth reading precisely because of how it was produced. The 7% to 14% band comes out of EY's own property and casualty claims quality assessments, which are closed-file reviews. In one top US insurer's litigated book, that review put leakage at 10% of total paid. Nearly two thirds of the files reviewed carried some assessed leakage. More than 85% of the leakage EY assessed sat in three places: coverage determination, litigation prevention, and evaluation and resolution. That last sentence is the useful one, because it says the leak is concentrated in decisions rather than in payments.

The rest of the numbers in circulation deserve naming. The figure repeated most often on this topic - roughly 6% of claim payments, about $67 billion a year - traces back to an article that does not contain it, which is why our State of Claims Automation 2026 report excludes it by name and prices the credible range per homeowners file instead. Below EY's band sits a 2% to 4% figure that circulates as conventional industry wisdom with no study attached. Above it sits 20% to 30%, which is one consultancy's own audit findings with no disclosed sample, method or denominator. Every coefficient we are willing to stand behind, with its source, is listed on the claims leakage calculator, including the one thing it deliberately leaves out: a leakage rate that varies by line of business, because no published band does.

The strongest by-line statement anyone has put in print came from a consultant writing in Canadian Underwriter in 2015. It names every line this post covers, and it carries no numbers, which is exactly the shape of the evidence base eleven years later.

Leakage is a residual. It is the difference between what a claim should have cost under the policy as written and what it did cost, and it accumulates through four mechanisms: the amount was evaluated wrong, a recovery was available and not pursued, a fraudulent or inflated element was paid, or the file stayed open long enough that handling cost and claimant behaviour changed the number. Which of those four dominates is set by line-specific conditions, and that is where the real per-line answer lives. Speed and leakage are the same problem: every one of the four mechanisms gets worse the longer a file sits without the evidence that would resolve it.

Set the four mechanisms against the four line groups and the asymmetry is obvious. On auto physical damage the live mechanisms are valuation and salvage, and the file closes in weeks. On auto bodily injury the live mechanisms are damages evaluation and litigation timing, and the file closes in years. On property the mechanism is scope and depreciation under surge conditions. On casualty it is a reserve set against an incomplete evidence record. Same four mechanisms, four completely different operating problems.

One number does travel across all of them, and it is the one most leakage conversations skip. NAIC's 2025 full-year industry results report loss adjustment expense of $86.0 billion against $551.8 billion of net losses incurred, which is 15.6% derived, up from 15.3% on the prior year figures. Handling cost is not a rounding error on the leakage question, it is a sixth of losses. For the per-line version of it, the place to look is the NAIC Report on Profitability by Line by State, which is where a carrier can get loss adjustment expense ratios line by line rather than blended.

CCC's Crash Course 2026 report, published March 31, 2026, is the reference dataset for this line, and three findings in it move the valuation question. Total loss frequency reached 23.1% of claims, a new industry high. 28.3% of repairable estimates now include calibrations, which is a line item that did not meaningfully exist a decade ago. And there were 12 million fewer vehicles six years old or newer in operation as of the third quarter of 2025 than in 2020, which raises the share of the book where an older vehicle meets a modern repair cost and tips into a total. CCC publishes its own dataset, so the figures are theirs, but on auto they are the reference set.

The customer-facing side says the same thing from the other direction. The J.D. Power 2026 U.S. Auto Claims Satisfaction Study has repairable cycle time at 19.3 days, down from 22.3. Inside that average, vehicles from 2015 and older with no advanced driver assistance ran 17.9 days while vehicles from 2019 onward with three or more ADAS features ran 21.5 days. Total losses were 27% of claims in that study, up from 24%. And the number that matters for leakage: only 58% of total-loss customers were satisfied with the valuation they were offered. A valuation that two in five claimants dispute is a valuation that generates supplements, escalations and handling cost.

Commercial auto physical damage is the one piece of commercial auto that works. AM Best put its 2024 combined ratio at 88.6, with about $1.5 billion of underwriting profit, inside a commercial auto segment that lost about $4.9 billion overall. Short-tail, physical, verifiable. That contrast within one commercial product is the cleanest proof available that the line label on a claim tells you far less about leakage risk than the coverage part does.

