Complete guide · 2026
AI for Bodily Injury Claims: The Complete 2026 Guide
How insurers, TPAs, and defense teams use AI to review medical records, answer demands, and manage risk on injury claims, with current statistics, state rules, checklists, and an ROI calculator.
Andrej EvtimovamaiseThe short answer
AI for bodily injury claims is software that reads an entire injury claim file (medical records, bills, demand packages, police reports, pleadings, and adjuster notes), organizes it into a source-linked chronology and summary, and answers adjusters' questions with a citation to the exact page. The adjuster still decides coverage, liability, value, and settlement.
Key takeaways
- The pressure is real: personal auto bodily injury severity reached $28,278 per claim in 2024, up 66% since 2015, and 57% of BI claimants had an attorney by 2022.
- Four kinds of AI touch injury claims: rules-based valuation (the Colossus era), predictive models, document and generative AI, and agentic AI. They solve different problems.
- The highest-value job is reading. Files of several hundred to several thousand pages are where AI saves the most time and catches what people miss.
- Source links are non-negotiable. Every statement in an AI summary should open the page it came from, because omissions and hallucinations still happen.
- Regulators expect governance, not bans. 29 U.S. jurisdictions have adopted the NAIC AI Model Bulletin, which explicitly covers claim administration.
- Start with one painful workflow such as medical record review or time-limited demand response, measure it, then scale.
What is AI for bodily injury claims?
AI for bodily injury claims is a set of technologies that help insurers, third-party administrators (TPAs), independent medical examiners, and defense counsel understand injury claim files faster and more completely. It applies to third-party liability claims where someone is hurt: personal and commercial auto liability, general liability, workers' compensation, and medical malpractice.
In practice, a modern system does four jobs:
- Read everything. It ingests the full file, including scanned and handwritten pages, and turns it into searchable text.
- Organize it. It splits bundled PDFs into individual documents, classifies them, removes duplicates, and puts treatment in order by date of service.
- Extract what matters. It pulls out diagnoses, procedures, providers, treatment gaps, pre-existing conditions, billed amounts, and liability facts.
- Answer with evidence. It writes summaries and answers questions, and every statement links back to the page that supports it.
What it is not is an autonomous settlement engine. The most credible tools are built to inform the adjuster's judgment, not replace it, and regulators expect insurers to stay accountable for every decision "regardless of the tools and methods" used, in the words of the NAIC Model Bulletin on the Use of AI Systems by Insurers.
How is it different from general claims automation?
General claims automation focuses on high-volume, low-complexity work: first notice of loss intake, payments, routing, and simple property or auto physical damage claims. Photo-based damage estimation is a good example. Bodily injury is different. The facts live in long, messy, unstructured documents written by many different authors, the medical questions are subtle, and the financial stakes are high. A tool that is excellent at estimating a dented bumper from a photo cannot tell you whether a lumbar disc herniation was pre-existing.
Who uses AI on injury claims?
- Carrier BI adjusters and casualty claims teams who review records, set reserves, and respond to demands.
- TPAs that handle liability and workers' comp claims for self-insured clients.
- Claims leaders and litigation managers who need early warning on severe or litigation-prone files.
- Independent medical examiners and defense counsel who need a fast, neutral chronology.
- Plaintiff law firms, on the other side of the table, which now use AI to assemble demand packages at scale.
Why bodily injury claims need AI in 2026
Bodily injury has become the most expensive and least predictable part of auto liability. Costs are rising faster than medical inflation, more claimants are represented, litigation is more common, and the experienced adjusters who used to absorb that complexity are retiring.
average personal auto BI liability claim in 2024, up from $17,014 in 2015 (Triple-I / ISO)
of BI claimants were represented by an attorney by 2022 (Insurance Research Council)
of auto injury claimants litigated in 2022, up from 10% in 2017 (IRC)
of auto liability dollars paid go to bodily injury, though fewer than 1 in 4 damage exposures carry a BI claim (CCC Crash Course 2026)
median time to close for represented BI claimants, more than double unrepresented ones (IRC)
in nuclear verdicts against corporate defendants in 2025, from nearly 200 verdicts of $10M or more (Marathon Strategies)
Claim severity is outrunning medical inflation
According to Triple-I, citing ISO data, the average personal auto bodily injury liability claim rose from $17,014 in 2015 to $28,278 in 2024. CCC's Crash Course 2026 reported average paid BI severity up 10.3% year over year and 32% over four years.
The Insurance Research Council's 2026 auto injury study, based on more than 7.4 million claims, found that the average BI payment grew about 7.8% a year between 2017 and 2022, roughly double the rate of medical-care inflation. Over the same period, the amount paid per dollar of medical bills rose from $1.80 to more than $2.30.
