An AI Helped Deny Your Claim. Ask for the Decision Record.

Consider a health-plan denial letter that says a treatment was “not medically necessary.” It identifies the service and gives an appeal deadline, but it does not say whether software extracted a diagnosis from the clinical notes, matched billing codes against a coverage rule, assigned a risk score or merely routed the file to a clinician.

That letter is the object to work from. The useful challenge is not a demand to explain artificial intelligence in general. It is a request to reconstruct this decision: which records entered the system, which criteria controlled the result, where those records came from, whether software changed or summarized them, and what a human reviewer can reconsider.

The applicable rights depend on the type of coverage, the plan documents and the jurisdiction. A self-funded employer health plan, Medicare Advantage plan, state-regulated insurance policy and public-benefit program do not share one appeal system. This explainer describes a method for reading and challenging an AI-assisted outcome; it is not legal, medical or financial advice.

Start with the adverse decision, not the model

An explanation of a model describes the system across many cases. It might identify the categories of data used, the purpose of a score, known limitations or the way humans oversee outputs. That information can reveal structural problems, but it may say nothing about why one claim failed.

An explanation of an individual outcome connects the system to the denial letter. It should identify the material facts attributed to the claimant, the coverage provision or eligibility rule applied to those facts, and the path from the submitted record to the result. If a clinical note was summarized as “no prior conservative treatment,” for example, the decisive issue may be a missing document or a faulty extraction rather than the model’s overall accuracy.

The distinction matters because a model card, which is documentation describing a model’s intended use and limitations, does not establish what happened in a particular run. Neither does a statement that “automated tools may assist our reviewers.” Those disclosures address system design. A challenge needs the decision record.

Begin by marking every factual assertion in the denial letter. Note the named service, diagnosis, dates, provider, eligibility status and stated reason. Compare each item with the submitted claim, clinical record or benefits application. Keep the original letter, envelope, portal messages and screenshots showing when documents were uploaded, since appeal periods and proof of submission can matter.

Ask for the inputs and controlling criteria

A focused written request can ask the insurer, plan administrator or agency to identify the records considered and the criteria applied. Use the language from the letter rather than alleging that an algorithm made the decision before that is known.

For the “not medically necessary” letter, the request could say: “Please provide the records, clinical criteria, plan provisions, internal guidelines and other material relied upon in reaching this determination. Please identify any factual finding that was decisive, including any finding that required treatment, documentation or eligibility information was absent.”

That request tracks rights that already exist in some benefit systems, whether or not AI was involved. Federal rules for plans governed by the Employee Retirement Income Security Act require an adverse benefit determination to include the “specific reason or reasons” for the denial and reference the “specific plan provisions” on which it rests. For covered health claims, the rules also address access to relevant documents and any internal “rule, guideline, protocol, or other similar criterion” relied upon. When a denial turns on medical necessity or an experimental-treatment exclusion, the notice must provide an explanation of the scientific or clinical judgment, or state that one will be supplied free on request.

Those requirements are enforceable rules within their scope, not a general federal right to inspect every algorithm. They also do not guarantee access to source code, model weights or every internal log. A decision log, meaning a record of inputs, software outputs and reviewer actions during one case, may be operationally valuable without being a document the claimant can always compel.

Medicare Advantage adds a more specific constraint. Current Centers for Medicare and Medicaid Services rules allow an algorithm or software tool to assist a coverage decision, but the organization must base the decision on the individual patient’s circumstances and the applicable coverage criteria. The tool cannot substitute its own criteria for Medicare coverage rules. A challenge should therefore connect any generalized estimate, such as an expected recovery period, to the patient-specific record that the reviewer was required to consider.

Correct the record before arguing about automation

Once the insurer identifies the material facts, inspect their provenance, meaning where each item originated and how it entered the file. A provider may have submitted an outdated diagnosis code. Optical character recognition, software that converts scanned pages into searchable text, may have dropped a handwritten qualification. A benefits system may have matched the wrong employment status or failed to associate an attachment with the claim.

Send corrections in a format that a new reviewer can verify. Quote the disputed finding from the denial letter or case file, state the corrected fact, attach the source document, and identify the page or section that supports it. If the correction comes from a provider, ask the provider to send it through the channel the plan accepts rather than relying only on a patient-uploaded copy.

