TL;DR: Insurance claims automation works best when it takes over the reading, matching and first-draft work of a claim (FNOL intake, document ingestion, coverage checks against the policy file, triage scoring, subrogation and litigation flags, file summaries) and leaves the judgment calls with adjusters. Straight-through processing is real for simple property and auto claims, but most of the money in a casualty or specialty book sits in claims that will never go straight through. The carriers getting results pick one countable unit first (cost per claim, days to first reserve, recoveries per 1,000 closed files) and fix the data underneath before they add agents. Bain's 2025 claims assessment found only 4% of P&C insurers have scaled AI, so the gap is execution, not ideas.

A bodily injury claim comes in on a Tuesday. By Friday it has a police report, two ER bills, a recorded statement, a photo set and a letter of representation from a plaintiff firm. The adjuster who owns it has somewhere north of 120 other open files. The letter of rep sits in a shared inbox for six days before anyone links it to the claim, the reserve still reflects a soft-tissue injury, and the policy's additional insured endorsement that would have pointed to a second carrier never gets read. Nobody made a bad decision. The file simply arrived faster than a person could read it.

That is the problem insurance claims automation is supposed to solve, and it is a different problem from the one most vendor pages describe. Most of what ranks for "AI claims processing" is about making a windshield claim or a small homeowners claim close in a day. Useful, but it is not where a claims VP at a casualty carrier, an MGA running a specialty program, a TPA or a self-insured risk manager loses sleep. The expensive part of the book is the claim that gets an attorney, a misread reserve, a missed recovery or a coverage question nobody caught at intake.

This guide is written for claims VPs, COOs and CEOs at carriers, MGAs, TPAs and self-insured programs. It covers what you can actually automate across the claim lifecycle, where straight-through processing stops paying off, how to score triage and litigation risk, what claims leakage really is, a medical professional liability angle, what to measure, and a 90-day plan. If you want the broader view of where carrier AI budgets are going this year, our insurance AI and data trends for 2026 post covers that. This one stays inside the claims department.

Key takeaways

• Only 4% of P&C insurers have scaled AI across the organization, per Bain & Company's 2025 claims maturity assessment of 81 insurers, even though 78% use generative AI in some form. Pilots are common. Production is rare.

• Attorney involvement is the claim cost driver to watch. The Insurance Research Council found the share of auto injury claimants represented by attorneys rose from 40% in 2017 to nearly 50% in 2022, and represented bodily injury claimants waited a median of nearly 440 days for closure, more than double unrepresented ones.

• Adjuster capacity is shrinking, not growing. The U.S. Bureau of Labor Statistics projects employment of claims adjusters, appraisers, examiners and investigators to decline 5% from 2024 to 2034, with openings driven by retirements and exits.

• Missed subrogation is a recoverable cost. The National Association of Subrogation Professionals estimates insurers miss recovery on roughly 15% of claims that could have been subrogated, as cited by Genpact.

• MPL is under real pressure. AM Best reported a 108% combined ratio for medical professional liability specialty writers in 2024, the tenth straight year of underwriting losses.

• Straight-through processing is a small-claim tool. The bigger return in casualty and specialty lines comes from earlier, better-informed human decisions on the complex claims, not from removing humans from simple ones.

• Name the countable unit before you automate. Without a baseline like cost per closed claim or days from FNOL to accurate reserve, you cannot prove ROI, and the project stalls at renewal.

What is insurance claims automation, really?

Insurance claims automation is the use of software, and increasingly AI agents, to do the repeatable work of a claim: capturing the first notice of loss, reading and classifying documents, matching the loss to the policy, scoring severity and complexity, routing the file, flagging fraud, litigation and recovery signals, and drafting summaries and reserves for a human to approve. It is not one product. It is a set of jobs that today get done by adjusters, claims assistants, nurse reviewers and supervisors reading the same paper over and over.

Three generations of tooling are worth separating, because buyers often conflate them.

• Rules and workflow. The claim system itself (Guidewire ClaimCenter, Duck Creek Claims, Majesco, Origami Risk, Ventiv, or a homegrown system at many TPAs) with business rules, diaries and assignment logic. This is the backbone. It does exactly what it is configured to do and nothing more.

• Predictive models. Severity scores, fraud scores, litigation propensity models. These have been around for a decade at larger carriers. They are only as good as the structured fields they see, and much of what matters in a claim is not in structured fields.

