The AI claims evidence trail: a justification record is not an investigation record

The question to ask about an AI claims evidence trail is not who owns it. It is how the trail gets produced, because two different artifacts go by the same name. A justification record shows what a system considered and where a score came from. An investigation record also shows what was checked and what was ruled out, including the checks that came back clean. US claim-file rules ask for the second one, and they asked for it decades before any of this software existed.

The sharpest live version of the claim belongs to a named vendor. On September 22, 2026 CLARA Analytics launched Agentic Intelligence, described in its release as "the industry's first end-to-end AI workforce built for complex casualty and workers' comp claims", with one sentence that plants a flag: "Adjusters keep the decision; CLARA owns the evidence trail behind it." That sentence is correct about what matters. The evidence trail is the thing worth owning. What belongs in it is the part nobody has pinned down.

So the disagreement is not with the claim. It is with what the word "trail" is allowed to mean. Score provenance explains an output. A claim file has to account for an insurer's activities on one claim, including the activities that turned up nothing, and it has to do that years later in front of an examiner who was not there. We have worked through what the claim-file rules asked for before AI state by state, and where agentic claims deployments actually sit relative to investigation.

The release is specific about scope and about what the transparency looks like. Agentic Intelligence is built into the CLARAty.ai platform and "puts a team of AI agents on every claim, from first notice of loss through resolution, while adjusters stay in control of every decision." Its published capability bullets are "Complete Lineage and Outcome 'Ground Truth'", "Data Freshness and Quality", "Cohort Contextualization", "Transparent Rationale" and "Interactive Deep Dives Across Structured and Unstructured Data". Mubbin Rabbani, the company's Chief Product Officer, is quoted saying "Insurers get turn-key access to an advanced AI workforce backed by robust guardrails and human validation." The data claim is a training claim: "a 10-year data moat of over 7 million claims" describes what the models learned from, not a customer count and not an outcome.

Three things deserve stating plainly before any argument. The release publishes no accuracy, speed, cost-saving or ROI figure, which is normal for a launch announcement and not a tell. "Adjusters keep the decision" is a deliberate human-in-the-loop design choice, and a defensible one. And the vocabulary is shared: Claims Journal's September 29, 2026 round-up covered the launch alongside Cozmo AI, which "deploys 'AI employees' to handle operational work from FNOL to invoice settlement", plus Finch's "full-claim intelligence platform" and Shift Technology's agentic Shift Claims. "AI workforce" is category language now, not one company's quirk.

The same column states the transparency claim in the company's own terms: "According to the company, adjusters stay in control of decisions. The system gives adjusters access to where a score came from and how current underlying data is." That is score provenance, and score provenance is a real and useful thing to ship. It is also a property of a recommendation. The claim file is a record of an insurer's activities on one claim. Carriers evaluating this layer against other options will find the module-by-module version of that comparison in our CLARA Analytics alternatives breakdown.

The definitions that decide this are not ours. NAIC Model Regulation 902, the model unfair property and casualty claims settlement practices regulation, defines "Documentation" to include "all pertinent communications, transactions, notes, work papers, claim forms, bills and explanation of benefits forms relative to the claim" (Section 3.E), and defines "Investigation" as "all activities of an insurer directly or indirectly related to the determination of liabilities under coverages afforded by an insurance policy or insurance contract" (Section 3.G). All activities. Then Section 4.B sets the test: "Detailed documentation shall be contained in each claim file in order to permit reconstruction of the insurer's activities relative to each claim."

Read Section 4.B against a lineage-complete score and the gap is structural rather than a matter of implementation quality. Lineage reconstructs an output: these inputs, this version, this freshness, this rationale. Section 4.B asks for reconstruction of activities. An activity that was performed and returned nothing produces no output to trace, so it leaves no trace at all unless something deliberately writes it down. The better the score provenance, the easier it is to mistake a complete account of a recommendation for a complete account of a claim.

This is the distinction Hesper is built on, and the reason the output is a file rather than a score. The evidence file Hesper produces per claim carries an executive summary, an evidence package with source documents, a timeline reconstruction, statement analysis with the inconsistencies named, a coverage position, risk scoring with its rationale and recommended next actions. We verify, not just flag. We investigate, not just score. And the same evidence engine stands behind the coverage position at stage 3, the settlement valuation at stage 6 and the subrogation screen at stage 8, which is what "investigation-grade evidence behind every decision" means in practice: the depth built for SIU work sits behind the ordinary decisions too.

Model 902 Section 5.C is the clearest instance in the model text of a negative finding doing the work: "A claim shall not be denied on the basis of failure to exhibit property unless there is documentation of breach of the policy provisions in the claim file." The file has to carry proof about something that did not happen. Model Act 900 points the same way from the conduct side: Section 4.F makes it an unfair practice to be "Refusing to pay claims without conducting a reasonable investigation", which is a finding about the absence of activity, provable only from a record of what activity there was.