The Insurance Research Council's Auto Injury Insurance Claims: A Study of Increasing Claim Severity, published July 30, 2026, is the most useful per-line dataset on this topic. It covers 7.4 million auto injury claims from nine insurers representing 43% of the US private passenger auto market, from mid-2017 to mid-2022. The average bodily injury payment moved from about $14,000 to above $20,000, a compound 7.8% a year. Settlement leverage went from $1.80 per dollar of medical bills to above $2.30. Attorney representation across coverages rose from 40% to nearly 50%, with BI claimants up 11 percentage points. The litigation rate went from 10% to 18%. And the median time to closure for a represented BI claimant is nearly 440 days, against less than half that for an unrepresented one.

Those two figures are not strictly comparable and the comparison is still the point. One is a median closure interval for a represented injury claimant, the other is an average repair cycle. Set side by side they describe about 23 times the days open between two halves of the same auto policy. The order of magnitude is what matters, because an operation built for a 19-day file is the same operation handling the 440-day file, with the same adjuster, the same queue and usually the same diary intervals.

The newer IRC survey work explains why the long side starts long. Public Opinions on Attorney Involvement in Auto Insurance Claims: 2026, published October 1, 2026, found about 70% of recent claimants consulted or hired an attorney, nearly 50% engaged counsel within a few days of the accident and 80% within weeks. More than 90% had seen or heard attorney advertising in the past year, 60% believe that advertising increases liability claims and lawsuits, and about 80% of claimants who consulted an attorney were referred to a medical provider for treatment.

That timing is the whole operational argument on this line. If representation attaches within days and treatment follows a referral, the evidence that decides the damages evaluation is being created by the other side while the file is still in intake triage. The macro picture matches: the Casualty Actuarial Society and Triple-I analysis of auto liability inflation put the 2014 to 2023 impact on personal auto liability losses and defence and cost containment expense at $76.3 billion to $81.3 billion, with personal auto liability severity compounding at 3.6% from 2014 to 2019 and 9.8% from 2020 to 2023. CCC has average paid bodily injury severity up 10.3% year over year and 32% over four years.

This is where the evidence sequence decides the number. A damages evaluation built after a demand lands is a reaction to someone else's record. One built from the file as it accumulates - treatment dates reconciled against the mechanism of loss, provider patterns checked across claims, statements cross-referenced, the coverage position cited to the endorsement - is a position. The capacity problem underneath it is well documented: about 25% of flagged claims get a full workup today because 200+ cases per investigator rations the rest, which is the argument in how uninvestigated claims drain profitability. Those are Hesper internal benchmarks for the manual baseline, and on a 440-day file the rationing is invisible until the file resolves.

Property cycle time is improving, which is worth saying before the harder news. The J.D. Power 2026 U.S. Property Claims Satisfaction Study, drawn from 5,093 homeowners surveyed between December 2024 and December 2025, has the average claim running 40.7 days from first notice to final payment, down 3.4 days year over year, and 29.6 days to complete repairs, down 2.8. Overall satisfaction rose 20 points to 702. Digital adoption is doing real work: 38% filed initial notice digitally, 49% submitted photos and 45% received updates digitally. Almost one in five customers still said the experience was not great.

Verisk's Quarterly Property Report for the second quarter of 2026 is where the mix shift shows. US property claim assignment volume was 1.24 million, down 12.2% year over year and 13.1% below the five-year average, while catastrophe claims made up 43% of assignments against 34% five years ago. Those two facts together are a derivation worth stating plainly: if the total is falling while the catastrophe share is rising, the decline is concentrated in non-catastrophe work, so the average property file arriving in 2026 is likelier to be a catastrophe file than at any earlier point in that series. Trade coverage of the same report adds a severity reading of $17,085 at report date with projected maturation to between $18,794 and $19,323, and a US reconstruction labour composite up 3.9% against materials up 2.1% between June 2025 and June 2026, which you can read in the Verisk report itself.

The peril mix behind that is documented too. Swiss Re Institute's sigma 1/2026 put global insured natural catastrophe losses at US$107 billion in 2025, with severe convective storms at US$51 billion, wildfires at a record US$40 billion, and floods at only US$3.4 billion against a US$15.4 billion prior five-year average. Secondary perils made up a record 92% of total global insured natural catastrophe losses, and about 49% of the US$220 billion of economic losses was insured, the highest insured share on record. A book whose catastrophe load is shifting from hurricane to hail and wildfire is a book whose claim population is shifting from a few large files to many medium ones, which is a different leakage profile with the same annual loss total.