• Language models and agents. Systems that can read a 400-page medical record, a police narrative or a demand letter, pull out what matters, compare it against the policy and the claim history, and write a first draft. This is the new part, and it is why claims automation now reaches into casualty and specialty lines where it never could before.

Our position: the third generation only works if it can see what the first two already hold. An agent that reads a demand letter but cannot see the policy endorsements, the prior claims on the same claimant, or the billing status of the policy is producing confident summaries of half the picture.

How do you automate claims processing, step by step through the lifecycle?

You automate claims processing by breaking the lifecycle into its reading and matching tasks and handing those to software, while keeping decisions on coverage, liability, reserves above authority and settlement with people. Here is what that looks like at each stage, in the order a file actually moves.

FNOL intake

FNOL automation is the most mature piece. Claims arrive by phone, web form, broker email, agency portal, TPA feed and, for self-insured programs, an internal incident report. Automation here means capturing the loss consistently regardless of channel: parsing the email or call transcript, extracting date of loss, location, parties, vehicles or premises, injury description and reporter, and matching the claim to the right policy and insured on the first try. The quiet win is deduplication. The same loss reported by the insured, the broker and a claimant's attorney should become one claim with three contacts, not three claims.

Document ingestion

This is where the volume lives. A single casualty file can collect police reports, medical records and bills (UB-04 and CMS-1500 forms), repair estimates from CCC or Mitchell, contractor invoices, recorded statement transcripts, photos, demand packages, letters of representation, ISO ClaimSearch results and correspondence. Automation classifies each document, extracts the fields that matter (dates of service, CPT and ICD-10 codes, billed versus paid amounts, treating providers, officer narrative, citation issued), attaches it to the right claim and exposure, and alerts the adjuster only when something changes the picture.

Coverage verification against the policy file

Coverage review is where a lot of leakage starts. Automation compares the reported loss against the actual policy as issued: the declarations, the forms and endorsements, the limits, retentions and deductibles, the policy period, additional insureds, exclusions and any manuscript language. The agent's job is not to make the coverage call. Its job is to surface the clauses that matter and the facts that touch them, so a coverage specialist reads two pages instead of sixty. This depends entirely on the policy file being complete and correct, which is why we treat policy file integrity as a prerequisite, not an afterthought.

Triage and severity/complexity scoring

Triage decides who gets the file and how fast. A good triage score blends what is known at FNOL (injury type, venue, vehicle or premises type, policy limits, claimant age) with what the documents reveal in the first two weeks (treatment pattern, surgery recommended, attorney retained, prior claims). The output is a routing decision: fast track, desk adjuster, field, complex or large loss unit, nurse case management, or SIU review.

Reserving support

Agents should not set reserves. They should make reserving faster and better informed. That means a draft reserve recommendation with its reasoning laid out: the medical specials to date, the treatment trajectory, comparable closed claims in the same venue, attorney involvement and policy limits. The adjuster or supervisor accepts, edits or rejects it. Every reserve change still carries a human name.

Fraud and SIU signals

Fraud detection is pattern recognition across claims: the same provider, attorney, tow company or body shop appearing in unrelated losses, a policy bound days before the loss, treatment patterns that do not match the injury, inconsistencies between the recorded statement and the police narrative. The Coalition Against Insurance Fraud's 2022 study put the annual cost of insurance fraud in the U.S. at $308.6 billion across all lines, a figure some researchers have criticized as hard to verify. Whatever the true number, the operational point holds: SIU teams are small, and their time should go to the referrals with real evidence attached, not to every claim a rule happened to trip.

Litigation and attorney involvement flags

The moment a letter of representation arrives, the economics of a claim change. Per the Insurance Research Council's study of more than 7.4 million auto injury claims, bodily injury claimants with attorneys netted $1.40 per dollar of medical expense after legal fees versus $1.80 for unrepresented claimants, and waited more than twice as long. Automation here does two things: it detects the letter of rep or demand the day it arrives and links it to the right claim, and it scores open claims for the likelihood of representation so that adjusters can make early contact while it still matters.

Subrogation recovery identification

Subrogation gets missed because the signal is buried in the narrative: "the other driver ran the light," "the water heater failed after a recent install," "the forklift was rented." Automation reads every file, including closed ones, for third-party liability facts and product, contractor or landlord involvement, then queues the candidates for a recovery specialist. This is the core of our claims and subrogation agent.