The fraud-suspicion provisions raise the bar again. Model 902 Section 7.A relieves an insurer of part of the written-denial requirement "Where there is a reasonable basis supported by specific information available for review by the insurance regulatory authority that the first party claimant has fraudulently caused or contributed to the loss", and Section 7.B allows additional time on a materially identical condition. The entitlement to treat a claim as suspicious is conditioned on what the file can put in front of a regulator. California states the duty in both directions: 10 CCR 2695.7(d) requires every insurer to "conduct and diligently pursue a thorough, fair and objective investigation" and forbids persisting "in seeking information not reasonably required for or material to the resolution of a claim dispute". Proving either half needs a record of what was sought and what it returned.

Negative findings go missing for a reason that has nothing to do with intent. A cleared check gives the adjuster nothing and costs one more line to type, so under caseload pressure it is the first thing to stop being written down. The arithmetic is the arithmetic of manual investigation: 14+ days per case and 200+ cases per investigator, both Hesper internal benchmarks, which is why only about 25% of flagged claims get a real workup and the rest get a disposition. Run 15+ investigation phases in parallel instead, in minutes, not weeks, and the economics invert. The phase ran whether or not anyone had time to annotate it, coverage of flagged claims goes from about 25% to 100% on the same benchmark, and the cleared checks land in the file as findings rather than as nothing.

None of this is new law, and none of it was written with AI in mind. Model 902 Section 4 opens by stating that "Each insurer's claim files for policies or certificates are subject to examination by the Commissioner of Insurance or by the Commissioner's duly appointed designees", then adds the reconstruction requirement in 4.B and a dating requirement in 4.C: "Each relevant document within the claim file shall be noted as to date received, date processed or date mailed." Section 7.A requires that a denial on a specific policy provision, condition or exclusion reference that provision, be given in writing, and be documented in the file. California's 10 CCR 2695.3(a) states the reconstruction test in its own words, requiring files to contain all documents, notes and work papers "in such detail that pertinent events and the dates of the events can be reconstructed and the licensee's actions pertaining to the claim can be determined".

Read together, these are documentation requirements expressed as conduct requirements. The regulator does not ask whether the file is tidy. It asks whether an outsider can rebuild what the insurer did, in order, with dates, and whether the written reason given to the claimant rests on something in the file. An automation layer inherits that question the moment it starts doing the activities.

The arithmetic is short and consequential. 10 CCR 2695.3(b)(1) requires claim data to be "accessible, legible and retrievable for examination" and available "for all open and closed files for the current year and the four preceding years", and (b)(3) requires files to be maintained on the same horizon: the current year plus 4 preceding years, 5 years in all. Model 902 Section 4.A asks for the current year plus 2, or 3 years in total. Either way the examinable life of the file is measured in years while a model version's life is measured in release cycles.

That is also where logging obligations and claim-file obligations separate. Article 12 of the EU AI Act requires that "High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system", and its worked example for biometric systems lists the period of each use, the reference database, the input data that matched and the people who verified the result. Every one of those is an event in a system. A US carrier that satisfies Article 12 has not thereby satisfied Model 902 Section 4.B: one provision asks about the system over its lifetime, the other about the insurer's activities on one claim. The same mismatch shows up in litigation, where the manual claim file fails the same test as often as an automated one.

The California Department of Insurance's targeted market conduct examination of the California FAIR Plan Association is the most itemised public example. The review period covered homeowners and dwelling fire claims closed between 1 January 2017 and 18 March 2021, a period holding 9,755 claims; examiners reviewed 259 sample files and cited 418 total violations, about 1.6 cited violations per file examined (derived). The same period produced 173 California consumer complaints related to those claims. The Department's July 31, 2025 announcement of legal action describes the same 418 violations across 259 claims reviewed, with 118 tied to smoke damage claim denials, and names "Failing to investigate claims fairly" and "Denying legitimate claims without reasonable basis" among the failures. The examination report notes that its alleged violations have not undergone a formal administrative or judicial process.

The distribution is the finding. Across these five allegations the investigation standard draws the most citations and the file-maintenance standard the fewest, 38 against 4. The exam's single most-cited line sits outside this table and makes the same point: 85 citations under 10 CCR 2695.7(c)(1), the duty to put in writing, and then restate every 30 days, the reasons more time is needed on an unresolved claim. The report also records that documentation failures around betterment, depreciation and salvage were "identified in the previous examination report" and "were also identified in the current examination", which is what a structural gap looks like from the regulator's side rather than a bad quarter.

A fresher action points the same way. On May 4, 2026 the Department announced legal action over the handling of Los Angeles wildfire claims, finding 398 violations of state law in 114 claims out of a sample of 220, against roughly 11,300 residential claims handled. That is about 52% of sampled files carrying at least one alleged violation (derived). The lead finding category is headed "Slow and inadequate investigation", and the cited failures include not beginning to investigate within 15 days and not providing required written denials for hygienist and environmental testing. These are allegations rather than adjudicated findings, and neither action should be read as more than what the Department alleged. What the pattern establishes is where the exposure sits: in the thoroughness of the investigation and the written basis for the decision.

