Plaintiff firms got AI first: the demand package is assembled in hours and examined in weeks

The average third-party medical expense submitted on an auto bodily injury claim reached $32,300 in the first quarter of 2026, up from $24,300 in the first quarter of 2023, according to CCC data reported by Claims Journal in August 2026. That is the demand side: what was billed and presented, not what was paid. It is the artifact landing in an adjuster's queue.

The package carrying that number got faster to build. The plaintiff bar adopted generative AI and compressed demand-package assembly from days into hours. Carrier-side examination of those packages did not compress. It still runs at the speed of a human reading records, one file at a time. Everything else in this post follows from that asymmetry, including why auto bodily injury severity keeps climbing on a shrinking base of claims, and why the exposure sits in a category almost nobody investigates.

This post walks the mechanics: how the auto loss dollar shifted to bodily injury, what is actually inside a demand package and which parts of it are checkable, why the plaintiff bar's adoption of AI is legitimate work, not misconduct, why the industry has framed the response as a document-comparison problem when it is a claim-examination problem, and what the statutory clock does to a carrier that cannot examine at volume.

Two framing notes before the data. Social inflation claims are a real and separately caused phenomenon, and AI is not their origin; a full section below concedes that and explains why it makes the capacity argument stronger, not weaker. And the category at issue is buildup rather than hard fraud, which is a distinction with operational consequences we cover in our pillar on claims automation.

Bodily injury has taken over the auto loss dollar. CCC's August 2026 analysis found BI claims were 52.3% of the combined dollars paid across the two claim categories, bodily injury and auto property damage, up from 44.4% in 2022. The mix shifted while total claim counts moved the other way.

CCC's own Crash Course 2026 report, published March 2026, states the same shift against a different base: bodily injury now accounts for 52.4% of total liability dollars paid, "despite only about one in four property damage exposures carrying an associated BI exposure." One in four exposures is producing more than half the dollars. That is a severity story, not a frequency story, and it is the shape a claims VP should recognize immediately from the loss-ratio side.

The frequency numbers underneath it diverge sharply, and the metric names are not symmetric, so both need saying precisely. Using ISS Fast Track data cited in the CCC report, bodily injury claim frequency rose 4% over the last two years while property damage paid claim frequency decreased 12.6% over the same period. Fewer paid property damage claims, more bodily injury claims, and a much larger average bodily injury number.

On severity, CCC reports that the average personal BI claim payout rose 21% over the same two years. Separately, and over a different window, CCC reports that the average amount paid per BI claim rose 9% to 10% per year to reach more than 30% over four years, at more than $30,000 per feature. The source does not define "feature," and a feature is not the same unit as a claim file, so the $30,000 figure should be read as a per-feature average rather than a per-claim one.

The submitted-expense line is the one to sit with, because it measures a different thing from all the others. Every other row is an outcome the carrier participated in producing. The submitted number is the input, authored entirely on the other side of the table. From Q1 2023 to Q1 2026 it grew roughly a third, a growth rate derived from CCC's two published figures. That is a statement about what is being presented for examination, and it says nothing on its own about what was justified.

Which is the point. Nobody can say what share of that growth is medical cost inflation, what share is more thorough documentation of real injuries, and what share is buildup, because the examination that would answer the question does not happen at the volume required to answer it. The number went up. The measuring instrument did not change.

A demand package is the bundle claimant counsel sends a carrier to open settlement negotiations on a third-party bodily injury claim. It contains medical records, itemized billing, a treatment chronology, lost-wage documentation, the property damage estimate, a liability narrative, and a general damages figure. Most of that is factual assertion. The rest is legal argument.

The split matters more than anything else in this post. A factual assertion can be tested against evidence that exists independently of the person making the assertion. A legal argument cannot, and should not be. Treating a demand's general damages number as an inflation signal is a category error, and carriers that make it end up in exactly the kind of bad-faith exposure the last section of this post describes.