Claims file summarization

Every time a file changes hands (reassignment, supervisor review, roundtable, reinsurance reporting, defense counsel handoff, audit) someone rereads it from the top. A current, cited summary of the claim (facts, coverage position, liability assessment, injuries and treatment, reserves and their history, open diaries, recovery potential) saves hours per handoff and makes roundtables faster. The citation part is not optional. A summary that cannot point back to the page it came from will not be trusted, and should not be.

What is straight-through processing in insurance claims, and where does it stop?

Straight-through processing (STP) means a claim moves from FNOL to payment without a human touching it. It works for high-volume, low-severity, low-ambiguity claims: glass, towing and rental, small first-party property, some travel and pet claims, and parametric covers. For those, a clean policy match, a validated estimate or invoice under a threshold and a passed fraud screen can close the claim in hours.

Here is the arguable part. Many carriers are chasing STP rates as the headline metric for their claims automation program, and for a casualty, specialty or MPL book that is the wrong target. A bodily injury claim, a general liability premises claim, a professional liability claim or a workers' compensation lost-time claim should never go straight through, and pushing a higher STP rate on a book where 20% of claims drive 80% of losses moves the cheap claims faster while leaving the expensive ones untouched.

For those lines, the better framing is "straight-through preparation." The agent does all the reading, matching and drafting so that when the adjuster opens the file, the coverage clauses are highlighted, the medical chronology is built, the reserve draft is waiting and the subrogation and litigation flags are already set. The human still decides. They just decide on day three instead of day thirty.

What is claims leakage, and how much does it cost?

Claims leakage is the gap between what a carrier paid on a claim and what it should have paid under the policy and good claims practice. It includes overpayment, paying uncovered items, missed deductibles or other insurance, missed subrogation and salvage, late reserving that drives poor settlement decisions, and avoidable expense such as unnecessary independent medical exams or litigation that earlier contact would have prevented.

How big is it? Honestly, nobody knows your number but you. A July 2026 Insurance Thought Leadership article put leakage at 7% to 14% of total claims payouts, but it did not cite a study for that range, and most leakage figures in circulation trace back to vendor estimates or old consulting audits. We would rather you measure it than quote it. The usual method is a closed-file audit: pull a stratified random sample of closed files, have senior examiners score each against your best practices, and calculate the dollars that should not have gone out the door.

Common claims leakage examples

• Coverage leakage. Paying a loss excluded by an endorsement nobody read, or missing that another carrier's additional insured coverage should respond first.

• Recovery leakage. Closing a file without pursuing subrogation, contribution or salvage. NASP's estimate that about 15% of subrogable claims are missed belongs here.

• Medical bill leakage. Paying billed charges on a liability claim without review for unrelated treatment, duplicate billing or charges well above reasonable value.

• Timing leakage. A reserve that stays low until the demand arrives, by which time the settlement value has moved. Swiss Re has reported that U.S. liability claims grew 57% over the past decade, and delay compounds that trend.

• Expense leakage. Defense counsel retained on a claim that could have settled pre-suit, or duplicate vendor assignments on the same loss.

Notice that almost every item on that list is a reading problem. The information to prevent it was in the file. Nobody had time to connect it.

Will AI replace claims adjusters?

No. AI will change what adjusters spend their day on, and it will let a smaller team carry a larger book, but the judgment calls in claims (coverage positions, liability assessments, settlement authority, conversations with injured people) are exactly where you want an accountable person. The labor math is going the other direction anyway: the BLS projects adjuster employment to fall 5% from 2024 to 2034, and the roughly 21,600 openings it projects each year come from people leaving the field, not new seats.

Caseload is the real pressure. Rising Medical Solutions' 2022 Workers' Compensation Benchmarking Study, as cited by Milliman, found 32% of respondents reported indemnity caseloads of 126 or more claims, well above the commonly recommended ceiling. When an adjuster carries that many files, the expensive claims get the same attention as the cheap ones, which is to say not enough.

Regulators are also explicit that humans stay in the loop. The NAIC's Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted by 24 states as of March 2025 according to law firm Quarles, asks insurers to maintain a written AI governance program, document how AI is used across underwriting and claims, oversee third-party vendors, and consider the extent of human involvement in final decisions. A claims automation design that cannot show who approved what will have a hard time in a market conduct exam.

What should not be automated in claims?