The operative sentence in the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted December 4, 2023, is this: "Actions taken by Insurers in the state must not violate the UTPA or the UCSPA, regardless of the methods the Insurer used to determine or support its actions." Section 1 adds that an insurer's conduct, "including its use of AI Systems to make or support actions and decisions that impact consumers, is subject to investigation, including market conduct actions." The bulletin then ties control depth to, among other factors, "(iii) the extent to which humans are involved in the final decision-making process" and "(v) the extent and scope of the insurer's use or reliance on data, Predictive Models, and AI Systems from third parties". Its AI systems programme guidance names "case management, claim administration and payment, and fraud detection" in scope, and lists traceability, "the auditability of these measurements" and "Data and record retention" among the expected assessments.

Two of its third-party provisions matter for anyone buying this layer. Section 4.1 expects due diligence on a third party's AI systems so that "decisions made or supported from such AI Systems that could lead to Adverse Consumer Outcomes will meet the legal standards imposed on the Insurer itself", and Section 4.2(b) suggests contract terms requiring the vendor to "cooperate with the Insurer with regard to regulatory inquiries and investigations related to the Insurer's use of the third-party's product or services". Ownership of the trail, in the only sense a regulator recognises, is not transferable. The vendor can produce and hold the record; the carrier answers for it. Our post on how regulators now ask for the model by name covers the governance half of that obligation.

The state layer is uneven and worth getting right. On the NAIC's own implementation map, status as of April 1, 2026, 25 jurisdictions have adopted the model bulletin and 4 more have insurance-specific regulation or guidance of their own: California's Bulletin 2022-5, Colorado's 3 CCR 702-10 effective November 13, 2023 with amendments effective 10/15/25, New York's Insurance Circular Letter No. 7 issued July 11, 2024 and addressed to underwriting and pricing rather than claims, and a Texas bulletin from 2020. The map notes it does not determine whether any state's action contains all elements of the model.

Texas has moved since that status date, and the newer bulletin settles the human-in-the-loop question for everyone in the category. Commissioner's Bulletin B-0003-26, dated June 12, 2026, states that "If a regulated entity uses AI to make a consequential decision, TDI expects a person to review and agree with all decisions before action is taken", and points to Texas Insurance Code Chapter 542 on the processing and settlement of claims. A regulator expects a person to agree with the decision. CLARA's "adjusters keep the decision" is consistent with that, and so is Hesper's position: not human-replacement, adjusters and investigators keep decision authority and the agents do the legwork. The disagreement in this category is not about whether a human reviews. It is about what that human is handed, which is the subject of which decisions stay human.

Model Act 900 states its purpose as setting "standards for the investigation and disposition of claims arising under policies or certificates of insurance issued to residents of [insert state]", then adds: "It is not intended to cover claims involving workers' compensation, fidelity, suretyship or boiler and machinery insurance." Model 902 Section 2 carries the same carve-out for the regulation, which sets "minimum standards for the investigation and disposition of property and casualty claims". So of CLARA's three published lines, auto liability and general liability sit squarely inside the unfair-claims frame, while workers' compensation is excluded from the model act and the model regulation by their own terms and governed instead by each state's comp statute, its benefit clocks and its audit regime. That is an observation about which rulebook applies, not a criticism of the platform.

For a carrier, TPA or MGA the practical question is coverage of the book, not coverage of a line. Hesper runs auto, property, workers' compensation, general liability, liability and bodily injury, and pet claims, for carriers, TPAs, MGAs and self-insured programs across personal and commercial lines, which means one evidentiary standard rather than one per vendor. The same point in reverse is why we wrote up handler-assist agentic AI versus autonomous investigation, and why every claim in the eight-stage claims lifecycle carries the same file forward: clean claims resolve straight through, suspicious claims get an investigation-grade workup, and both leave the same kind of record.

The claim is now common enough to be worth testing in procurement rather than in a market conduct examination. Eight questions, each tied to something already in the rules, turn the argument above into a diligence artifact. Ask them of us too.

1. Does the record show activities or only conclusions? Model 902 Section 4.B asks for reconstruction of the insurer's activities relative to each claim, not reconstruction of the recommendation.
2. Are checks that returned nothing recorded as findings? Section 3.G defines investigation as all activities, and Section 5.C shows the model contemplating documentation of something that did not happen.
3. Can the file be rebuilt in year 5, after the model version that produced it has been retired? California keeps it examinable for the current year plus 4.
4. Does the written basis for a denial come out of the same record? 10 CCR 2695.7(b)(1) wants all bases plus the factual and legal basis for each reason given.
5. Who holds the record if we switch vendors, in what format, and who can read it after the contract ends?
6. Will you commit in writing to cooperate with a market conduct inquiry? The NAIC model bulletin Section 4.2(b) suggests exactly that contract term.
7. Which lines of business is this evidentiary standard available on, and which rulebook applies to each of them?
8. What does the adjuster receive at the moment of decision: a recommendation with its rationale, or a completed record of what was checked and what was ruled out?

Question eight is the one that separates the artifacts, and it is the one Hesper answers with a file rather than a score. Adjusters review evidence-backed files instead of building them. Every agent action is logged with its sources, reasoning and timestamps, so the record is defensible to regulators, reinsurers and capacity partners, and the decision stays with the licensed human who signs it. Evidence behind every decision. CLARA is right that the evidence trail is worth owning; the rulebook decides what has to be in it.