Claims Automation: The Complete Guide to Automating the Claims Lifecycle in 2026

Claims automation is the execution of claim work by software and AI agents across the full lifecycle - intake, triage, coverage analysis, evidence gathering, valuation, settlement, payment, and recovery - with a licensed human making every step that is genuinely a decision. The useful definition is not fewer people. It is that every step which is a lookup, a comparison, a reconciliation, or a document read gets done by a machine in minutes, and every step that is a judgment stays with an adjuster who now has the evidence assembled in front of them instead of spending the week assembling it.

Most definitions of insurance claims automation are a parts list - robotic process automation, OCR, intelligent document processing, NLP, machine learning - and a parts list describes what the software is made of rather than what it finishes. Two products with identical technology lists can own completely different amounts of a claim. The definition that survives an evaluation is a definition by stage and by decision right: which stage does this reach, and which calls does it leave to a person. Both views are below, the stack layer by layer after the stage table.

Three things get sold as claims automation and are not. A chatbot at first notice of loss is intake, not automation of the claim. Straight-through processing on the simplest fifth of the book is real and valuable, but it is a rule about which claims to skip, not a capability that works the hard ones. A workflow engine that routes tasks between the same humans doing the same work has automated the routing and nothing else. Each of the three helps. None of them touches the stages where cycle time and loss cost actually accumulate.

It is worth separating three levels, because vendors and carriers use one word for all three. Task automation does one step: extract the fields from an ACORD form. Stage automation owns an entire stage end to end: everything triage does, from severity scoring through assignment. Lifecycle automation runs every stage against one shared evidence record, so nothing is re-keyed and no stage starts by rebuilding context the previous stage already had. Most carriers in 2026 sit at task and stage. The distance between stage automation and lifecycle automation is where the remaining money is, because two automated stages that do not share state still hand off through a human.

For the narrower snapshot of which specific tools are shipping today and which parts of the pipeline still have no vendor, see insurance claims automation in 2026: what is automated, what is still manual. This guide is the structural version of that map.

A claim is two things running at once: a clock and a ledger. The clock is cycle time, from the loss until the file closes, and it is what the policyholder experiences. The ledger is loss cost, every dollar that leaves on that claim, and it is what the underwriting result experiences. Almost every claims automation decision is a bet on one of the two. The good ones move both, because the same delays that stretch the clock are usually the same gaps that leak the ledger.

The clock is getting worse in property. The J.D. Power 2024 U.S. Property Claims Satisfaction Study put the average property claim cycle time at 23.9 days, more than six days longer than its 2022 study, with catastrophe claims at 34.2 days. The 2026 study shows some improvement, but the average homeowner still waits 40.7 days from first notice of loss to final payment. That is the line where a slow claim means a family is not back in their house.

Auto is the counterfactual that proves automation works when it is aimed at the right stage. J.D. Power reported auto repair cycle time peaking at 23.1 days in 2023 and falling to 19.3 days by 2025. Auto is the line where estimating, parts procurement, and payment were automated hardest and earliest. Where the automation landed, the clock moved. Where it did not, it did not.

Cycle time is not a soft metric that only shows up in survey scores. Accenture found that 30% of dissatisfied claimants switched carriers, with 60% citing settlement speed and 45% citing the closing process. A slow claim is a cancelled policy on a delay, and the acquisition cost to replace that policyholder is spent by a different department than the one that caused the delay.

The ledger is quieter and larger. EY's Property and Casualty Claims Transformation practice puts claims leakage - the gap between what a claim should have cost and what it did - at 7-14% of carriers' total claims spend, based on its claims quality assessments. On $500M of annual claims spend, that range works out to $35M-$70M a year. That is arithmetic on EY's published range rather than a benchmark for any specific carrier, but it is the right order of magnitude to hold in your head when a claims automation business case is being argued in basis points. For the mechanics of how leakage forms, see our pillar on claims fraud leakage and the practical version in reducing claims leakage. To measure it on your own book, file by file and coded to a stage, use the claims leakage audit checklist, which also explains why a sampled leakage rate describes only the files that were opened.