Read the last two rows again, because they are the guardrail on everything that follows. A high general damages demand is not evidence of anything except that claimant counsel is doing the job the claimant hired them to do. A liability theory the carrier disagrees with is a dispute, not a red flag. The five rows above them are different in kind: each one is a factual claim about the world, and each one has a source of truth sitting outside the envelope.

Note what the independent-evidence column actually requires. Prior claim history is a cross-carrier database lookup. Provider billing patterns are a comparison across other claims and, ideally, other carriers. Employment verification is an outbound contact and a payroll record. Imaging order dates against visit dates is a document-level reconciliation inside the records themselves, and the failure modes there are specific enough that we wrote a separate piece on medical record fraud in insurance claims. Each of those is a different system, a different query, and in a manual workflow a different hour of somebody's day.

That is why examination is slow in a way that assembly never was. Assembling a demand package means gathering documents the claimant already has a right to. Examining one means reaching outside the package into six or seven independent sources and reconciling what comes back. The two tasks are not mirror images, and automating one does not automate the other.

Adoption of generative AI in demand-package assembly moved faster than adoption of automation in demand-package examination. That is a statement about tooling timelines, not about conduct. A Thomson Reuters survey cited by Claims Journal in August 2026 found 41% of law firms said they were using generative AI, up from 28% the year before.

The mechanics of what that adoption compressed were described by Erik Bahnsen, director of casualty industry analytics at CCC, in the same Claims Journal coverage.

The work he is describing is the paralegal work: gathering bills and records, assembling the chronology, citing the documentation, drafting the letter. Vendors sell it explicitly. EvenUp, a plaintiff-side demand generation platform, advertises drafting demands in minutes, a dataset of more than 250,000 verdicts and settlements behind its valuations, and published throughput of 10,000 cases processed weekly. Those are the vendor's own marketing figures, and they are worth taking at face value for what they indicate about direction of travel.

Now put the two sides of the same file next to each other, task by task. The claimant-side column describes capability that vendors publicly sell. The carrier-side column describes what most US P&C carriers actually do with a third-party bodily injury demand in 2026.

Four of the five rows have a vendor answer on the claimant side. The fifth row has no vendor answer on either side, and it is the only row that produces a fact rather than a document.

There is a second-order effect worth naming. When assembly time falls, the number of files a firm can carry rises, and the average file that gets a full demand package gets smaller. Work that used to be economically viable only on serious-injury files becomes viable on mid-severity ones. The result is not primarily bigger demands on the same claims. It is fully documented demands on claims that previously would have settled on a phone call and a two-page letter. The carrier now receives a professionally assembled evidentiary package on a file it was never going to investigate.

Demand validation, as the claims-technology market currently defines it, is document comparison: checking a demand letter against the medical records, billing, and depositions that accompany it. Claim examination is a different task. It tests the demand's factual assertions against evidence the claimant did not supply, which is where prior claims, provider patterns, and employment records live.

The industry named the first problem well. In a sponsored content piece published by Claims Journal on 4 August 2026, Amy Mingopoulos, writing for Wisedocs, described the pressure precisely.

That is an accurate description of a real operational gap, and Wisedocs named it in public before anyone else did. Their medical record review platform automates medical record review at scale for carriers, TPAs, law firms, IMEs, and government, with published claims of automating and scaling 70% of the medical record review workflow. CCC is building into the same space: its April 2025 announcement pairs CCC Injury Evaluation Solutions with EvolutionIQ's Medhub for demand package organization and deduplication, and medical document synthesis that extracts and synthesizes insights from extensive medical records, with every AI-surfaced item traceable and auditable.

Both of those are useful. Both are the correct response to a document-volume problem. And here is the observation this post is built on, offered as our reading and not as anything either source said: neither the Claims Journal news coverage of the severity data nor the Wisedocs piece uses the words fraud, SIU, or investigation anywhere. We checked. The industry looked at a wave of AI-assembled demand packages, correctly identified that human review capacity could not keep up, and framed the answer as reading the documents faster.

