To the Federal Trade Commission:

The Commission’s proposed policy statement asks whether companies may engage in deceptive practices when they present AI systems as accurate, objective, or reliable while secretly manipulating, constraining, or steering those systems contrary to consumers’ reasonable expectations.

Insurance claims present an especially important application of that concern.

I respectfully urge the Federal Trade Commission to open a focused investigation into the use of artificial intelligence in insurance claims handling.

The risk is not limited to one defective algorithm making one incorrect recommendation. A modern insurance claim may move through several AI systems embedded within an insurer-designed claims process:

• One AI interprets the First Notice of Loss.
• Another summarizes the reported facts.
• Another classifies the claim’s severity or complexity.
• Another selects its handling path.
• Another evaluates photographs and generates an estimate.
• Another assesses repairability, total-loss value, liability, or fraud risk.
• Another summarizes policy language or recommends a coverage application.
• Another recommends payment, escalation, further investigation, or closure.

Each system can narrow the facts, choices, and conclusions available to the next system or employee.

By the time the person called the “adjuster” receives the claim, multiple automated systems may already have framed the loss, selected the workflow, generated an estimate, and pointed the employee toward a proposed outcome.

The appropriate subject of investigation is therefore not an individual AI model in isolation. It is the combined claims decision system:

I. Measurable Warning Signs Justify a Deeper Investigation

Available outcome data does not prove that AI caused every underpayment, denial, or claim closure.

It does show that insurance claims are producing results serious enough to justify federal investigation.

A. Homeowners claims closed without payment have increased dramatically

An analysis of insurers’ statutory filings found that the national homeowners claims closed-without-payment rate increased from approximately 25.7% in 2004 to 39% in 2023 and 42.1% in 2024.

Among approximately 6.8 million homeowners claims received and closed in 2024, more than four in ten closed without payment. Fourteen large insurers reported rates between 40% and 51%. ()

A claim closed without payment does not automatically establish misconduct. The category can include losses below the deductible, excluded losses, duplicate claims, consumer withdrawals, insufficient documentation, and other legitimate outcomes.

But those explanations do not eliminate the trend.

A change from approximately one-quarter of claims to more than four out of ten represents a major shift in consumer outcomes. It requires more detailed examination than the broad “closed without payment” category presently allows.

B. Auto-liability claims show a similar long-term trend

A review of New York private-passenger auto-liability filings found that insurers closed approximately 48.6% of claims without payment in 2025, compared with 33.6% in 2005.

That is a movement from approximately one in three claims to nearly one in two. Allstate reportedly closed 55.5% of the relevant claims without payment, while two Progressive underwriting companies each reported 47.4%. ()

C. Right to Appraisal finds thousands of dollars beyond initial insurer evaluations

Texas Watch, a nonprofit consumer advocacy group, reviewed 1,246 disputed auto claims that proceeded through Right to Appraisal.

For repairable vehicles, appraisal increased the claim value by an average of approximately $5,307, or 131% above the insurer’s final offer.

For total-loss claims, appraisal added an average of approximately $3,889, or 26%.

Across all claims examined, independent appraisal identified more than $5 million in additional claim value.

These were disputed claims, not a random sample of every auto claim. The results should not be projected across all claims or insurers.

But the study demonstrates something significant: when qualified independent people reviewed contested insurer evaluations, they frequently identified thousands of dollars in additional damage or value.

That leads to an important consumer-protection question:

The Commission should not assume that artificial intelligence caused these outcomes. It should determine whether AI, insurer-selected rules, fragmented workflows, and restricted human authority are contributing to, accelerating, or concealing them.

II. One Consumer’s Claim

In an AI-handled claim, the insurer’s estimate totaled $438.10 on a policy with a $500 deductible.

After the consumer took the vehicle to a qualified collision repair facility, the shop estimated the damage at $2,697.40.

The difference was $2,259.30.

This was not an ordinary supplement or a modest professional disagreement. The customer had been told that the loss did not exceed the deductible, and the claim file was closed.

When I managed field estimators, supplement percentage was used as a competency and performance measure. An estimator whose work was regularly supplemented by more than approximately 10% of the original estimate could trigger review for additional training or deficiencies in first-inspection quality.

Here, the additional amount identified by the shop was approximately 516% of the original insurer estimate!

The estimate itself also contains a glaring internal contradiction.