Underneath leakage sits fraud. The Coalition Against Insurance Fraud put the cost of insurance fraud in the United States at $308.6 billion a year in its 2022 study. Fraud is not a separate problem from leakage; it is the part of leakage that has intent behind it, and it is the part that requires an investigation rather than an audit to resolve.

The direction of travel is not in dispute. McKinsey's Claims 2030 work projects that more than half of claims activities could be automated by 2030. The question a claims leader has to answer in 2026 is not whether, it is which half, and in what order.

The claims lifecycle runs eight stages. Different carriers name them differently and some split or merge a stage, but the work is the same everywhere. The table below is the summary of this entire guide: what is genuinely automatable at each stage in 2026, and what is still a human's job.

Read the right-hand column carefully. It is short, and almost every entry in it is a decision rather than a task. That is the shape of an honestly automated claims operation: the machine does the work, the human makes the calls, and every call arrives with the evidence already assembled underneath it. Any vendor whose right-hand column is empty is describing a product that does not exist and a regulatory posture that will not hold.

Most definitions of insurance claims automation are a parts list: robotic process automation, optical character recognition, intelligent document processing, natural language processing, machine learning, computer vision. The list is accurate and close to useless for deciding what to buy, because those technologies do not do the same kind of work and do not reach the same stages. Two products with identical technology lists can own completely different amounts of a claim. It is more useful to read the category as four layers, in the order they arrived.

Read that table by what each layer assumes. RPA assumes the work is already fully specified and the only cost is a human doing the typing. Document AI assumes the bottleneck is reading a page. Scoring assumes the bottleneck is knowing which claims deserve attention. Rules assume the decision reduces to conditions someone can write down beforehand. All four assumptions hold at the front of the lifecycle, which is why stages 1, 2, and 5 automated first and automated well.

None of the four holds at stage 4. Investigating a flagged claim is not re-keying, not a single document, not a score, and not a rule. It is a sequence of dependent lookups and comparisons in which each result changes what the next step should ask: the police report moves the timeline, the timeline contradicts a statement, the contradiction sends you back to the provider's billing history. That is the shape of work the fourth layer does and the first three cannot, and it is the reason a lifecycle assembled entirely from the first three still stalls for 14 or more days in the middle of itself.

Stage 1: FNOL and intake

First notice of loss is the most automated stage in the industry and the one carriers are proudest of, which is exactly why it is worth being precise about what the automation bought. Digital FNOL moved intake from a phone call transcribed by a person into structured data captured at the moment of loss. That matters less because it saved the call and more because everything downstream now starts from clean typed fields instead of free text that each later stage has to re-read.

• Extraction from ACORD forms, adjuster emails, carrier and agent portals, loss photos, and voice transcripts directly into the claim record
• Duplicate-claim and prior-loss checks at first touch, before a file number is assigned
• First-pass severity and complexity flags that determine whether the claim needs a human at all
• Automated acknowledgement and document-request letters, which is where a measurable slice of the cycle-time clock is won or lost in the first 48 hours

What does not automate at intake is the loss nobody modelled. A claim that does not fit a template should be handed to a person immediately rather than forced through one. A good intake automation is measured by how fast and how visibly it escalates, not by how much it swallows.

Stage 2: Triage and assignment

Triage classifies each claim by severity, complexity, litigation potential, and fraud risk, then routes it - straight through if it is clean and small, to a specific adjuster by license and workload if it is not, to SIU if it scores as suspicious. Carriers have largely solved this. Scoring that took an experienced adjuster twenty minutes of reading now happens in seconds, and the routing rules are auditable. The deep version of this stage is in the claims triage automation playbook.