Property is also the only line where a state regulator publishes a per-line litigation rate, which makes Florida property the most thoroughly documented line-and-jurisdiction combination in the country. The Florida Office of Insurance Regulation's July 2026 Property Insurance Stability Report has Florida's share of nationwide homeowners insurance lawsuits falling from 79% in 2020 to 41% in 2025, while its share of national homeowners claims fell from 9% to 5%. Florida homeowners underwriting moved from $1.5 billion of losses in 2020 to $1.7 billion of gains in 2025, and at least 20 new carriers entered after the 2022 and 2023 reforms. For the baseline, NAIC data reported in 2021 had Florida at just over 8% of US homeowners claims opened in 2019 and more than 76% of all litigation against insurers nationwide.

Read that as a measurement finding rather than a policy one. The same peril, the same policy form and the same claim can carry a litigation exposure that differs by a factor of several depending on the statute in force, which is why a national per-line leakage rate would be meaningless even if someone published one. For the per-file dollar framing, our State of Claims Automation 2026 report prices the published range against the $13,349 average US homeowners claim closed with payment. And the capacity side of a catastrophe surge is the subject of headcount buys throughput, not depth per file, which is the relevant constraint when 43% of your assignments arrive in clusters.

NCCI's 2026 State of the Line is the cleanest illustration in US property and casualty of why a blended number hides the mechanism. The calendar year 2025 combined ratio for workers' compensation was 91. The accident year 2025 combined ratio was 102, which NCCI expects to develop favourably toward 97 as claims settle and close. One line, one year, three different answers depending on which lens you use. Medical claim severity rose 4% in 2025 and indemnity claim severity rose 4%, while lost-time claim frequency fell 2% and net written premium slipped 0.2%. The redundant industry reserve stood at $14 billion, down from $16 billion in 2024. California, at about 20% of the national market, carried an accident year combined ratio of 129. The primary document is NCCI's 2026 State of the Line At a Glance; the figures above are as reported by Insurance Journal.

The medical share question has a current answer too. NCCI's September 2026 Asked and Answered brief puts the preliminary medical share of workers' compensation losses for accident year 2025 at 53%, unchanged from 2024 and down from a peak of 56.6% in accident year 2016, with the metric broadly flat over two decades. Indemnity is currently growing faster than medical because wage growth is outpacing medical inflation. For a claims operation that matters: the half of the loss dollar that responds to medical management is not the half that is accelerating.

The litigation split inside the line is the sharpest per-line number in this entire post. A WCRI study of more than 950,000 claims found attorney involvement adds $7,700 to $12,400 to a claim, with involvement at 34% overall among workers with more than seven days of lost time, 14% of temporary-disability-only claims, 38% of claims with permanent partial disability only, 64% of claims with permanent partial disability or a lump sum, 76% of lump-sum-only claims and 79% where there is both permanent partial disability and a lump sum. Our own arithmetic on two of those: 14% against 64% is a 4.6 times difference in litigation exposure between two claim types inside the same line of business, which is a wider spread than most people assume exists between entire lines.

The same study splits involvement by injury type, and the pattern holds: neurological pain carried 61% involvement with a $40,670 average benefit, inflammations 43% at $29,913, carpal tunnel 40% at $21,204 and spine conditions 36% at $15,199. A workers' compensation book is therefore not one population for audit purposes. It is at minimum a short-duration population where the leak is medical pricing and return-to-work timing, and a litigated population where the leak is the evidence record behind compensability and apportionment. Sampling them together produces a number that describes neither.

AM Best's commercial auto analysis is the cleanest evidence that this mechanism is systemic rather than firm-specific. The commercial auto liability combined ratio was 113 in 2024 and 113.3 in 2023. It has been above 100 every single year since 2014, reaching 113 five times. The segment lost about $4.9 billion on underwriting in 2024 and about $5.5 billion in 2023, and AM Best puts the industry $4 billion to $5 billion under-reserved on the liability side. Companion coverage of the same analysis adds that 14 of the top 20 commercial auto insurers ran above a 100 combined ratio in 2024, a fourteenth consecutive year of underwriting losses, with a ten-year average loss of slightly over $2.9 billion a year.

The verdict environment explains the reserve problem. Marathon Strategies' 2026 edition of Corporate Verdicts Go Thermonuclear counted nearly 200 nuclear verdicts of $10 million or more in 2025, up 40.7% on 2024 and the highest since 2009, totalling about $25.6 billion. Forty of those were thermonuclear verdicts of $100 million or more and four exceeded $1 billion. They came from 97 courts across 28 states and 68 industries. Product liability alone accounted for 29 nuclear verdicts totalling $12 billion, Georgia produced 10 totalling $4.8 billion and Texas 29 totalling $3.3 billion. The insurance sector itself took 5 verdicts for $390 million.