Do not automate decisions that carry legal, financial or human consequences you would need to defend: coverage denials and reservation of rights letters, liability determinations, reserve changes above an adjuster's authority, settlement offers on injury claims, SIU referrals that will lead to a claim being contested, and any conversation with a claimant who is grieving, hurt or angry. Automate the reading and preparation for each of those. Keep the decision and the signature with a person.

A practical way to draw the line:

1. Agent decides and acts. Document classification, data extraction, duplicate detection, claim-to-policy matching on exact identifiers, diary creation, routine acknowledgment letters, payment of validated small invoices under a defined threshold.

2. Agent drafts, human approves. Triage assignment on borderline scores, reserve recommendations, file summaries, coverage clause extraction, subrogation referrals, demand response outlines.

3. Agent flags, human investigates. Fraud indicators, litigation propensity, coverage conflicts, possible bad-faith exposure (an approaching time-limited demand, for instance).

4. Human only. Denials, liability decisions, settlement authority, claimant conversations on serious injury or death claims.

Write this matrix down before you buy or build anything. It is also the document your compliance team and your regulator will want to see.

How does claims automation apply to MPL and specialty commercial lines?

Medical professional liability and specialty commercial lines are where claims automation is hardest and where it pays the most per claim. Frequency is low, severity is high, files are long, and the outcome often turns on details buried in hundreds of pages of clinical records. AM Best reported that MPL specialty writers ran a 108% combined ratio in 2024 with a $586 million underwriting loss, their tenth straight year in the red, and TransRe's verdict tracking counted 57 medical malpractice verdicts of $10 million or more in 2023, per reporting in MDedge.

In MPL, the claim usually begins as a potential claim or incident report, a records request or a notice of intent, long before a suit is filed. The reading load is enormous: the EHR record exported from Epic, Cerner (Oracle Health) or MEDITECH, nursing notes, imaging reports, medication administration records, consent forms, peer review materials (handled under privilege rules), and the deposition transcripts that follow. Where automation helps:

• Medical chronology. A dated, page-cited timeline of the patient's care, built in hours rather than the days a nurse reviewer or paralegal needs, with gaps and deviations flagged for the reviewer.

• Standard-of-care questions surfaced early. Not a judgment on negligence, which belongs to experts and counsel, but a list of the decision points the plaintiff will likely focus on: delayed diagnosis, missed follow-up on an abnormal result, communication handoffs.

• Insured and exposure mapping. Which physicians, groups, facilities and employed providers are involved, which policies and limits apply (claims-made with tail, entity versus individual coverage, consent-to-settle clauses), and whether another carrier shares the exposure.

• Venue and severity context. How comparable claims in the same venue and specialty have resolved, which feeds the reserve recommendation.

Specialty commercial lines (management liability, E&O, cyber, construction defect, trucking) share the same shape: long documents, layered coverage, multiple insureds and carriers, and outcomes that depend on reading the contract and the facts together. The common thread is that these are reading-heavy, judgment-heavy files. Automate the reading.

Why the data layer underneath decides whether claims AI works

Claims AI fails most often not because the model is weak but because it cannot see across the systems a claim touches. Policy data lives in the policy admin system (Guidewire PolicyCenter, Duck Creek Policy, or a legacy mainframe). Claims live in ClaimCenter or a TPA's system. Billing sits elsewhere. Documents live in an ECM like OnBase or in shared drives and email. Broker submissions sit in an agency system like Applied Epic. Vendor data (medical bill review, independent adjusters, defense counsel billing) arrives as feeds and PDFs. For MGAs and TPAs, the policy may live at the carrier and the claim at the TPA, with a bordereau spreadsheet as the only bridge.

These systems do not share IDs, they define the same things differently (is "insured" the named insured, the additional insured or the driver?), and they update on different clocks. An agent that cannot reliably answer "is this the same claimant as the one in the 2023 slip-and-fall?" or "which version of this policy was in force on the date of loss?" will produce confident, wrong output.

The fix, explained without the jargon:

• Connect the systems you already run rather than replacing them. The claim system stays the system of record. The intelligence layer reads from it and writes back recommendations through its normal workflow.

• Resolve the entities. Decide, with rules your team can audit, when two records refer to the same person, provider, vehicle, location, policy or law firm. This is called entity resolution, and it is what makes fraud rings, repeat claimants and subrogation targets visible.

• Model the relationships. A knowledge graph stores how things connect: this claimant, treated by this provider, represented by this firm, on this policy, with this endorsement, in this venue. Agents reason over that map instead of guessing from one document at a time.