The failure mode here is specific, common, and expensive: better triage produces more referrals, and the teams receiving those referrals did not grow. A carrier that improves fraud triage without expanding investigation capacity has bought a longer queue, not a lower loss cost. This is the mechanism behind why most flagged claims are never fully investigated, and it is the single most under-modelled second-order effect in claims automation programs. Straight-through processing compounds it: every claim that closes untouched has moved past the last person who might have looked, so the flagged lane concentrates rather than shrinks as the rest of the lifecycle speeds up. That dynamic is the subject of automated claims investigation, the step that grows as everything else speeds up.

Stage 3: Coverage verification

Coverage verification is the stage where language models changed what is technically possible. A policy with a stack of endorsements is a long document that has to be read against one specific set of loss facts. That is now genuinely automatable: read the form and every endorsement, apply them to the facts on the file, produce a coverage position with each element cited to the clause that governs it, and draft the reservation-of-rights or coverage letter that follows. What used to be a half-day of a senior adjuster's reading is now a first draft in minutes.

What stays human is the decision and the ambiguity. A coverage denial is a regulated act with unfair-claims-practice exposure attached to it, and ambiguity that turns on jurisdictional case law is not a document-reading problem. The correct output of an automated coverage stage is a cited draft position, not a verdict. If a vendor's coverage module issues denials, that is not a feature, it is a liability.

Every other stage in this guide has a software incumbent. Intake has FNOL platforms. Triage has scoring engines. Estimating has CCC, Mitchell, and Solera. Payment has payment rails. Detection has FRISS, Shift Technology, and Verisk. Investigation has no incumbent. At most carriers, in 2026, the answer to what happens after a claim is flagged is still a person, a case file, and 14 or more days.

The arithmetic is unforgiving. A typical carrier runs roughly one investigator per 200+ cases. A manual SIU investigation takes 14 or more days of that investigator's time when it is done properly - records to pull, statements to compare, a timeline to build, a report to write. In Hesper's own benchmarking, that ratio produces a full investigation on roughly 25% of flagged claims. The other 75% are paid, denied, or parked. Not because someone assessed them as clean, but because there was no capacity to find out. The backlog mechanics are in the claims investigator backlog guide.

This is why more detection has diminishing returns. Detection answers which claims to look at, and that has not been the constraint for years. FRISS, Shift Technology, and Verisk are good at what they do and they are complementary to an investigation layer rather than substitutes for it. A better score on a claim nobody has capacity to work does not change the outcome of that claim.

The reason investigation resisted automation for so long is that it looks like judgment from the outside. It is not, mostly. An investigation decomposes into discrete phases, most of which are evidence work rather than decision work, and most of which are independent enough to run at the same time:

• Evidence collection - claim documents, medical records, repair estimates, police and fire reports, prior claims through NICB and ISO ClaimSearch
• Document forensics - metadata, compression artifacts, and internal-consistency checks that establish whether a submitted document is what it claims to be
• Statement cross-referencing - comparing every account given by every party against each other and against the physical evidence
• Timeline reconstruction - a single chronology assembled from timestamps across every source, which is where most contradictions surface
• Open-source and public-records research - the publicly available information that confirms or contradicts the claim narrative
• Provider and network analysis - billing patterns, treatment protocols, and the shared addresses, phones, and attorneys that reveal a ring rather than a claim

Hesper runs 15+ of these phases in parallel and completes an investigation in minutes rather than 14+ days, with every assertion in the final report cited to the specific evidence that produced it. The compression comes from parallelism, not from cutting phases - a human investigator does one thing at a time because a human can only do one thing at a time. The architecture is covered in parallel processing across investigation phases and AI agent architecture for claims investigation.

What does not automate here is the determination, and it should not. A fraud finding carries a signature, and the person who signs it has to be able to defend it in a deposition, an examination under oath, or a suspicious activity report to a state fraud bureau. An EUO is a live proceeding. Credibility assessment of a person in a room is not an evidence-processing task. The agent assembles the case; the investigator decides it. See the defensibility standard for fraud investigation AI and the general counsel's view for what that signature has to survive.