That silence is the finding. Reading the documents faster is necessary and insufficient, because the questions that determine whether a demand is supported are mostly not answerable from the documents in the demand.

Four of the nine questions are simply not reachable from inside the envelope, and they are the four with the highest signal on buildup. This is a scope distinction, not a criticism. Document intelligence answers document questions well. It was never designed to reach a cross-carrier claim history or a payroll record. CCC produced the severity data this entire post is built on and deserves credit for publishing it; the estimating-and-document-processing versus investigation line is one we draw at more length in our comparison with CCC Intelligent Solutions.

The Insurance Research Council defines fraud as a "specific material misrepresentation of the facts of a loss" and buildup as "the inflation of an otherwise legitimate claim, such as through unnecessary medical treatments or diagnostic procedures." Buildup is the category that lives inside a demand package, and it is more common than fraud.

Those definitions come from IRC's Fraud and Buildup in Auto Injury Insurance Claims, 2008 Edition, covering 2007 claim data. The operational difference is that hard fraud has a smoking gun and buildup does not. A staged collision has a moment where the story breaks. An extra six weeks of chiropractic care on a real injury has no such moment. It has a pattern, and patterns are only visible against comparison data.

The prevalence split from that same 2007 claim data is the clearest published statement of which category is larger.

The more recent edition, using 2012 claim data and reported by Insurance Journal in February 2015, found that 21% of bodily injury claims and 18% of PIP claims closed with payment had the appearance of fraud and/or buildup. Claims with the appearance of buildup accounted for 15% of dollars paid for BI and PIP claims in 2012. The excess-payment estimate was between $5.6 billion and $7.7 billion, representing between 13% and 17% of total payments under the five main private passenger auto injury coverages. The study covered more than 35,000 auto injury claims closed with payment, contributed by 12 insurers representing 52% of the US private passenger auto market.

For the broader loss context, the Coalition Against Insurance Fraud puts insurance fraud across all lines of US insurance at $308.6 billion a year, and reports that fraud occurs in about 10% of property-casualty insurance losses. That 10% is the pool. Buildup is the part of it that does not look like anything until you compare it to something.

Why buildup is the least investigated category

Now the arithmetic that determines what actually gets worked. A manual special investigations workup runs 14+ days per case and roughly $2,500 per case against Hesper's published internal benchmarks. Set that against a bodily injury feature averaging more than $30,000. The workup consumes about 8% of the payout before a dollar of it is recovered or denied, a figure derived by dividing the two.

Eight percent is not an absurd number in isolation. It is an impossible number as a default policy, because the SIU has to spend that 8% before knowing whether there is anything to find, on a category where the base rate of a clean finding is high and the expected recovery on a positive finding is a fraction of the feature. No SIU director builds a referral queue that way, and none should be asked to.

So the queue gets built the only way it can be: largest exposure first, clearest indicators first, prosecutable cases first. Mid-severity buildup fails all three tests. It is not the largest exposure, it has no clear indicator until somebody looks, and no district attorney is taking it. This is the structural reason manual SIU teams fully investigate roughly 25% of flagged claims, and we catalogue the rest of the reasons in why flagged claims never get investigated.

One further note on where buildup concentrates. It is rarely evenly distributed across providers. Treatment patterns cluster, referral chains repeat, and the same names recur across files that share nothing else. Provider-pattern analysis is the highest-yield check on a demand package and also the one least available to any single-file review, because it is a question about the population, not the file. We walk that pattern in detail in medical mill fraud investigation.

The aggregate consequence is leakage rather than headline fraud losses: dollars paid above what the claim should have cost, spread thinly across a very large number of files, each individually too small to justify the look. That is the mechanism we describe in our pillar on claims fraud leakage, and bodily injury demand packages are one of its largest single reservoirs.