• Keep lineage. Every agent output should trace to the source page and system it came from, which is what makes it auditable for supervisors, reinsurers and regulators.

If this sounds like a two-year platform project, it does not need to be. Start with the systems one claim workflow touches and add from there. We wrote more about this in what an AI context layer is.

What should you measure, and what is the ROI of claims automation?

Measure one countable unit you can baseline today and check monthly. If you cannot name it, you are not ready to automate. The useful candidates for claims:

• Loss adjustment expense per closed claim (allocated and unallocated, by line).

• Days from FNOL to a reserve within 20% of ultimate incurred (reserve accuracy and speed in one number).

• Subrogation recoveries per 1,000 closed claims, and the share of closed files screened for recovery.

• Days from letter of representation received to linked and acknowledged.

• Open files per adjuster and adjuster hours per claim on the targeted claim type.

• Leakage rate on closed-file audit, measured the same way before and after.

A worked example (hypothetical numbers)

Take a hypothetical regional commercial auto and general liability carrier closing 12,000 claims a year with $180 million in paid losses. Assume these illustrative inputs, which you should replace with your own:

1. Subrogation. A closed-file audit finds 3% of closed files had a missed recovery opportunity, with an average recoverable amount of $8,000 and a 50% realistic collection rate. If an agent screens every file and surfaces two-thirds of those opportunities, that is 12,000 x 3% x 2/3 x $8,000 x 50% = $960,000 a year in added recoveries.

2. Adjuster time. Adjusters spend an average of 2.5 hours per claim on reading, indexing and summarizing documents. Cutting that by 40% on 12,000 claims saves 12,000 hours a year, roughly six full-time adjusters' worth of capacity at about 2,000 hours each. That capacity goes to the complex files, not to layoffs, which is how it shows up as lower severity rather than lower payroll.

3. Earlier attorney contact. If faster linking of letters of representation and earlier outreach on high-propensity claims shaves even 1% off bodily injury severity on a $60 million BI book, that is $600,000.

In this hypothetical, the identifiable annual benefit is roughly $1.5 million in hard dollars plus six adjusters of capacity, before any change in leakage on coverage or medical bills. Set that against what the program costs you (licenses or build, integration, change management, and your team's time) and you have a business case a CFO can test. The point of the exercise is not these numbers. It is that each line ties to a unit you can count before and after.

One warning from experience: delivered capability is not realized ROI. If adjusters do not open the summaries, if supervisors ignore the triage scores, if the subrogation queue has no owner, the software works and the money does not move. Plan for adoption the same way you plan for integration. We have written about why AI pilots stall before they reach EBIT.

How to start: a 90-day plan for claims automation

Start with one claim type, one countable unit and one team, and prove it in 90 days before expanding. The status quo is the real competitor here: most claims departments do not fail at AI, they just never decide.

Days 1 to 30: pick the target and baseline it

1. Choose one workflow where the reading load is high and the outcome is measurable. Good first candidates: subrogation screening on closed auto and property files, letter-of-representation detection on BI claims, or medical chronology for MPL incident files.

2. Name the countable unit and pull 12 months of baseline data from the claim system.

3. Run a small closed-file audit (100 to 200 files) to measure the current miss rate by hand. This becomes your ground truth.

4. Write the decide/draft/flag/human-only matrix for this workflow and get claims leadership and compliance to sign it.

5. Map the systems this workflow touches and who owns each.

Days 31 to 60: connect and run in shadow mode

1. Connect the claim system, the policy file and the document store for the targeted claim type. Resolve the core entities (claim, claimant, policy, insured, provider, firm).

2. Run the agent in shadow mode against live and recently closed claims. It produces flags, drafts and summaries that the team reviews but does not act on yet.

3. Have two or three senior adjusters grade outputs weekly: right, wrong, useful, noisy. Tune thresholds based on their grades, not the vendor's defaults.

Days 61 to 90: go live with one team and measure

1. Turn on the workflow for one unit. Outputs land inside the claim system as diaries, notes or tasks, not in a separate tool adjusters must remember to check.

2. Track the countable unit weekly against baseline, plus adoption: how many flags were opened, accepted or rejected, and why.

3. At day 90, decide with numbers: expand to the next team or claim type, fix and rerun, or stop.

Build or buy?