Two findings from the 2026 CLM Litigation Management Study, as reported by Peter Venetis of ValueMomentum, turn that into an operating instruction. Auto and trucking has overtaken general liability as carriers' single largest litigation spending driver, at 33% of total litigation spending, up from 24% in 2023. And a mean of 87% of non-workers'-compensation litigated claims settle, against a verdict rate of 2.9%. If 87 out of every 100 litigated files settle, the file is almost always negotiated rather than tried, which means the evidence record is the entire leverage position. The study itself is published in full.

The inflation arithmetic sits underneath both. The Casualty Actuarial Society and Triple-I put the 2014 to 2023 inflation impact on commercial auto liability losses and defence expense at $42.7 billion to $55.8 billion, which with the personal auto liability figure makes a combined $118.9 billion to $137.2 billion, or 9.9% to 11.5% of $1.2 trillion of total net losses over that decade. Commercial auto severity rose 78% from 2014 to 2023 while CPI rose 29%. A reserve set in 2026 against a 2026 evidence record will be tested by a 2029 verdict environment, and nothing about the closed-file audit cycle surfaces that error in time to act on it.

Salvage and subrogation is the exception that proves the thesis of this post, because here a regulator's own journal does publish per-line ratios. Bisco and Fier, writing in the NAIC Journal of Insurance Regulation in 2023, aggregated Schedule P filings from 1996 to 2021 and found recovery at 4.5% of net claims paid across the full sample and 6.2% across firms recording any recovery, with 77.1% of property-liability insurers recording any recovery in a given year, which leaves roughly one in four recording none. By line, auto physical damage rose from about 11% of net claims paid in 1996 to about 20% in 2021, dominated by salvage on total losses. Personal auto liability sat at 2.0% and commercial auto liability at 1.1%. The maximum observed recovery was nearly 90% of net claims paid. These are firm-year averages winsorised at the 1st and 99th percentiles, with medians much lower at 1.85% and 3.15%, and 28% of personal auto liability recovery is collected in the first year after the loss.

Our own arithmetic on those published ratios: about 20% against 2.0% is roughly ten times, and about 20% against 1.1% is roughly eighteen times, so a dollar of recovery opportunity is far likelier to be captured on an auto physical damage file than on either auto liability line. The mechanism is not mysterious. Salvage on a total loss is automatic, physical and screened by the estimating system. Subrogation on a liability file requires somebody to identify the responsible party, assemble the evidence and file the demand inside the statutory window, and that is discretionary work on a desk that is already full. The reason it goes missing is in missed subrogation opportunities are lost at intake, not at closure. The maximum observed recovery of nearly 90% is the proof that the ceiling is an operating choice and not a structural limit.

This is also where the vendor landscape splits by line, and it is worth being precise about it. CCC sells AI subrogation with demand packaging and auto fraud screening, and publishes the most detailed per-line driver data on auto anywhere in this post. The fraud screening is described in auto terms, and the driver data is an auto book: nothing in Crash Course travels to a workers' compensation or general liability file. CLARA Analytics is strong on exactly the casualty drivers this post identifies, and its public materials describe decision support alongside the claims system rather than moving a claim end to end. EvolutionIQ, now a CCC company, describes guidance for adjusters on open injury and disability claims. Five Sigma's Clive is the closest thing to a multi-line AI claims platform, and its published Risk agent flags suspicious activity and recommends further investigation rather than conducting it, with subrogation not among the published agents. All four are credible in their lane. A multi-line carrier or TPA that needs the same depth on four lines currently assembles it from four vendors and six bureaus by hand.

That assembly job is the one Hesper AI is built for. It is the AI claims resolution platform: agents take every claim from first notice to final recovery, with investigation-grade evidence behind every decision and fraud detection built in, across auto, property, workers' compensation, general liability and pet. The per-line point is narrow and specific. One evidence file per claim - sourced, timestamped, reconstructable - serves the coverage position, the damages evaluation, the fraud finding and the recovery demand at the same time, which means the physical damage file gets its salvage screen and the liability file gets its subrogation screen from the same pass that produced the coverage position. Every file screened for subrogation and salvage, not the ones somebody had time for. Clean claims resolve straight through; suspicious claims get an investigation-grade workup, with 15+ investigation phases running in parallel. Those last figures are Hesper internal benchmarks.

