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 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 this software existed. 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.
CLARA planted a flag on the evidence trail, and the flag is in the right place
Answer
What did CLARA Analytics announce in September 2026?
On September 22, 2026 CLARA Analytics launched Agentic Intelligence, which it calls the industry's first end-to-end AI workforce built for complex casualty and workers' comp claims. Its release states that adjusters keep the decision and CLARA owns the evidence trail behind it. The published lines are workers' compensation, auto liability and general liability.
CLARA's release is correct about what matters: the evidence trail is the thing worth owning. "Adjusters keep the decision" is a defensible human-in-the-loop choice. And "AI workforce" is category language now: Claims Journal's September 29, 2026 round-up covered the launch alongside Cozmo AI, Finch and Shift Technology's agentic Shift Claims.
The same column states the transparency claim: "The system gives adjusters access to where a score came from and how current underlying data is." That is useful score provenance. It is also a property of a recommendation, not a record of an insurer's activities on one claim. Our CLARA Analytics alternatives breakdown compares the options module by module.
A justification record and an investigation record are different artifacts
Answer
What is the difference between a justification record and an investigation record?
A justification record shows what a system considered and where its score came from: inputs, model version, data recency, rationale. An investigation record adds what was checked and what was ruled out, including findings that came back clean. The first describes a recommendation. The second describes the insurer's activities on one claim.
The definitions that decide this are not ours. NAIC Model Regulation 902 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.
Section 4.B then 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."
A justification record answers why. The claim file has to answer what.
Six questions a reviewer puts to an AI-handled file, and which of the two records can answer each one.
The claim file rules ask what the insurer did, not what the model concluded.
California's 10 CCR 2695.3(a) states the same test in its own words: files must hold 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". The table sets out the main provisions.
The gap is structural, not a matter of implementation quality. Lineage reconstructs an output: these inputs, this version, this rationale. Section 4.B asks for reconstruction of activities. A check that was performed and returned nothing produces no output to trace, so it leaves no trace unless something deliberately writes it down.
This is why Hesper's output is a file rather than a score. The evidence file Hesper produces per claim carries an evidence package with source documents, a timeline reconstruction, statement analysis, a coverage position and risk scoring with its rationale. We verify, not just flag. We investigate, not just score.
Negative findings are part of the record, not a byproduct
Answer
Why do negative findings belong in the claim file?
Because the rules make absences load-bearing. Model 902 defines investigation as all activities related to determining liabilities, requires documentation permitting reconstruction of those activities, and in Section 5.C requires documentation of a breach in the file before a denial for failure to exhibit property. A check that returned nothing is an activity.
Model 902 Section 5.C is the clearest instance: "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.
Negative findings go missing under caseload pressure, not by intent. A cleared check gives the adjuster nothing and costs one more line to type. Manual investigation takes 14+ days per case with 200+ cases per investigator, both Hesper internal benchmarks, which is why only about 25% of flagged claims get a real workup.
Run 15+ investigation phases in parallel instead, in minutes, not weeks, and the economics invert. 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.
A check that comes back clean is a finding, so it gets written down
The phases run in parallel and every result is recorded, including the ones that found nothing.
- 01Phases in parallel15+ investigation phases run at once, not one case at a time
- 02Every result keptChecks that return nothing are written down as findings
- 03Sources attachedEach value cites its document, page or database, and its timestamp
- Coverage decision
- Finding or referral
- Denial basis in writing
The record outlives the model version that produced it
Answer
How long does an AI-produced claim record have to stay reconstructable?
In California, for the current year plus the 4 preceding years, so 5 years in all. NAIC Model 902 Section 4.A sets the model standard at the current year plus 2. Both horizons outlast the model version that produced the record, which is why the artifact has to be readable without the system that made it.
The examinable life of the file is measured in years. A model version's life is measured in release cycles. That gap is the problem: the record has to be readable long after the system that wrote it has changed.
The file stays examinable for 5 years. Model versions do not.
California keeps claim files open to examination for the current year plus the 4 preceding years.
Logging obligations and claim-file obligations separate here too. 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". That asks about the system. Model 902 Section 4.B asks about the insurer's activities on one claim, so satisfying one does not satisfy the other.
The same mismatch shows up in litigation, where the manual claim file fails the same test as often as an automated one.
