Connected Claims USAFind us at booth 502, Signia by Hilton Orlando, Nov 3-4Booth 502 · Orlando, Nov 3-4Booth 502 · Connected ClaimsBook a meeting
Blog›Guides
GuidesNovember 15, 2026·13 min read·Nitish Badu·Updated November 16, 2026

The claim file review checklist, in the order the file gets tested

A claim file is the evidence that the decision was defensible, read by people who were not there. Here is the review checklist in the order four readers test it, and what fails in published exams.

NB
Nitish Badu · COO and Co-founder
November 15, 2026·13 min read·Updated November 16, 2026
The numbers behind this
17items on the published Lloyd's market claims file review listLMA Coverholder Audit Scope, Version 3.0, Q4 2019. The same document names six rules for choosing which files get pulled and reports the claims testing summary across 11 sections, A to K, as a weighted average percentage
28%error ratio on the collision claims sample in a 2025 market conduct examPennsylvania Insurance Department examination of Allstate Fire and Casualty Insurance Company, issued November 4, 2025: 31 violations cited across 21 of 75 files, sampled from a population of 39,918 collision claims. The highest claims error ratio in that exam
$4,675,000in fines across 270 state enforcement actions against P/C insurers for claims violations in 2019Insurance Journal, October 8, 2020, reporting Wolters Kluwer Compliance Solutions analysis. Improper or incomplete documentation of claim files ranked ninth among the top ten market conduct criticisms
3.45%omission rate in an LLM summarisation pipeline, against 1.47% hallucinationAsgari et al., npj Digital Medicine, 2025. Hallucinations ran at 1.47% across 12,999 clinician-annotated summary sentences; omissions ran at 3.45% of the source transcript sentences. Clinical documentation, not claims

A claim file is not a record of the claim. It is the evidence that the decision was defensible, and it is read by people who were not there. So a useful review checklist follows the sequence a later reader works through, not the sequence the claim followed.

The checklist below runs in test order. Before it come the sampling rules that pick which files get pulled, and what actually fails in published examinations, which is almost never the decision. The lifecycle this review cuts across is mapped in our guide to automating the claims lifecycle.

A claim file is not a record of the claim

Answer

What is a claim file review checking for?

A claim file is the evidence that the decision was defensible, assembled for readers who were not there. A review tests that evidence item by item and returns a pass or fail on each one. That is why the checklist is ordered by what gets tested later, not by the order in which the claim happened.

The operative standard is one sentence long. Commonwealth Automobile Reinsurers Appendix M (2021 revision) reproduces the NAIC claims examination standards, and one line appears under three separate headings: "The file must clearly document the basis for the decision and result."

Read that line as a reviewer and the job inverts. It does not ask whether the decision was right. It asks whether the basis for it is on the page. A correct denial with no stated ground fails.

The artifact standard is reconstruction. California's file and record documentation rule requires all documents, notes and work papers in enough detail that pertinent events and their dates can be reconstructed and the licensee's actions determined. The real test of any file is a stranger holding it two years after the adjuster who worked it left. More on that in what the claim-file rules asked for before AI.

Four people read this file after you close it

Answer

Who reads a closed claim file, and what does each reader test?

A closed claim file is read by a market conduct examiner, a client or capacity partner's claims auditor, opposing counsel in a coverage or bad-faith dispute, and your own QA. Each one opens a different page first and applies a different standard, and the checklist is the union of their four tests.

The review categories are published. Washington wrote them into regulation: WAC 200-150-050, effective November 22, 2015, names what a claims audit covers, including coverage, reserves, file documentation, settlements and subrogation.

The delegated-authority market publishes the method too. The LMA Coverholder Audit Scope, Version 3.0 lists 17 items for a claims file review and reports the result across 11 sections as a weighted average percentage. For an MGA head of claims, that percentage is what a capacity partner reads before renewing capacity.

ReaderWhat they open firstStandard appliedWhat fails
Market conduct examinerDenials and files closed without paymentState unfair claims settlement practices rules, read against the NAIC claims standardsThe basis of the decision is not stated in the file; a required written notice is absent
Client or capacity partner auditorA risk-weighted sample: nominal reserves, small reserves with large costs, denials, files with recoveriesA published scope such as the LMA Coverholder Audit Scope: 17 file review items reported across 11 sectionsAuthority limits and referral triggers; reserve accuracy; subrogation never identified
Opposing counselThe claim log and the datesThe state claim-file documentation rule, plus the carrier's own claim guidelinesGaps in the chronology; decisions with no recorded reason
Internal QAWhatever the scorecard samples that monthThe carrier's own rubric, usually mapped to the NAIC standardsWhatever the adjuster had no time to write down

For the examiner the stakes are public. Insurance Journal reported Wolters Kluwer Compliance Solutions analysis showing that in 2019, state regulators issued 270 enforcement actions against property-casualty insurers for claims violations, carrying $4,675,000 in fines.