It includes repair operations and two entries labeled “Basecoat Reduction” at negative 0.5 hour each. Together, those entries produce negative one refinish hour.

There is no positive refinish time from which that reduction could logically be taken. The estimate does not simply omit paint labor. It affirmatively deducts refinish labor and associated material value, reducing the total estimate.

A competent human reviewer should have questioned:

• Why repair operations appeared without positive refinish time.
• Why negative paint time was included.
• Whether refinish-material value was being improperly reduced.
• Why the estimate fell just below the consumer’s deductible.
• Whether the estimate bore a reasonable relationship to the visible damage.

Had the customer not sought a second opinion, the customer could reasonably have concluded that the loss did not exceed the deductible and that no insurance payment was available.

This example presents at least two troubling possibilities.

First, the estimate or surrounding process may have been intentionally structured or constrained to produce an amount below the deductible.

Second, the alleged human review may have been so poorly trained, fragmented, restricted, or inattentive that it offered no meaningful protection.

Both conditions may also have operated together: a system oriented toward a low result and a human employee who lacked the knowledge or authority to question it.

I am not asking the Commission to infer intentional misconduct from one estimate.

I am asking the Commission to investigate what an estimate like this reveals about the claimed human safeguard.

If an employee reviewed this estimate:

• What did that review consist of?
• Did the employee understand collision estimating?
• Did the employee inspect the vehicle or review all available evidence?
• Did the employee recognize the negative refinish calculation?
• Did the employee have authority to correct it?
• Did changing it require supervisory approval?
• Was the below-deductible result independently audited?
• Did the insurer compare the initial estimate with the final shop evaluation?
• Was the substantial deficiency used to retrain the model or reviewer?

If the human lacked the training, information, or authority to identify and correct an estimate this inaccurate, then that person did not meaningfully supervise the AI-supported process.

The person merely transmitted and legitimized its output.

III. AI and Claims-Process Design Are Inseparable

Insurance AI should not be evaluated as a tool operating alongside an otherwise independent adjuster.

AI is increasingly embedded throughout the claims architecture.

An AI-generated First Notice of Loss summary may omit or mischaracterize an important fact. A triage system may then classify the claim as minor or routine. That classification may place the claim into a low-touch process. A photo-estimating system may evaluate only visible damage. Another system may recommend a valuation or coverage application based on the limited information inherited from the earlier steps.

Each output becomes an input, assumption, or boundary for what follows.

An employee receiving the claim near the end of that sequence may never know that the original summary was incomplete, the classification was wrong, or the damage evaluation omitted necessary operations.

The claim may appear to have received several reviews. In reality, each system and employee may have relied on the same initial assumption.

This creates a compounding error problem.

A claim does not become more accurate merely because several systems touched it. Multiple automated touches can make one flawed conclusion appear increasingly authoritative.

Claims-process design can steer outcomes even if the AI model itself is unchanged

An insurer can influence the consumer outcome without changing the technical core of an AI model.

The insurer controls or influences:

• Which facts are collected at intake.
• Which information is summarized or excluded.
• Which claims are classified as simple.
• Which photographs or documents are evaluated.
• Which estimating rules and thresholds apply.
• Which claims are routed to human investigation.
• What information employees can see.
• How claims are divided among task queues.
• What settlement or coverage authority employees possess.
• When escalation is required.
• Which employee behaviors are rewarded or penalized.
• Whether overriding the system requires additional documentation or approval.

This architecture can make one result faster and easier to accept while making a different result slower and professionally riskier.

Accepting the automated recommendation may require one click.

Correcting it may require a written explanation, supervisory approval, reassignment, or several levels of escalation.

That is steering through process design.

The Commission should therefore ask not merely whether the AI was accurate in isolation, but whether the combined system was designed so that an incomplete or low outcome became the path of least resistance.

IV. “Human in the Loop” Is Not Meaningful Human Review

Insurers and technology vendors frequently claim that AI supports rather than replaces human judgment.

The phrase “human in the loop” can be technically accurate while materially misleading.

A human may be present but see only one narrow task.

The person may rely on an AI-generated summary rather than original evidence.

The person may lack the experience to identify an incomplete estimate or coverage analysis.

The person may not possess authority to modify the result.

The person may face added scrutiny or performance consequences for disagreeing with the system.