The same approach applies when an insurer used a consumer report for underwriting or pricing. Under the Fair Credit Reporting Act, an insurer that takes adverse action based in whole or in part on information in a consumer report generally must provide an adverse-action notice identifying the consumer reporting agency and explaining the right to obtain and dispute the report. The reporting agency, not the insurer, usually handles correction of the report itself. That is a separate route from appealing the insurance outcome, and both may need attention.

Returning to the medical-necessity letter, a generic claim that AI is biased will rarely repair a missing physical-therapy note. The stronger record says that the denial relied on an absence, shows that the treatment occurred, documents when the proof was submitted and asks the reviewer to rerun the coverage analysis using the corrected file.

Make human reconsideration testable

“Human in the loop” is weak language. A person may have clicked approve after reading a generated summary, or may review only cases that cross a software threshold. Ask what the human will do on appeal.

A useful request seeks reconsideration by someone authorized to change the outcome, using the original materials plus the corrections. For a clinical denial, the plan’s procedures may provide for review by an appropriate health professional who was not involved in the initial determination and is not subordinate to that person. The appeal should ask the reviewer to address the disputed factual finding and explain how the cited coverage criterion applies to the individual record.

Also ask whether an automated system generated, ranked or summarized information material to the decision. If so, request the output associated with the case, the input fields that produced it and any confidence flag or exception presented to the reviewer. A confidence score estimates how strongly a system supports an output; it is not proof that the output is correct. Ask whether staff could override the output and whether anyone did.

There is no broad US rule requiring every insurer or benefits administrator to disclose each use of AI to each affected person. State insurance departments have begun applying governance expectations modeled on the National Association of Insurance Commissioners’ artificial-intelligence bulletin, including expectations around testing, documentation and risk management, but a model bulletin is not itself law everywhere. State adoption and enforcement differ.

That distinction keeps the challenge accurate. Existing claims rules may compel reasons and relevant criteria even when they never mention AI. Newer AI governance measures may require insurers to manage models internally while giving an individual claimant no direct right to a model file. Proposed obligations should not be presented as available remedies.

Escalate through the route named in the notice

The denial letter should identify the internal appeal route and deadline. Follow its submission instructions, retain proof, and state if the matter qualifies for an expedited review under the applicable rules because delay could threaten health or function. Do not wait for a separate data dispute if the appeal clock continues to run.

After internal review, some health-plan disputes may qualify for an external review by an independent organization. Public-benefit programs can provide agency hearings or other due-process procedures, while state-regulated insurance complaints may go to the state insurance department. The correct route turns on the coverage arrangement, so the plan document and denial notice matter more than a generic AI complaint form.

Human review has a real cost. It can add time, require a clinician or specialist, and still reproduce the initial result when the reviewer sees the same incomplete record. Its value comes from authority and evidence: the reviewer must be able to depart from a software recommendation, consider corrected material and give a reason tied to the governing rule.

For the original “not medically necessary” letter, the endpoint is concrete. The final file should show which clinical criterion controlled, which patient facts the reviewer accepted, how the corrected document changed or failed to change the analysis, and who had authority to issue the reconsidered decision.

Questions people ask

Do I have a right to know that AI was used?

Not under one general US rule covering every insurance or benefits decision. Disclosure duties vary by program and state, while older claims rules may require specific reasons, plan provisions and relied-upon criteria without requiring a complete description of the software.

Should I ask for the algorithm or source code?

You can ask, but source code is rarely the most useful first target and may not be available through the claims process. Request the case-specific inputs, software output, controlling criteria, factual findings and reviewer actions, which are more likely to show whether bad data or an inapplicable rule drove the result.

What if the insurer says a human reviewed the decision?

Ask who had authority to change the outcome, what materials that person reviewed and whether the reviewer could disregard the software output. A human click does not resolve an inaccurate input or an unsuitable criterion; the reconsideration should address the disputed fact and apply the governing coverage rule.

Can correcting my data reverse the denial automatically?

No. A corrected record removes one source of error, but the insurer may still conclude that the claim fails under the applicable policy or program rule. Submit the correction through the accepted channel, preserve proof and continue the appeal within the stated deadline rather than assuming the systems will synchronize.