Which files get pulled

Answer

How do reviewers decide which claim files to sample?

Reviews run on samples, and the sample is chosen on risk rather than at random. The Lloyd's market scope names six selection rules: spread across the period and the authorities granted, small reserves with large costs, nominal reserves, description keywords showing unusual activity, denials, and files with recoveries identified.

Nobody reviews the whole book. The Lloyd's market scope names six pull rules, and if you know them you know which of your files are being judged on behalf of all the others.

  • Spread. A sample across the period under review and every authority granted.
  • Small reserves, large costs. Where reserving discipline shows.
  • Nominal reserves. The classic sign that nobody has re-read the exposure since it was set.
  • Keywords in the description. Wording that indicates unusual activity.
  • Denials. Where the basis of the decision has to be in the file and most often is not.
  • Recoveries. Files where a recovery or subrogation interest was identified.

The sample can be thin. In the Pennsylvania Insurance Department's market conduct examination of Allstate Fire and Casualty Insurance Company, issued November 4, 2025, the automobile collision sample was 75 files drawn from a population of 39,918 claims. The examiners cited 31 violations across 21 of those files and reported an error ratio of 28%.

One more cited file moves that published ratio by more than a point. And the ratio is what decides whether a finding reads as isolated or as a practice.

01Sampling math

75 files set the published finding for 39,918 claims

An examiner samples on risk, so the error ratio that reaches the public report is computed on a sliver of the book.

  1. Population39,918Automobile collision claims in the exam period
  2. Sample pulled75Collision claim files the Department reviewed
  3. Files cited21Carrying 31 violations between them
Files with a violation21Collision sample
÷
Files sampled75
=
Error ratio reported28%Highest claims ratio in that exam
Read it asA share of files carrying at least one violation, not a share of claims paid wrong.
Source: Pennsylvania Insurance Department, Market Conduct Examination Report of Allstate Fire and Casualty Insurance Company, as of September 8, 2025, issued November 4, 2025, Section VI, Claims. The collision sample was 75 files drawn from a population of 39,918; 31 violations were cited across 21 of them, an error ratio of 28%. Ratios in the same exam ran 15% on comprehensive, 11% on automobile property damage and 1% on total loss.

A leakage audit and a file review ask different questions

A leakage audit asks whether too much was paid. A file review asks whether the decision can be defended. The same file can pass one and fail the other, so a clean leakage result is not an audit-ready file. The money version is in our claims leakage audit checklist.

The review order: coverage, investigation, quantum, conduct, closure

Answer

In what order should a claim file be reviewed?

Coverage first, because every later finding is void if the coverage position was wrong. Then whether the investigation was proportionate to the exposure, then whether the number is supported, then how the insured was treated and told, then whether the file closed cleanly and the recovery was screened.

Workflow order and review order are not the same sequence. A reviewer works gates, and each gate decides whether the next one is worth testing. A quantum finding on a claim that was never covered is not a finding, it is a second problem.

02The review order

Coverage is tested first because every later finding dies with it

Five gates in the order a reviewer applies them, with the artifact each one has to find in the file.

The closed file
Policy in force on the date of lossContact and diary logEstimates, bills and payment recordsLetters and written notices
Each gate decides whether the next one matters
  1. 01CoveragePolicy version, provision relied on, position stated in writing
  2. 02InvestigationProportionate to exposure, with the checks recorded either way
  3. 03QuantumEvery adjustment carries its stated basis in the file
  4. 04ConductContacts, timing and the written notices the rules require
  5. 05ClosureClosed with notice given and the recovery screen recorded
Pass or fail recorded per item, with the artifact named
Review resultPass or fail, item by item
  • Artifact named
  • Gap named
  • Repeat finding flagged
Why this orderA quantum finding on a claim that was never covered is not a finding. It is a second problem.
Gates assembled from three published sources: the LMA Coverholder Audit Scope, Version 3.0, Q4 2019, which lists 17 items for a claims file review and reports 11 scored sections; WAC 200-150-050, effective November 22, 2015, which names twelve claims-administration categories; and the NAIC claims examination standards reproduced in Commonwealth Automobile Reinsurers Appendix M, 2021 revision. The grouping and the wording of the gates are ours.

The fifteen questions below are grouped into those five gates. The grouping and the wording are ours.