The per-line measurement machinery in US insurance regulation is already built. The NAIC Market Conduct Annual Statement instructions are published separately for each line, including homeowners and private passenger auto. Carriers report, per line, claims closed with payment and closed without payment bucketed by days to closure, from 0 to 30 days out past 365, the median days to final payment, the claims closed without payment because the amount claimed fell below the deductible, and the suits opened and closed. That is a per-line claims-handling dataset filed annually with every state that collects it.

Read the element list again and notice the shape of the gap. Every element measures timeliness, disposition or litigation. None measures accuracy. A carrier can report a median days-to-payment in the top quartile of its peer group on every line it writes while paying the wrong amount on a quarter of those files, and nothing in the statement would show it. Fair-claims-settlement-practice regimes have the same shape: they impose acknowledgement and decision deadlines, documentation standards and communication duties, not accuracy standards. Leakage is the quantity the regulatory apparatus does not ask for, which is precisely why it has no agreed measurement standard and no published per-line rate.

Two consequences follow for anybody running the measurement themselves. First, the NAIC Market Regulation Handbook's sampling guidance holds that results cannot be generalised beyond the field the sample was drawn from, so a sampled leakage rate describes the files you opened and nothing else - the sample-size arithmetic is worked through in the stage-by-stage claims leakage audit checklist. Second, the record that proves a decision was right is the same record a market conduct examiner, a capacity partner or a reinsurer asks for, and the requirements for that record predate AI by decades, which is the argument in what the claim-file rules asked for before AI. Build the per-line audit so its output satisfies both readers at once and the work gets paid for twice.

Stratify by line first and by cause second, because the cause codes that dominate one line barely appear in another. Depreciation and pricing-data errors drive a property book. Damages evaluation and litigation timing drive a bodily injury book. Total-loss valuation and salvage drive a physical damage book. Compensability and apportionment drive a workers' compensation book. Run one blended sample across all four and the cause distribution you get back is an artifact of your line mix, not a finding about your operation.

Then run the audit in the same order every time, so the output is comparable across lines even though the inputs are not.

1. Define the population by line and coverage part, not by claim number range, and record the denominator you used before you look at a single file.
2. Draw the sample stratified by severity band, because a random draw from a long-tail line will miss the files where the money is.
3. Re-work each file against the policy as written: coverage position, damages evaluation, recovery screen, payment accuracy, in that order.
4. Price the gap per file, cite the evidence for each finding, and code the finding to the lifecycle stage where it originated rather than to the handler.
5. Pair the result with the line's published driver metric for the same period, so a rising per-file gap can be separated from a rising severity environment.

Step five is the one most audits skip, and it is where a per-line programme earns its budget. A bodily injury book whose per-file gap widened in 2025 while representation rates rose 11 percentage points and settlement leverage moved from $1.80 to above $2.30 per dollar of medical bills has an environment problem partly outside its control. A physical damage book whose per-file gap widened while total-loss frequency rose to 23.1% has a valuation problem it owns. Same measured widening, two different interventions, and only the driver pairing tells them apart.

The operational conclusion is the one the drivers point to in every line. Each of the four mechanisms is an evidence problem before it is a payment problem: a coverage position taken without reading every endorsement, a damages evaluation built after the demand arrives, a recovery not screened because nobody had the time, a fraudulent element paid because 25% of flagged claims get a real workup and 200+ cases per investigator rations the rest. Manual investigation runs 14+ days per case, so coverage gets rationed by arithmetic rather than by judgement. Those are Hesper internal benchmarks for the manual baseline.

Running 15+ investigation phases in parallel takes that case to minutes, not weeks, at a fraction of the manual cost, and moves coverage of flagged claims from about 25% to 100%. Adjusters review evidence-backed files instead of building them, and every decision stays with the licensed human.

Every claim, first notice to final recovery, with evidence behind every decision. For a leakage programme that means one evidence file per claim that serves the coverage position, the damages evaluation, the fraud finding and the recovery demand, built the same way on an auto physical damage file as on a commercial liability file, with the per-line driver metrics attached so the number means something. Start with the definition if you need it - the claims leakage definition - then run the claims leakage calculator on your own book, where every coefficient carries its source and the per-line rate is the one input we refuse to invent.