California requires an insurer to accept or deny a claim within 40 calendar days of receiving proof of claim, and separately requires a thorough, fair and objective investigation. Those two obligations run at the same time on the same file. The examination window is fixed by regulation while assembly time on the other side collapsed.

The text is worth reading in full, with its lead-in clauses intact, because the lead-ins are where the operational constraint actually sits. All three subsections below are from California's Fair Claims Settlement Practices Regulations at 10 CCR 2695.7.

Read the scope precisely: the 40-day clock runs from receipt of proof of claim, not from receipt of a demand letter, and subsection (b)(4) carves out exceptions. Those are not the same trigger and conflating them produces a wrong answer about what is due when. But the structure holds regardless of trigger date. Subsection (d) requires a thorough investigation. Subsection (b) caps the time available to finish one. Subsection (c)(1) turns every extension into a written, dated, repeating admission on the file that the carrier is not done yet.

Here is the arithmetic in the vise. A 14+ day manual investigation consumes roughly 35% of a 40-calendar-day window on a single file. The investigator holding that file is carrying 200+ open cases and completes roughly 10 investigations a month. The window is fixed, the workup is long, and the caseload is a multiple of the throughput. There is no scheduling solution to that system of constraints. It is arithmetic, and the only variable that moves is which files get examined at all.

Massachusetts frames the same duty as a list of prohibited omissions. G.L. c. 176D, section 3(9) opens: "Unfair claim settlement practices: An unfair claim settlement practice shall consist of any of the following acts or omissions:" Four of the enumerated omissions bear directly on a carrier that cannot examine a demand at volume.

• (b) Failing to acknowledge and act reasonably promptly upon communications with respect to claims arising under insurance policies
• (e) Failing to affirm or deny coverage of claims within a reasonable time after proof of loss statements have been completed
• (f) Failing to effectuate prompt, fair and equitable settlements of claims in which liability has become reasonably clear
• (g) Compelling insureds to institute litigation to recover amounts due under an insurance policy by offering substantially less than the amounts ultimately recovered in actions brought by such insureds

What under-investigating a large file now costs

In September 2025, Suffolk Superior Court Justice Debra Squires-Lee entered a final judgment of approximately $91 million in a matter involving Peerless, Liberty Mutual Fire, and Ohio Casualty Insurance Co. and John Rooney, an injured construction worker, as reported by Claims Journal. The judgment is a statutory doubling of an underlying award of roughly $45.49 million, itself a $26.6 million jury verdict plus pre-judgment interest, triggered by a finding that the insurer willfully failed to investigate and to make reasonable settlement offers. The judge characterized the award as "grossly excessive" and found she had "no discretion" to reduce it.

That is not a fraud case and should not be read as one. Its relevance here is directional and it points at the compliance officer more than the SIU director: the statutory penalty for under-investigating a large-severity file is now large enough that "we did not have the capacity" is not a posture anyone wants to defend. Investigation capacity has become a compliance exposure and not only a loss-cost lever. The documentation standard that goes with it is the subject of our piece on the defensibility standard for AI fraud investigation, and the California-specific filing obligations are in our 10 CCR 2698 SIU compliance guide.

The two-sided nature of the exposure is what makes this hard. Under-examine and you overpay buildup, quietly, forever. Under-examine and then deny anyway, or delay to buy time, and you are in bad-faith territory. The only exit that satisfies both a regulator and an actuary is a documented, thorough examination completed inside the window on every file that warrants one. That has been the correct answer for thirty years. It has never been operationally available at volume.

Social inflation claims are liability claims whose cost growth economic inflation does not explain. AI did not create that growth. The severity trend was compounding years before generative AI reached law firms, and the strongest evidence for that comes from the same body of research this argument relies on.

This section makes the case against the post. Four counter-arguments, in descending order of how much they should change your mind.