More attorneys, more litigation, longer claims
The same IRC study found that 57% of BI claimants were represented by 2022, the litigation rate for auto injury claimants nearly doubled from 10% to 18%, and represented claimants waited a median of about 440 days for their claim to close. Speed matters too. In Sedgwick's own book of business, 80% of pending auto BI claims had attorney representation within 14 days of first notice, and litigated BI claims cost about 5.5 times as much as non-litigated ones.
At the extreme end, Marathon Strategies counted nearly 200 nuclear verdicts of $10 million or more against corporations in 2025, totaling more than $25.6 billion, including $2.9 billion against automobile defendants. Triple-I and the Casualty Actuarial Society estimate that legal system abuse added $91.6 to $102.3 billion to personal auto liability losses between 2015 and 2024.
The plaintiff side is already using AI
Demand packages are increasingly assembled by software. EvenUp, the best-known plaintiff-side platform, says it works with more than 2,000 law firms and was valued at more than $2 billion after its 2025 funding round. When one side produces polished, 500-page demands in hours and the other side reviews them by hand, the side reading by hand loses time it often does not have, especially under a time-limited demand.
The adjuster capacity gap
The U.S. Bureau of Labor Statistics counts 389,700 claims adjusters, appraisers, examiners, and investigators and projects that number to fall 6% by 2035, while about 21,600 openings a year still need to be filled as people retire or leave. Sedgwick expects a quarter of claim adjusters to retire by the end of 2027. Experience is leaving the desk just as files get longer and more contested.
Four kinds of AI used on injury claims
"AI" covers very different technologies in claims. Mixing them up is the most common reason evaluations go wrong, so it helps to separate them before comparing vendors.
Rules-based valuation
Expert systems that score adjuster-entered injury codes to produce a value range. The Colossus era.
Predictive models
Machine learning trained on historical claims to score severity, litigation risk, or fraud.
Document and generative AI
OCR, classification, extraction, and language models that summarize records and answer questions.
Agentic AI
Teams of specialized AI agents that run multi-step reviews with checkpoints and human sign-off.
1. Rules-based valuation
Rules-based systems ask an adjuster to code the injuries and treatment, then apply severity points to produce a range for general damages. They are consistent and fast, but only as good as the coding, and they cannot read the file themselves. Their history (covered below) is the main reason regulators and claimants are wary of "black box" valuation.
2. Predictive models
Predictive models use structured claim data such as injury type, venue, attorney involvement, and treatment patterns to estimate the probability of an outcome: severity, litigation, or fraud. They are useful for triage and for routing complex files to experienced people early. Their blind spot is the same as rules engines: if the important facts are buried on page 312 of a scanned record, the model never sees them.
3. Document and generative AI
This is the category that changed the most since 2023. Optical character recognition (OCR) and vision models turn scans into text, classifiers sort pages into document types, and large language models (LLMs) summarize records and answer natural-language questions. Done well, it removes the hours spent finding and ordering information. Done carelessly, a general-purpose chatbot can produce fluent summaries that omit or invent facts, which is why source-linked answers matter so much (more on that below).
4. Agentic AI
Agentic AI chains several specialized steps together: one agent sorts and deduplicates records, another extracts medical facts, another checks bills against treatment, another drafts the summary, and a quality-assurance step checks that every statement has a source. Anthropic's widely cited guide to building effective agents recommends exactly this kind of design for high-stakes work: simple, composable steps with checkpoints and a pause for human review. In claims, the human checkpoint is the adjuster.
| Approach | What it reads | What it produces | Explainability | Best for | Watch out for |
|---|---|---|---|---|---|
| Rules-based valuation | Injury and treatment codes entered by the adjuster | A general damages range | Medium: rules are fixed, but tuning is often opaque | Consistency on routine soft-tissue claims | Garbage in, garbage out; rigid adherence to outputs |
| Predictive models | Structured claim data | Risk scores and probabilities | Low to medium | Triage, routing, early severity flags | Bias in historical data; misses facts in documents |
| Document and generative AI | The documents themselves | Chronologies, summaries, answers | High when every statement links to a source page | Medical record review, demand review, Q&A | Omissions and hallucinations without source links |
| Agentic AI | The whole file plus claim system data | Multi-step reviews, drafts, alerts, reports | High with step-level audit trails | End-to-end file review, large loss reports, demand response | Over-automation; needs clear human checkpoints |
Looking for a Colossus alternative?
Colossus is the best-known rules-based BI valuation system. A 1999 Connecticut legislative research report describes it as a program originally developed in Australia and marketed in the U.S. by Computer Sciences Corporation (CSC), and CLM Magazine describes roughly 600 injury profiles that assign severity points to adjuster-entered data. It does not assess specials, liability, or offsets, and each insurer "tunes" it with its own settlement history.