Professor Daniel Schwarcz’s forthcoming essay, Procedural Bad Faith and AI Claims Handling, argues that increasingly automated claims systems create a distinct procedural injury when insurers deny, reduce, or delay claims without meaningful human review, adequate explanation, or a genuine opportunity for the insured to be heard.

The paper explains that insurers often characterize algorithms as assisting human reviewers, even though automated systems may effectively displace judgment and leave humans providing only cursory oversight of outcomes already generated, structured, or predetermined by technology.

It also explains why merely placing a person after the algorithm has acted is a structurally weak safeguard. The automated recommendation anchors the reviewer. Approving it becomes the path of least resistance. Overriding it requires explanation and responsibility. Automation bias encourages overreliance, while organizational throughput pressures can further discourage disagreement. The central question is not simply whether a human was in the loop, but whether the loop was designed so the person could realistically do anything other than agree.

Meaningful human review should require:

• Relevant training and experience.
• Access to the complete claim.
• Access to original evidence, not merely generated summaries.
• Knowledge of the material limitations of each AI system.
• Sufficient time to evaluate the claim independently.
• Actual authority to change the result.
• Freedom from retaliation or negative performance consequences for disagreeing.
• Accountability for the overall consumer outcome.

Human presence is not the same as human judgment.

V. Current State Proceedings Show That the Concern Is Not Hypothetical

The FTC does not need to begin this investigation without existing evidence.

Current state proceedings separately allege underpayments, improper denials, inadequate investigations, undisclosed claims standards, and limitations on adjuster authority.

The allegations have not been finally adjudicated. The insurers are entitled to respond and defend their conduct.

But the underlying evidence may allow the FTC to examine whether AI, insurer programming, and claims-process design contributed to the challenged outcomes.

A. California Department of Insurance action involving State Farm

In May 2026, the California Department of Insurance announced enforcement action following a market-conduct examination of State Farm’s handling of claims arising from the 2025 Los Angeles wildfires.

The Department reported reviewing 220 claims and identifying 398 alleged violations in 114 of those files. That means at least one alleged violation was identified in 51.8% of the sampled claim files—more than half of the claims reviewed. The cited problems included slow or inadequate investigations, unreasonably low settlement offers, claim underpayments, delays, and denials. A defect rate appearing in more than half of a regulatory sample warrants immediate root-cause analysis. The appropriate question is no longer whether individual adjusters made occasional mistakes, but whether a common system, workflow, rule, technology, training deficiency, or authority structure contributed to repeated failures. (California Department of Insurance)

The FTC should seek evidence from the California Department of Insurance concerning:

• Whether AI evaluated fire, smoke, or property damage.
• Whether automated tools generated or constrained settlement recommendations.
• Whether claims were routed through automated classifications.
• Whether employees had access to complete files.
• Whether they could depart from system-generated conclusions.
• Whether common software rules or vendor tools contributed to repeated problems.
• Whether regulatory intervention resulted in increased payments.
• Whether consumers received meaningful notice about automated involvement.

B. Oklahoma’s lawsuit against State Farm

In June 2026, the Oklahoma attorney general filed a lawsuit alleging that State Farm engaged in a coordinated scheme to wrongfully deny or underpay legitimate hail- and wind-damage claims.

The state alleges that State Farm used an internal “Hail Focus Initiative” designed to reduce roof-replacement approvals and minimize claim payments through undisclosed claims-handling practices, restrictive standards, and outcome-oriented engineering reviews.

Although the allegations remain unproven, they are nevertheless relevant to the FTC’s inquiry because they raise the question of whether technology and process controls were used to implement or enforce the disputed standards.

The Commission should investigate:

• AI-supported inspection tools.
• Automated damage thresholds.
• Image-analysis systems.
• Claims-routing rules.
• Coverage decision trees.
• Employee scripts.
• Audit criteria.
• Management escalation requirements.
• Restrictions on adjuster settlement and coverage authority.

C. Oklahoma’s lawsuit against Allstate

In July 2026, the Oklahoma attorney general filed a separate lawsuit alleging that Allstate wrongfully denied or underpaid legitimate wind- and hail-damage claims.

The state alleges that Allstate altered its claims process by limiting field-adjuster authority, relying on third-party inspectors and reviewers, and applying restrictive internal standards that were not disclosed to policyholders.