Investigation is where this platform has unusual depth, and it is deliberately one stage of eight rather than the whole story. A carrier does not buy claims automation to get better fraud investigation; it buys claims automation to close claims faster and cheaper, and discovers that the stage blocking both is the one nobody automated. The full treatment of that stage is our pillar on autonomous AI claims investigation, with the operational view in SIU operations and the detection layer above it in insurance fraud detection.

Stage 5: Estimation and valuation

Estimating is the second most automated stage after intake, and auto is the proof. Photo-based damage assessment, parts pricing against live catalogs, and supplement adjudication are mature products with real accuracy on high-frequency, low-complexity damage. This is the clearest single explanation for the divergence in the J.D. Power numbers: auto repair cycle time fell from its 23.1-day peak in 2023 to 19.3 days in 2025, while property cycle time went the other way, rising to 23.9 days in 2024. Auto got the estimating automation. Property largely did not.

Property estimating is harder for structural reasons rather than technical ones. A vehicle is a catalog of known parts; a house is not. Interior damage is occluded, scope grows once a contractor opens a wall, and the variance between a photo estimate and a physical scope is material. What automates well in property is the first-pass scope, the line-item pricing, supplement review against the original estimate, and reserve recommendation. What does not is the total-loss call, complex structural damage, and anything that needs someone on the roof.

Reserve accuracy is the underrated part of this stage. A reserve set from a first-pass estimate and never revisited is a slow leak into the ledger, and it is the kind of drift that only shows up in a quarterly triangle. Automated reserve recommendation that re-runs whenever new evidence lands on the file is one of the least glamorous and highest-return automations available.

Stage 6: Settlement and litigation

Settlement is where automation is least mature and most often oversold. The negotiation itself does not automate. What automates is everything that feeds it: digesting a 400-page demand package into the facts that matter, benchmarking the valuation against comparable resolved claims, drafting the response, and flagging litigation risk early enough to change the strategy rather than react to it. The other side of the table automated first. Plaintiff firms now assemble demand packages with generative AI while carriers still examine them by hand, and average third-party submitted medical expenses on auto bodily injury claims reached $32,300 in Q1 2026, up from $24,300 in Q1 2023, per CCC. The asymmetry is covered in what AI-built demand packages do to claim severity.

This is disproportionately worth doing because evaluation and litigation are where the ledger bleeds hardest. In EY's case study of one top US P&C insurer, more than 85% of the leakage it assessed sat in three areas: coverage determination, litigation prevention, and evaluation and resolution. Anything that shortens the time between a demand landing and a substantive, well-priced response is working directly on the most expensive part of the book. The corollary is that a demand package sitting unread for three weeks is one of the most expensive queues a carrier operates.

Stage 7: Payment and closure

Payment is close to solved as a mechanical matter. Digital payment rails settle same-day, and the operational lift of issuing money is no longer a bottleneck anywhere. The remaining automation opportunity in this stage is not speed, it is the check before the money leaves: payee verification against the claim record, anomaly detection on amount, payee, and pattern, and a hold on anything that does not reconcile. A fast payment to the wrong payee is a worse outcome than a slow one.

Closure is the quiet stage and the one most likely to be automated badly. A file that closes without a complete record is a file you cannot defend in an audit and cannot recover on later. Accenture found that 45% of dissatisfied claimants cited the closing process, which makes closure quality a retention issue rather than a file-hygiene one. Automated closure should run a completeness audit - every decision cited, every document filed, every recovery avenue screened - before it marks anything closed. That evidence discipline is the same one described in generating an audit-ready report.

Recovery is the most under-automated stage in the lifecycle and the easiest to justify economically, because the dollars in question have already left the building. Subrogation and salvage are the only stage where automation recovers cash rather than avoiding cost, and the finance conversation is correspondingly short.

One industry estimate (SecondLook, 2021) puts missed subrogation at roughly $15 billion a year, with about 15% of claims closed without their recovery potential ever being identified. It is a vendor estimate rather than an audited figure. That second number is the automatable one. Identification is a screening problem, not a legal one: read the loss facts, determine whether a third party contributed, check whether the statute is still open, and flag it before the file closes. See the subrogation definition for the mechanics.