That tuning became the center of controversy. In 2010, Allstate agreed to a $10 million settlement with 45 states after a multistate market-conduct examination. Regulators found inconsistent tuning across regions but "no systemic underpayment," and Allstate agreed not to require or reward settlement at the Colossus value and to tell claimants the tool might be used.
The lessons translate directly to modern AI:
- Read the file, not just the codes. Modern document AI works from the records themselves, so the quality of the output no longer depends on how carefully someone keyed in injury codes.
- Show your work. Every fact and suggestion should link to its source page, so the adjuster, a supervisor, a regulator, or a court can see why.
- Keep humans in charge of value. AI should prepare the evaluation, not dictate it. Incentives tied to matching a machine number are exactly what regulators objected to.
- Govern changes. Document model versions, testing, and overrides, as the NAIC bulletin now expects.
How AI for bodily injury claims works, step by step
Under the hood, a credible system runs the same sequence an experienced adjuster or nurse reviewer would, only faster and without skipping pages. Here is the pipeline from upload to answer.
- 01Ingest & OCR
- 02Split & classify
- 03Deduplicate
- 04Extract & code
- 05Chronology
- 06Analyze
- 07Knowledge graph
- 08Cited answers
Step 1: Ingest and OCR every page
Files arrive from the claims system or by upload. OCR and vision models convert scanned, faxed, and handwritten pages into searchable text. This matters because injury files are rarely clean: a single demand package can bundle hundreds of pages from a dozen providers in one PDF.
Step 2: Split and classify documents
The system splits bundled PDFs into individual documents and labels each one: emergency department record, MRI report, physical therapy note, operative report, bill, police report, demand letter, pleading, and so on. amaise's Core module, for example, classifies more than 50 document types.
Step 3: Remove duplicates
Demand packages repeat records. The same MRI report can appear three times, attached to different letters. Deduplication keeps repeated records from looking like separate treatment events and shrinks what a person has to read.
Step 4: Extract and code medical facts
Diagnoses, procedures, providers, dates of service, restrictions, and prescriptions are extracted and normalized, typically to ICD-10-CM for diagnoses and CPT for procedures. This is where engineering matters: a 2024 study in NEJM AI found that GPT-4 reproduced exact ICD-10-CM codes only 33.9% of the time when asked directly. Reliable coding comes from retrieval against coding knowledge bases plus validation, not from free-form generation.
Step 5: Build the treatment chronology
Events are ordered by date of service, not the date a document was printed or sent. Gaps in treatment are left visible rather than smoothed over, because a two-week delay before first treatment or a three-month break in care can matter to causation and value.
Step 6: Analyze causation, pre-existing conditions, and bills
The system flags prior injuries and degenerative findings, inconsistencies between providers' histories, and treatment that does not match the reported mechanism of injury. It reconciles bills to records and dates of service, spots duplicate charges, and applies offsets such as MedPay or PIP. Many carriers benchmark charges against data sources such as FAIR Health's property and casualty benchmarks, and Medicare conditional payments must be resolved before settlement; Milliman notes that Medicare can recover double damages if it is not repaid within 60 days.
Step 7: Store the facts in a knowledge graph
Instead of a pile of text, the extracted facts are stored as linked entities: the claimant, each provider, each diagnosis, each date, each bill, each source page. That structure makes cross-document questions possible ("Which providers documented prior neck pain?") and lets the system spot contradictions. amaise calls its per-claim knowledge graph CaseDNA.
Step 8: Answer with citations
Summaries and answers are generated from retrieved facts and pages, a pattern known as retrieval-augmented generation (RAG), first described by Lewis and colleagues in 2020. In a claims context, the non-negotiable feature is page-level provenance: click a statement, see the page.
Illustrative example. Each citation opens the source page in the file.
Why source links are non-negotiable
The research on AI medical summarization is encouraging but not perfect. In a Stanford study published in Nature Medicine, physicians rated the best adapted model's summaries as equivalent to expert summaries 45% of the time and superior 36% of the time. A large clinical evaluation of nearly 13,000 note sentences found a hallucination rate of 1.47% and an omission rate of 3.45%, and 44% of the hallucinations were classed as major.
Translated to a 500-page injury file, small error rates are still meaningful. The practical safeguards are the same ones a good supervisor would insist on: constrained extraction instead of free writing, a source link on every statement, explicit flags when something cannot be sourced, and a human who checks the facts that drive the decision. amaise's own guide shows how to check a claims AI answer against the source page in under a minute.