The structure of these allegations closely resembles the risk created by AI-supported claims handling:

• One person gathers limited information.
• Another person or system interprets it.
• Coverage and payment authority are separated from the person who inspected the loss.
• Rules constrain what can be approved.
• The final result is nevertheless represented as an individualized insurer evaluation.

The FTC should seek the underlying process maps, training materials, vendor contracts, software specifications, authority guidelines, audit criteria, system records, and employee communications developed in these proceedings.

VI. Modern Claims Operations No Longer Follow State Boundaries

Traditionally, claims departments and adjusters were more likely to be located within the states and communities whose losses they handled.

Adjusters developed familiarity with local laws, construction and permitting standards, repair markets, courts, regulators, and policyholder expectations.

That structure has changed.

Insurers increasingly operate through remote and consolidated claims departments handling broad territories composed of multiple states. One remote claims unit, vendor, software platform, AI model, or centralized workflow may influence claims across an entire region or nationwide.

The employee handling a claim may be hundreds or thousands of miles from the loss. During the same workday, that employee may handle claims governed by several states’ laws while relying on centralized scripts, software prompts, AI summaries, remote photographs, and insurer-wide authority rules.

This creates a mismatch between the national structure of modern claims operations and a regulatory framework divided among individual states.

A state insurance department may see consumer complaints and claim files within its own borders. But the system that produced those results may have been:

• Designed in another state.
• Managed by a national claims department.
• Configured by an out-of-state vendor.
• Deployed across dozens of jurisdictions.
• Updated simultaneously for thousands of claims nationwide.

One state may identify inaccurate estimates. Another may identify improper denials. A third may see complaints involving the same platform or workflow.

VII. Insurer-Specific Programming Makes State-by-State AI Rules Insufficient

Claims systems are not standardized products operating identically from insurer to insurer.

Each company creates a unique architecture by choosing:

• Its vendors.
• Its AI models.
• Its software configurations.
• Its data fields.
• Its intake scripts.
• Its claim categories.
• Its routing thresholds.
• Its estimating rules.
• Its coverage prompts.
• Its authority levels.
• Its escalation requirements.
• Its performance measurements.

Even when two insurers use the same technology vendor, they may configure the product differently and connect it to different workflows.

It is therefore unrealistic to expect each individual state department of insurance to independently develop complete technical rules for every insurer-specific combination of models, vendors, rules, workflows, and human authority structures.

State regulators remain essential. They possess claim files, complaints, market-conduct authority, and local enforcement powers.

But state regulation alone cannot provide a consistent national baseline for systems that operate across state borders.

No single state regulator can easily determine:

• Whether the same vendor product is producing similar errors elsewhere.
• Whether an insurer uses materially different configurations across states.
• Whether a nationwide model update changed outcomes in multiple jurisdictions.
• Whether a national claims center is applying the same problematic workflow across a broad territory.
• Whether repeated errors originate with the model, the vendor, the insurer’s configuration, or the human process surrounding it.

Federal involvement should create the common consumer-protection floor that state regulators cannot establish individually.

The FTC can address national representations about AI accuracy, objectivity, and meaningful human review. It can examine vendors and insurer practices operating across state lines. State departments of insurance can then apply those standards through claim-level investigations and enforcement of state claims-handling law.

The modern claims process is centralized, remote, proprietary, and national.

Oversight cannot remain fragmented and unable to see the complete system.

VIII. Requested FTC Investigation

I respectfully ask the Commission to open a coordinated investigation with state departments of insurance, state attorneys general, and the National Association of Insurance Commissioners.

For selected claims and insurer systems, the FTC should reconstruct the full AI decision chain:

1. Which AI and automated systems touched the claim?
2. What did each system interpret, summarize, classify, generate, or recommend?
3. What information did each system receive?
4. What information was omitted, unavailable, or outside its capability?
5. Did one AI output become an assumption for another system?
6. Which insurer-specific rules and thresholds applied?
7. What information was presented to each employee?
8. Could the employee access the original evidence and full claim?
9. Was the employee trained to recognize the system’s limitations?
10. Did the employee have authority to reject or correct the recommendation?
11. Did disagreement require escalation or additional documentation?
12. Did override behavior affect performance evaluations?
13. Did any qualified person independently review the complete claim?
14. Was the consumer informed that automated systems materially influenced the decision?
15. Was the outcome later changed after appraisal, repair-facility review, regulatory complaint, or litigation?