What automates in recovery is the screen on every closed file, the liability analysis behind a demand, the demand letter itself, arbitration filings, and statute-deadline tracking. What does not automate is the decision to litigate a recovery and the negotiation with the adverse carrier. The reason this stage stays manual at most carriers is not difficulty; it is sequencing. Recovery sits at the end, after the adjuster's attention has moved to the next file, and a screening step that depends on someone remembering to do it is a screening step that does not happen at scale.

The cost of that sequencing shows up in the industry data. Roughly one in four US property-liability insurers records no salvage or subrogation recovery in a given year (Bisco and Fier, Journal of Insurance Regulation, 2023). And because an Arbitration Forums panel decides on the evidence attached at filing, the documents that win a liability demand enter the file at FNOL, coverage and investigation, long before anyone opens a recovery queue. That is why the screen belongs at intake rather than at closure - see why missed subrogation opportunities are lost at intake, not at closure.

A claims automation evaluation is more informative when it is run backwards. Instead of asking what a platform can do, ask what it refuses to do and why. Five things should stay human, and a vendor that claims otherwise is either describing a roadmap or misunderstanding the regulatory surface.

1. The coverage decision. A machine can read every endorsement and cite the controlling language. Issuing a denial is a regulated act with unfair-claims-practice exposure, and it belongs to a licensed adjuster.
2. The fraud determination. The finding gets signed, and the signature has to survive a deposition, an EUO, or a referral to a state fraud bureau. The agent builds the case; the investigator concludes it.
3. Negotiation and settlement authority. Authority is delegated to people under a documented framework. Automating the preparation for a negotiation is high value; automating the authority is a governance failure.
4. Anything requiring physical presence. Scene inspections, roof scopes, in-person recorded statements, salvage evaluation on the lot. No amount of model quality substitutes for someone standing there.
5. The exception nobody modelled. Every claims book has a long tail that fits no template. The test of a claims automation program is not how much it handles - it is how fast and how visibly it escalates what it should not.

The regulatory floor under all five is about the record more than the reasoner. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023 and taken up by 26 jurisdictions as of 31 August 2026 per the NAIC adoption map, requires insurers to run a written AI systems program covering governance, risk management, and internal audit, and to extend that oversight to third-party AI vendors. It lists human involvement in final decisions as one of the factors that set how much control an insurer needs, not as a mandate, and the unfair-claims rules underneath it apply however a decision was made. The practical effect on claims automation is twofold: the carrier stays accountable for every decision its agents support, and the system has to be able to show which input produced which output on any given claim. The claim-file rules asked for that second part long before AI: NAIC Model Regulation 902 requires documentation that lets an examiner reconstruct the insurer's activities on each claim, as covered in what the claim-file rules asked for before AI. For the buyer-side version, see the compliance officer's deployment guide and how a chief risk officer evaluates claims AI.

New Jersey's A5494, a bill introduced on 14 September 2026, would bar AI from making the final decision to deny an auto, homeowners or flood claim, but it says nothing about what the person signing that denial must be handed. Florida's House passed a version with that evidence standard 108-0 in March 2026, and it died in the Senate - see what New Jersey's AI claim denial bill regulates, and what it leaves out. No property and casualty rule in force today reserves an investigation step to a human, so which calls stay human is a design decision, and the variable that matters is what evidence the reviewer sees at each checkpoint. That design question is worked through in which decisions stay human, and what they get handed.

Most of what is published about claims automation describes it as something the industry has already done. Two 2026 surveys describe close to the opposite: tools almost everywhere, and almost nobody running them at scale.