  1. Coverage. Is the policy version in force on the date of loss in the file?
  2. Coverage. Is the coverage position in writing with the provision relied on, and a reservation of rights issued where needed?
  3. Coverage. Where coverage was limited or declined, does the file carry the factual and legal basis rather than the conclusion alone?
  4. Investigation. Was the investigation proportionate to the exposure, and can a reader see what was checked?
  5. Investigation. Are the checks that came back clean recorded, or only the ones that found something?
  6. Investigation. Are contacts, statements and third-party records dated, with no unexplained gaps?
  7. Quantum. Does every adjustment for betterment, depreciation or salvage carry its justification in the file?
  8. Quantum. Where a settlement rested on a written scope or estimate, was the insured supplied a copy of each document, on the record?
  9. Quantum. Are reserves supported by what the file knew on the date they were set, and is each change reasoned?
  10. Conduct. Was first contact made inside the required timeframe, and is every required written notice in the file?
  11. Conduct. Where a decision ran past a statutory clock, is the delay explained in writing on the cadence the rule sets?
  12. Conduct. Did the handler act inside the authority granted, and are referrals above authority documented?
  13. Closure. If the claim closed without payment, is the written notice to the claimant in the file?
  14. Closure. Was the file screened for subrogation and salvage, with a negative result recorded?
  15. Closure. On a recovery, was the insured's deductible reimbursed, and is the calculation shown?

Every item is there because somebody later tests it. Items seven and eight come straight from California exam citations, and item thirteen from a Pennsylvania examiner's concern. Item eight is also where most contested property files start, as property claim estimate disputes shows. What the file has to hold when a claim turns into a referral is in the SIU referral package.

Review gateThe question askedWhat fails review
CoverageWas there coverage, on which version of the policy, and is the position in writing?A coverage position with no provision cited, or a policy version absent from the file
InvestigationWas the work proportionate to the exposure, and is it recorded?Checks performed but never written down, and clean results recorded nowhere
QuantumIs the number supported, and did the insured get what the number rested on?Depreciation or betterment with no stated basis; an estimate the insured was never supplied
Conduct and communicationWas the insured contacted, told, and told again on time?A missing written notice, or a delay with no written explanation on the required cadence
Closure and recoveryDid the file close cleanly, and was recovery screened?A closure without payment and without notice; a subrogation interest with no recorded disposition

What fails: the finding is usually an absence

Answer

What do claim file reviews most often find?

In published examinations the finding is rarely that the decision was wrong. It is that something required is missing from the file: the basis for an adjustment, a copy of the document a settlement rested on, a written notice at closure. A decision can be right and still fail review.

The Pennsylvania examination sampled four automobile categories at 75 files each. The error ratios came out twenty-seven points apart.

Claims error ratios by category, Pennsylvania market conduct examination of Allstate Fire and Casualty, issued November 4, 2025

Automobile collision (75 files sampled)28%
Automobile comprehensive (75 files sampled)15%
Automobile property damage (75 files sampled)11%
Automobile total loss (75 files sampled)1%

The examiners also recorded a concern rather than a violation: in two property damage files and four collision files, the claim closed without payment and no written notice went to the claimant. Nobody argued the claims should have paid. The file could not show the claimant had been told.

California's examination of the California FAIR Plan Association produced the same pattern in sharper form. Thirteen violations were cited because the association failed to document in the claim file all justification for adjustments for betterment, depreciation or salvage. Seven more were cited for settling on a written scope or estimate without supplying the insured a copy of each document the settlement was based on.

The report then records that the same documentation failures had been identified in the previous examination and were identified again in the current one. That is the finding no claims leader wants in print: not an error, a practice.

The exam finding is almost never that you decided wrong. It is that you cannot show you decided right.

What an AI summary gives you, and what a review still has to test

Answer

Can an AI summary replace a claim file review?

No, because they perform opposite operations. A summary compresses what is in the file. A review has to establish what is not in it and whether what is there supports the decision taken. Compression is the wrong operation on an absence, which is why summarisation research measures omission separately from hallucination.

Belem and colleagues, in From Single to Multi: How LLMs Hallucinate in Multi-Document Summarization (arXiv:2410.13961, October 2024, revised April 2025), asked models to summarise topic information that did not exist in the source documents at all. gpt-3.5-turbo still produced a summary 79.35% of the time and GPT-4o 44% of the time. An instrument that answers confidently about something absent is the wrong instrument for a question about absence.

03Summary vs review

A summary compresses. A review tests. These are opposite operations.

Both are useful. Only one of them returns a finding about something that is not in the file.