Larger carriers with a real data engineering team can build parts of this on their core platform and a model provider. Most MGAs, TPAs, self-insured programs and mid-size carriers will get there faster buying the data layer and agents and configuring them to their own claim practices. The honest test: if you do not already have clean, connected claim, policy and document data, building agents first is building on sand, whichever route you take.

Frequently asked questions

Do insurance companies use AI for claims?

Yes, widely in pilots and narrowly in production. Bain & Company's 2025 assessment found 78% of P&C insurers use generative AI in some capacity but only 4% have scaled it, and just 27% are pursuing a full claims transformation.

What is FNOL and can it be automated?

FNOL is the first notice of loss, the initial report of a claim. It is one of the easiest parts of claims to automate: capturing loss details from any channel, matching to the right policy and insured, deduplicating reports and routing the claim. The harder work comes after FNOL.

What percentage of claims can be processed straight through?

It depends heavily on line of business. Simple glass, towing, small property and some specialty personal lines can go straight through at high rates. Bodily injury, general liability, workers' compensation lost-time and professional liability claims generally should not, and chasing STP on those lines is the wrong goal.

How does AI triage insurance claims?

AI triage scores each claim on expected severity and complexity using FNOL facts plus what the documents reveal in the first weeks (injury type, treatment pattern, venue, attorney involvement, limits), then recommends routing to fast track, desk, field, complex loss, nurse review or SIU. A human confirms borderline routing.

Can AI predict which claims will involve an attorney?

It can score likelihood based on injury, venue, claimant contact patterns and delays, and it can detect letters of representation the day they arrive. That matters because the Insurance Research Council found represented BI claimants waited more than twice as long for closure as unrepresented ones.

How can AI help with subrogation?

AI reads every open and closed file for third-party liability facts (another driver at fault, a contractor, a product failure, a landlord) and queues candidates for a recovery specialist. Its biggest value is coverage: it can screen every file, where most closed-file reviews only sample a fraction.

Is it legal for insurers to use AI in claims decisions?

Yes, with governance. The NAIC Model Bulletin, adopted in 24 states as of March 2025 per Quarles, expects a written AI program, documentation, vendor oversight and attention to human involvement in final decisions. Some states have additional rules, especially in health claims.

How long does it take to see ROI from claims automation?

A focused first workflow can show measurable results within 90 days if you baseline a countable unit first. Recovery-focused workflows like subrogation screening tend to show dollars fastest because the benefit is a direct cash recovery.

Is your platform compliant with HIPAA and SOC 2?

OutcomeCatalyst is HIPAA-aligned and SOC 2 Type 2 aligned, with the formal audit underway and expected to complete before year-end. MPL and health-related claim files carry protected health information, so ask any vendor the same question.

Sources

• Bain & Company, 2025 claims maturity assessment, as reported in "Most P&C insurers still on pilot mode with gen AI" (Insurance Business, December 2025): insurancebusinessmag.com

• Insurance Research Council, "Auto Injury Insurance Claims: A Study of Increasing Claim Severity" (summary): insurance-research.org

• U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, "Claims Adjusters, Appraisers, Examiners, and Investigators": bls.gov

• National Association of Subrogation Professionals estimate, as cited in Genpact, "Maximize subrogation opportunities to strengthen the bottom line": genpact.com

• AM Best MPL market segment report, as reported in "MPL market faces 10th straight year of underwriting deficit" (Insurance Business, May 2025): insurancebusinessmag.com

• TransRe verdict data, as reported in "Mega Malpractice Verdicts Against Physicians on the Rise" (MDedge): mdedge.com

• Swiss Re, as reported in "Swiss Re's Business Unit CEOs Share Views on Key Risk Themes of 2025" (Carrier Management, January 2025): carriermanagement.com

• Coalition Against Insurance Fraud, "The Impact of Insurance Fraud on the U.S. Economy" (2022): insurancefraud.org

• James Ballot and Diane Brassard, "Claims Leakage Costs Insurers Billions Annually" (Insurance Thought Leadership, July 2026): insurancethoughtleadership.com

• Milliman, "Industry survey: Claims department challenges and artificial intelligence" (citing Rising Medical Solutions' 2022 Workers' Compensation Benchmarking Study): milliman.com

• Quarles, "Nearly Half of States Have Now Adopted NAIC Model Bulletin on Insurers' Use of AI" (April 2025): quarles.com

OutcomeCatalyst connects the systems you already run into a governed intelligence layer your team and your agents can reason over. Demos on this site use fictional data. To see this on your own book, start a conversation.