1. The severity trend predates generative AI entirely

This is the strongest objection and it is correct. The Insurance Research Council released a study in July 2026, "Auto Injury Insurance Claims: A Study of Increasing Claim Severity," covering more than 7.4 million auto injury claims closed with payment, contributed by nine insurers representing roughly 43% of the US private passenger auto market. As Insurance Business reported, average bodily injury claim payments rose from approximately $14,000 in 2017 to more than $20,000 in 2022, an annualized increase of 7.8%.

The study window runs mid-2017 to mid-2022. It closes before generative AI was in any law firm workflow. Over that same window, across all coverages, the share of claimants represented by an attorney rose from 40% to nearly 50%, with bodily injury claimants seeing the sharpest increase at 11 percentage points, and litigation rates rose from 10% to 18% of claimants. Represented bodily injury claimants waited a median of almost 440 days for their claim to close, more than twice as long as unrepresented claimants, and netted $1.40 per dollar of medical bills paid against $1.80 for those who went without an attorney.

Concede the point plainly: severity was climbing at 7.8% a year for reasons that have nothing to do with AI. Medical cost growth, rising attorney representation, litigation rates nearly doubling, and litigation funding are all independent drivers with their own histories. Any argument that AI caused the severity curve is contradicted by the curve's own shape before AI existed.

Then notice what conceding it does to the capacity argument. If demands were already growing at 7.8% annualized, if representation was already approaching half of claimants, and if assembly time then fell from days to hours on top of that, the volume of documented, professionally assembled demands arriving per adjuster went up on a trend that was already going up. The counter-argument does not weaken the capacity case. It removes AI as the thing to fix and leaves capacity as the only remaining variable a carrier controls.

2. AI is a force multiplier, not the cause

The people who published the observation say so themselves. In Claims Journal's coverage of the CCC data, the report's authors are described as considering AI's role as "more of a force multiplier" than the sole influence. That phrase belongs to the CCC report's authors, and it is the right framing. A force multiplier acting on a trend that was already compounding produces a larger number without being its origin.

3. Litigation dynamics are their own driver

The litigation environment has moved independently of anything in a claims department. A Triple-I and Casualty Actuarial Society analysis published in October 2025, authored by James Lynch and William Nibbelin, attributed $231.6 billion to $281.2 billion in increased liability insurance losses over the past decade, as of year-end 2024, to legal system abuse and related litigation trends, across personal auto liability, commercial auto liability, other liability (occurrence), and product liability (occurrence).

The personal auto liability cut is $91.6 billion to $102.3 billion, or 8.7% to 9.7% of booked losses over that decade. Note that personal auto carries the smallest percentage share of the four lines and one of the largest dollar figures, which is what happens when a low rate is applied to a very large book.

On the verdict side, Marathon Strategies' "Corporate Verdicts Go Thermonuclear 2025 Edition," reported by Insurance Journal, counted 135 nuclear verdicts above $10 million in 2024, a 52% increase over 2023, totaling $31.3 billion, a 116% jump over the prior year. Thermonuclear verdicts of $100 million or more increased to 49, five of which exceeded $1 billion. That data is scoped to corporate defendants, not to auto bodily injury verdicts, and should not be read as a personal auto number. It describes the environment in which every liability settlement is negotiated, which is a different and still relevant thing.

Broader still, CCC's Crash Course 2026 cites Swiss Re's Social Inflation Index showing a 33% increase in liability costs from 2020 to 2024 tied to social inflation alone. None of these is an AI effect. All of them were running before the first AI-drafted demand letter was sent.

4. Carriers are not standing still

Also true. Wisedocs is selling medical record review at scale into carriers and TPAs today. CCC has paired Injury Evaluation Solutions with EvolutionIQ's Medhub for demand package organization, deduplication, and traceable medical synthesis. Those are real capabilities shipping into real claims departments, and any carrier without something in that category is behind. The scope line drawn earlier stands: they answer document questions well and were not built to reach a prior claim history, a provider network pattern, or a payroll record.