On what is in the file
On what is missing
Output
AI summaryCompression
On what is in the fileReads, ranks and cites the page used
On what is missingNo page exists, so nothing shows
OutputA shorter file
File reviewA test, item by item
On what is in the fileChecks it against the rule applied
On what is missingRecords the absence as a failed item
OutputA pass or fail per item
Topic not in the source documents
gpt-3.5-turbo produced a summary anyway79.35%
GPT-4o produced a summary anyway44%

An absence has no page to cite, so only a checklist that names the required item can find it.

Source: Belem et al., From Single to Multi: How LLMs Hallucinate in Multi-Document Summarization, arXiv:2410.13961, October 2024, revised April 2025. Asked to summarise topic information that did not exist in the source documents, gpt-3.5-turbo produced a summary 79.35% of the time and GPT-4o 44% of the time, across two multi-document benchmarks. Summarisation research across five models, not a claims study.

Omission is measured separately for the same reason. In a 2025 npj Digital Medicine study, clinicians found hallucinations in 1.47% of 12,999 summary sentences against omissions in 3.45% of the source transcript sentences. That is clinical documentation rather than claims, and the rates moved with the prompt, a point taken further in why a claims AI accuracy figure expires.

Vendors have made real progress on the readable half of the problem. Wisedocs' public materials describe source-linked summaries with page-level citations, and CLARA Analytics reports 83% less time spent reviewing medical records. Page-level citation solves provenance for what is in the summary.

None of it answers the review question, because an omission has no page to cite. The reviewer is asking whether the item the rule requires is in the file. A faithful summary of a file with no documented basis for its depreciation is a faithful summary of a file that fails item seven. A related distinction is in why a justification record is not an investigation record.

Making the file decision-ready while the claim runs

Answer

How do you make a claim file decision-ready?

Record the checks as they happen, including the ones that came back clean, so a review verifies a file instead of reconstructing one. Decision-readiness is a property of how the file was built rather than how carefully it is read afterwards, and no amount of later reading installs it.

The cheap way to pass a review is to stop treating documentation as a task separate from the work. Write the coverage position when the policy is read and the basis when the adjustment is made. Log the clean database check when it returns nothing. Then the review is a verification pass, not a reconstruction project.

It rarely happens because the failing items have one beneficiary: a reader who has not arrived yet. The adjuster already knows why the depreciation was applied. Under load, a file keeps the items that move the claim and loses the items that defend it.

This is also why retrofitting fails. A file is built at intake or it is not built at all. What belongs at the front door and what does not is in the first notice of loss process.

That is the part Hesper is built around. On a suspicious claim its agents run 15+ investigation phases in parallel, which moves coverage from about 25% of flagged claims to 100% of them, in minutes, not weeks. The checks that came back clean are recorded alongside the ones that did not.

Every agent action carries its source, its reasoning and its timestamp, which is the form all four readers want. The human keeps the decision: adjusters review evidence-backed files instead of building them.

One file serves all of it. The same cited record stands behind the coverage position, the fraud finding, the settlement valuation and the subrogation demand. The product view is on the claim file review page, and straight-through claims processing explains why a file that never touched a human still has to pass the same checklist.

A file assembled against the checklist is the same file in all four readers' hands. Build for the reader who was not there and you have built for all of them at once.

Key takeaways

  • A claim file is the evidence that the decision was defensible, so the checklist is the questions four later readers will ask, in the order they ask them.
  • Samples are chosen on risk, not at random: in the Pennsylvania exam of Allstate Fire and Casualty, a 75-file collision sample from 39,918 claims produced a published 28% error ratio.
  • Published findings are overwhelmingly absences rather than wrong decisions: an undocumented adjustment, a missing copy of the estimate, a closure without written notice.
  • A summary compresses what is present while a review tests what is absent, so the cheap way to pass is to record the checks as the claim runs.

Frequently asked questions

A claim file review checklist is the fixed list of items a reviewer tests on one file, recording a pass or fail against each. The useful ordering is coverage, investigation, quantum, conduct and communication, then closure and recovery, because each gate decides whether the next one is worth testing.

Enough that somebody who was not there can reconstruct what happened and why. In practice that means the policy version in force on the date of loss, the coverage position with the provision relied on, dated contact records, the investigation performed including the checks that returned nothing, the justification for every adjustment, each required written notice, and the subrogation screen with its result.

AI can read a claim file faster than a person and cite the page it used. Review is a different operation. A summary compresses what is present, while a review has to establish what is absent and whether what is present supports the decision. Treat the summary as the index and the checklist as the test.

On the Hesper platform

Hesper AI runs every stage of a claim, from first notice of loss to subrogation recovery, with investigation-grade evidence behind every decision.

The platformFNOL intakeClaims triageFraud investigationSubrogation & recovery

Keep reading

← More articles on the Hesper AI blog

See Hesper AI on your documents

Request a demo and we'll run an analysis on your real document samples.