IX. Requested Federal Guidance

The FTC should establish a national consumer-protection baseline addressing the following principles.

1. Accuracy must be measured against the complete consumer outcome

An insurer or vendor should not represent a system as accurate merely because each component performs a narrow task correctly.

The relevant question is whether the combined claims process fairly and completely evaluated the loss.

2. Insurer-specific steering must be disclosed

Consumers and regulators should be told when AI outputs are influenced by insurer-selected rules, thresholds, preferred practices, or financial objectives.

3. Material AI involvement must be disclosed

Consumers should receive notice when AI materially influences:

• First Notice of Loss interpretation.
• Claim classification.
• Routing.
• Estimating.
• Valuation.
• Coverage analysis.
• Fraud referral.
• Liability assessment.
• Total-loss determination.
• Closure without payment.

4. Meaningful human review must be defined

An insurer should not claim that a decision was human-reviewed unless a qualified person had access to the complete evidence, understood the systems’ limitations, and possessed actual authority to alter the outcome.

5. Complete audit trails must be preserved

Insurers and vendors should retain:

• Input data.
• Model identity and version.
• Insurer configurations.
• Rules and thresholds.
• Outputs and recommendations.
• Employee actions.
• Attempted overrides.
• Escalation decisions.
• Final reasons for payment, reduction, denial, or closure.

6. First-inspection quality must be measured

Routine reliance on supplements should not substitute for measuring the accuracy and completeness of the first inspection.

Regulators should examine:

• Supplement frequency.
• Supplement size.
• Difference between original estimates and final repair costs.
• Operations repeatedly omitted.
• Claims initially placed below the deductible but later found to exceed it.
• Differences between virtual and physical inspections.
• Whether errors disproportionately reduce claim value.
• Whether significant errors result in retraining or system correction.

Speed, closure, and claim-severity metrics should not replace first-inspection accuracy.

An estimate produced quickly but materially below a qualified repair evaluation is not efficient for the consumer.

It is simply wrong faster.

7. Regulators must review claims consumers did not challenge

Most consumers do not invoke Right to Appraisal, hire a public adjuster, retain counsel, or file a regulatory complaint.

Regulators should independently sample unchallenged claims, especially claims closed without payment and claims estimated just below deductibles.

A system cannot be judged fair only by examining consumers who possessed enough knowledge and resources to expose its errors.

Conclusion

Insurance claims may be one of the clearest settings in which to investigate whether AI is presented as accurate and objective while insurer-designed systems quietly steer the result.

The warning signs are measurable:

• National homeowners claims closed without payment increased from approximately 25.7% in 2004 to 42.1% in 2024.
• New York private-passenger auto-liability claims closed without payment increased from 33.6% in 2005 to 48.6% in 2025.
• Texas Right to Appraisal data found average increases of approximately $5,307 for disputed repair claims and $3,889 for total-loss claims.
• Current proceedings in California and Oklahoma allege underpayment, improper denial, inadequate investigation, restrictive internal standards, and limitations on adjuster authority. (California Department of Insurance; Oklahoma Attorney General.)

These facts do not prove that every AI-assisted claim is deceptive.

They provide compelling reason for the FTC to investigate.

The public should not be reassured merely because an insurer says that an adjuster remains involved.

The insurer may know that the employee:

• Cannot see the complete claim.
• Handles only one isolated task.
• Relies on AI-generated summaries.
• Is not trained to identify the error.
• Lacks authority to change the result.
• Must obtain permission to disagree.
• Is evaluated on speed, severity, or conformity.

In that environment, AI does not merely assist a human decision.

The human gives institutional legitimacy to an outcome the combined AI and claims process has already made difficult to challenge.

Insurance should work accurately and fairly before a third-party appraiser, repair professional, public adjuster, regulator, or attorney becomes involved.

The consumer who quietly trusts the insurer’s evaluation is no less entitled to a complete and accurate decision.

Fair claim payment cannot depend on the squeaky wheel.

I respectfully urge the Federal Trade Commission to investigate the full chain through which AI-supported insurance claims systems convert limited information into consequential consumer outcomes; coordinate with state insurance departments and attorneys general that already possess relevant evidence; and establish federal guidance creating consistent standards for accuracy, disclosure, auditability, and meaningful human review.