Take all four counter-arguments together and the honest version of the thesis narrows to something more defensible than the headline. AI did not cause social inflation. AI changed the ratio between how fast a demand can be assembled and how fast one can be examined. That ratio was already unfavorable. It is now considerably more unfavorable, on a trend that was independently compounding, inside a statutory window that did not move.

Matching AI-assembled demand volume does not mean reading demands faster. It means changing the unit cost of examining one, so that a mid-severity bodily injury file clears the threshold where a full workup is economically rational. That is a coverage problem before it is a speed problem, and coverage is the number nobody reports.

Return to the arithmetic from the buildup section, because it is the whole argument in two lines. At roughly $2,500 per manual case, examining a bodily injury feature averaging more than $30,000 costs about 8% of the payout. At roughly $150 per case, the same examination is about 0.5% of it. Both percentages are derived from Hesper's published internal benchmarks against CCC's published per-feature average. An 8% examination cost forces triage. A 0.5% examination cost does not.

That is the entire mechanism. Not better detection, not a higher hit rate, not a smarter score. The set of claims worth examining is defined by the cost of examining them, and when that cost falls by an order of magnitude the set expands from the largest and most obvious files to every flagged file. Coverage moves from roughly 25% of flagged claims to 100% of them. Make every flagged claim investigable is the shortest statement of what changes.

Why a demand package is a parallel-processing problem

Look back at the anatomy table. A demand package is a multi-document artifact whose components are checkable against different, unrelated sources. Records against prior records and deposition testimony. Billing against fee schedules, CPT-to-note correspondence, and cross-claim provider patterns. Chronology against FNOL and police-report timing. Wages against employer and payroll verification. Injury severity against the property damage estimate. None of those checks depends on the output of any other.

A human investigator runs them sequentially because a person can only read one document at a time, and that sequencing is most of why a manual workup takes 14+ days. Hesper runs 15+ investigation phases in parallel on every flagged claim: document forensics, provider and prior-claims cross-checks, billing-code review, treatment-pattern analysis, statement and deposition reconciliation, timeline reconstruction with gap detection, OSINT, and financial pattern analysis, all at once instead of in a queue. The output is an audit-ready report a human SIU lead reviews and signs, not a score handed to an investigator who then does the work. The mapping between the demand package's components and the investigation phases is close to one-to-one, which is unusual and is why this particular artifact is a good fit for the approach.

The result is examination in hours rather than weeks, which matters less for its own sake than for what it does to the two constraints above it: the 40-calendar-day window stops being a scheduling crisis, and the 200+ case backlog stops setting the coverage ceiling. The investigator's role shifts from execution to decision-making.

Where this sits in the stack

The layered model is worth stating explicitly for a demand package, because four different vendor categories touch the same file and only one of them is doing examination.

Detection is upstream; investigation is downstream. A detection vendor scored that claim at FNOL and was right to; it had no view of a demand package that did not exist yet. Hesper is complementary to FRISS, Shift Technology, and Verisk, not a replacement, and the same is true of document intelligence at the lateral layer. The investigation row is the one with no software incumbent, which is why the default answer there has always been headcount and why the coverage number has been stuck at roughly a quarter for as long as anyone has measured it.

For a claims VP, the loss-ratio version is simple enough to write on one line: buildup is a percentage of paid losses on the fastest-growing component of your auto book, and the percentage you recover is bounded by the percentage of files you examine. For an SIU director, it is that the referral queue stops being a triage exercise. For a compliance officer, it is the one that matters most, and it is the note to close on.

The ability to complete a documented, thorough, objective investigation inside a 40-calendar-day statutory window, on every file that warrants one, is a compliance capability first and a loss-cost lever second. Subsection (d) has always required it. Subsection (b) has always capped it. What changed is that the volume of assembled, documented demands arriving on the other side went up, and the tooling that makes the two obligations simultaneously satisfiable finally exists at the investigation layer rather than only at the document layer.