The model that wrote the file will be retired before the file stops being evidence.
What a thin record costs, in the regulators' own counts
Answer
What do regulators actually cite when claim files fail?
Investigation quality and the written basis for the decision, more than data lineage. In one California targeted examination of 259 sampled files, 38 of the 418 alleged violations were for failing to conduct a thorough, fair and objective investigation, and 9 for failing to give written reasons for a denial.
The California Department of Insurance's targeted examination of the California FAIR Plan Association is the most itemised public example. Examiners reviewed 259 sample files of homeowners and dwelling fire claims and cited 418 total violations. The report notes that its alleged violations have not undergone a formal administrative or judicial process.
The Department's July 31, 2025 announcement of legal action names "Failing to investigate claims fairly" among the failures. The table lists five of the alleged violations with their counts.
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.
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. The lead finding category is headed "Slow and inadequate investigation". These are allegations rather than adjudicated findings.
The pattern shows where the exposure sits: in the thoroughness of the investigation and the written basis for the decision.
The AI rules scale the documentation bar to how much the AI does
Answer
What do the AI rules add to claim-file documentation?
They keep the liability with the insurer and scale expected controls to how much the AI does. The NAIC model bulletin states that an insurer's actions must not violate the unfair trade practices or unfair claims settlement practices acts regardless of the methods used, and sizes controls partly by the extent of human involvement and reliance on third parties.
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 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.
Texas settles the human-in-the-loop question. 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".
CLARA's "adjusters keep the decision" is consistent with that, and so is Hesper's position: adjusters and investigators keep decision authority and the agents do the legwork. The disagreement is not about whether a human reviews. It is about what that human is handed, which is the subject of which decisions stay human.
Line of business decides which rulebook you answer to
Answer
Does the unfair claims rulebook cover workers' compensation?
Not by its own terms. Model Act 900 Section 1 and Model Regulation 902 Section 2 both say they are not intended to cover claims involving workers' compensation, fidelity, suretyship or boiler and machinery insurance. Workers' comp claim files answer to a different rulebook, which matters when a platform's published scope leads with that line.
Model Act 900 says of its own scope: "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.
So of CLARA's three published lines, auto liability and general liability sit squarely inside the unfair-claims frame. Workers' compensation is governed instead by each state's comp statute, its benefit clocks and its audit regime.
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, which means one evidentiary standard rather than one per vendor. The table sets CLARA beside neighbouring vendors.
We cover handler-assist agentic AI versus autonomous investigation separately. 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.
Eight questions for any vendor that says it owns the evidence trail
Answer
What should a carrier ask a vendor that says it owns the evidence trail?
Whether the record shows activities or only conclusions, whether cleared checks are recorded, whether the file can be rebuilt in year 5 after a model version retires, where the written denial basis comes from, who holds the record after the contract ends, and which lines of business the standard covers.
The claim is now common enough to be worth testing in procurement rather than in a market conduct examination. Ask these of us too.
- Does the record show activities or only conclusions?
- Are checks that found nothing recorded as findings?
- Can the file be rebuilt in year 5, after the model version that produced it has been retired?
- Does the written basis for a denial come out of the same record?
- Who holds the record if we switch vendors, in what format, and who can read it after the contract ends?
- Will you commit in writing to cooperate with a market conduct inquiry?
- Which lines of business does this standard cover, and which rulebook applies to each?
- 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 separates the artifacts, and Hesper answers it with a file rather than a score. Every agent action is logged with its sources, reasoning and timestamps, and the decision stays with the licensed human who signs it. CLARA is right that the evidence trail is worth owning; the rulebook decides what has to be in it.
Key takeaways
- CLARA Analytics' September 22, 2026 launch states that adjusters keep the decision and CLARA owns the evidence trail behind it. What it publishes describes a recommendation.
- NAIC Model Regulation 902 defines investigation as all activities of an insurer, so the checks that came back clean belong in the file as findings.
- Enforcement follows investigation quality: in one California examination of 259 sampled files, 38 of 418 alleged violations were for failing to conduct a thorough, fair and objective investigation.
- California keeps a claim file examinable for the current year plus the 4 preceding years, long after the model version that produced it is retired.
- Human decision authority is shared ground: Texas Bulletin B-0003-26 expects a person to review and agree with consequential AI decisions. The real question is what that person is handed.