The WTW 2026 Advanced Analytics and AI Survey, published 19 March 2026, asked 59 property and casualty insurers in the United States and Canada where they actually are. Straight-through processing in claims workflow automation is in place at 14% of them, with another 36% planning to introduce it. Claims advanced analytics is used for fraud detection by 33% and for severity assessment by 29%, both of which respondents expect to reach 65-70% within two years. More than half already use large language models or generative AI somewhere. And the stated blockers are not model quality: 42% name data problems and inadequate IT support, only 20% report a well-defined analytics strategy, and only 12% regularly train employees on analytics. WTW also found the carriers furthest along ran combined ratios six points lower and premium growth three points higher than slower adopters between 2022 and 2024.

Sedgwick's property report, published 3 March 2026, measures the same gap from the maturity side. Between 58% and 82% of carriers use AI tools somewhere in their operations. 12% say they have fully mature AI capabilities. 7% say they have achieved scalable AI success. Nearly two-thirds acknowledge a gap between their AI vision and their reality, and 75% of claims professionals say AI needs human oversight. Sedgwick's own framing for the failure is the one that matters here: fragmented tools and inconsistent data, with AI layered on top of existing platforms rather than wired through them.

Put the two surveys next to each other and the conclusion is narrow but firm: adoption stopped being the constraint some time ago, and scale never started. That plateau is also older than AI. Deloitte's fourth Global Robotics Survey, reported in October 2018 from 530 business leaders, found just 4% of organisations running more than 50 software robots, barely up from 3% a year earlier, with process fragmentation named the single largest barrier to scale at 32%. Eight years and an entirely different technology later, Sedgwick's figure for scaled AI in claims is 7%, and the named barrier is still fragmentation. The technology changed. The thing that stops it from scaling did not.

Mapped onto the eight stages, the pattern is consistent. Automation scales where the work is uniform and the inputs are predictable, which is the front of the lifecycle, and it stalls where the work is a sequence of dependent judgements, which is stages 4, 6, and 8. A carrier with excellent intake automation and no investigation layer has not built half a lifecycle; it has built a faster road to the same queue. Our own stage-by-stage reading of where the industry sits is in the State of Claims Automation 2026 report.

The sequencing mistake is close to universal: carriers automate in the order vendors call on them, which is front to back, because the front of the lifecycle is where the demo is prettiest. Front to back is the wrong order. The right order is stalled stage first, because a lifecycle moves at the speed of its slowest stage regardless of how fast the others got.

1. Instrument before you automate. Measure cycle time stage by stage, not end to end. A single end-to-end average tells you nothing about which stage owns the days inside it. You cannot prioritize what you have not decomposed.
2. Fix the stalled stage, not the loudest one. If intake takes four hours and investigation takes 14 days, another hour off intake is noise. Rank stages by days contributed, not by how often people complain about them.
3. Automate the handoffs, not just the stages. Two automated stages that do not share state produce re-keying, and re-keying is where both time and errors reappear. Handoff automation is usually cheaper than stage automation and returns more.
4. Keep one evidence spine. Every stage should write to a single evidence record with source, timestamp, and provenance. This is what makes the audit trail a byproduct rather than a reconstruction project, and it is the only version that survives a DOI exam.
5. Put the human in the decision, not the workflow. If your target operating model still has adjusters approving routine routing steps, you automated the work and kept the clicking. Escalate exceptions to people; do not make people the transport layer.
6. Prove it on one line before scaling. Pick the line where the stalled stage hurts most, run it for a quarter against a measured baseline, and expand on evidence. Most AI claims programs stall in change management rather than model quality.

The change-management half of this is at least as hard as the technical half. The people-side playbook is in the claims operations manager guide to AI-augmented investigation, the budget and timeline version is in the Claims VP deployment playbook, and the finance case is in the CFO ROI memo. If you are trying to work out who at the carrier actually decides, see the buying center map. TPAs fund the program differently: the NAIC's TPA guideline permits fees based on the number of claims processed but bars fees contingent on the savings produced, so the case has to run per file - see claims automation for TPAs.

Most claims automation programs are measured on the metrics that are easy to move rather than the ones that matter. The set below is chosen so that each metric is hard to game and each one fails loudly when the automation is doing something cosmetic.

Two of these deserve emphasis because they are the ones carriers most often do not track. Flagged-claim coverage is the difference between buying detection and buying resolution, and a carrier that cannot state its coverage percentage does not know what its detection spend is producing. Reopen rate is the honesty check on everything else: it is trivially easy to improve cycle time by closing files prematurely, and reopen rate is the metric that catches it a quarter later. The fuller metric treatment is in SIU KPIs for 2026 and how to measure AI ROI in claims.

The other reason claims automation is hard to evaluate is that the published numbers do not compare. Collect the quantified claims from the pages that rank for this topic and you get, in one sitting: 100% accuracy and a 60% decrease in processing time from Automation Anywhere, citing an EXL case study at a US commercial P&C insurer; 95% faster claims at Trygg-Hansa and 60 days down to 3 at Select Health from SS&C Blue Prism; over 99% extraction accuracy and roughly 70% faster turnaround from Klippa and Doxis; 25-30% per-claim savings from Strada, presented as an example with no source attached; and 80% faster processing of low-severity claims plus intake falling from 10 days to 36 hours from Sedgwick. Every one of those may be true. None of them is comparable to any other, because none states what it was measured over.

One failure mode runs under all five rows. Any metric expressed per claim improves automatically when the claims you automate are the simple ones, even when nothing about the hard ones changed. A program that moves the cleanest 20% of the book to straight-through processing will report a lower average cost per claim, a shorter average cycle time, and a higher accuracy rate, and the 14-day investigation queue will be exactly where it was. That is not dishonesty on the vendor's part. It is a denominator nobody asked about.

The same standard applies to this guide. The flagged-claim coverage figures here are Hesper's own benchmarking rather than an industry survey, and they are labelled that way wherever they appear. Everything else carries a publisher and a date you can check: J.D. Power on cycle time, EY on leakage, Accenture on switching, the Journal of Insurance Regulation on recovery, WTW and Sedgwick on adoption. When a number in this category has neither a publisher nor a date attached, it is a brochure, and it should be priced as one.

Claims automation is sold as one category and bought as at least five. The vendors overlap far less than the shared label suggests, and most evaluation confusion comes from comparing products that occupy different stages. They are not spread evenly either: a stage-by-stage count of 2026 vendors finds seven serving intake and three at subrogation, the inverse of where leakage concentrates - see the best AI claims processing software in 2026, mapped by stage.

• Core claims systems - Guidewire, Duck Creek, Majesco, Snapsheet. System of record and workflow across all eight stages, deep in none of them. Their AI modules are stage features on top of a platform decision you already made.
• Estimating and repair networks - CCC, Mitchell/Enlyte, Solera. Stage 5, overwhelmingly in auto, and genuinely excellent at it. Not investigation, not coverage, not recovery.
• Detection platforms - FRISS, Shift Technology, Verisk. Scoring and matching that feeds stages 2 and 4. They tell you which claims deserve attention. They do not work the claim.
• Document AI - Ocrolus, Hyperscience, ABBYY. Extraction and classification that feeds every stage. Reading a document is not adjudicating it.
• Investigation and resolution - Hesper AI. Stage 4 at depth, wired into the stages on either side of it.

Hesper AI is the claims resolution platform: agents that run the lifecycle from first notice to final recovery, with a full investigation engine inside stage 4 rather than a referral out of it. Detection is built in, so it runs standalone or downstream of FRISS, Shift Technology, or Verisk - all three remain complementary. The distinguishing claim is not that it touches more stages than a core system does. It is that the one stage every other vendor routes around is the one it is built to finish. For the platform comparisons, see claims management systems compared, AI fraud platforms compared, and the hidden integration costs of legacy claims AI modules. Whichever vendor you shortlist, settle custody of the evidence record before you negotiate license cost: a Rhode Island claim closed in 2026 must stay producible through 2030, after a three-year contract signed the same year has ended. The exit and portability clauses to agree before signature are in build vs buy claims AI.