Before an AI-Drafted Letter Goes to the Payer
In an account published in November 2024, Wendell Potter described Jennifer Braunagel, then a 52-year-old bus dispatcher and receptionist with rheumatoid arthritis, who had been denied coverage for Actemra, an infusion drug. Meritain Health, an Aetna subsidiary that administered her health plan, first required her to try a less expensive medicine. After months of calls and failed appeals, she and her doctor’s office at the Arthritis Center turned to Claimable, a company that uses AI to write appeal letters. Claimable filed the appeal in less than a day, after a 30-minute intake with her, and the drug was approved.
The account does not reproduce the letter, and it reports that a state labor official also intervened before the approval. Her case shows a doctor’s office and a patient using an outside AI service to answer a payer.
The same kind of tool can also put an inaccurate statement about a patient into a letter. For a hospital, the question is who checks the letter before the payer sees it.
In September 2026, Doximity surveyed 3,379 of its U.S. physician members across 20 specialties about their use of artificial intelligence. Eighty-five percent said they use AI for work at least weekly, and 52% said they use it every day. Among the most common writing and administrative uses over the past year, insurance letters came third, at 32% of physicians, after professional correspondence (38%) and patient education (34%).
A letter to a payer is part of the record the payer uses to decide whether to pay. Prior authorization requests, letters of medical necessity and appeals of denials all go to a reviewer who compares what the letter says with the payer’s coverage criteria. The report does not break the 32% down by type of letter, and it gives no figure for how often those letters were checked before they were sent.
What the survey says about checking
The same report gives reasons to check. Three in four physicians said they had come across a factual inaccuracy in AI output at work in the past 12 months. Two-thirds said they independently verify AI-generated clinical information before using it all or almost all the time. That question asked about clinical information, and a letter to a payer is made of clinical information: diagnoses, test results, dates of service, treatments already tried.
A study published in March 2026 shows what a check can catch. Researchers had ChatGPT-5 draft prior authorization letters for 29 kidney-disease scenarios they wrote themselves. They rated the clinical reasoning strong in 26 letters and adequate in the other three. One letter, though, supported the requested drug by citing a trial of a different drug for a different condition, and six had diagnosis coding errors, most of them chronic kidney disease coded at the wrong stage. The authors concluded the drafts “are not yet reliable enough to be used without clinician verification.” The letters came from one model in September 2025, and the study did not measure whether payers approved any of the requests.
Fewer physicians reported clear rules for AI than reported permission to use it. Eighty-one percent of the physicians said AI use is permitted or supported at their hospital or organization, and 28% said their hospital or organization offers clear policies and approvals for its use. The figure was 53% among physicians at government hospitals and clinics and 14% among physicians in private practice.
The standard the reviewer applies
For Medicare Advantage, CMS has stated the standard the plan must apply. In a February 6, 2024, memo to all Medicare Advantage organizations and Medicare-Medicaid plans, CMS wrote that MA organizations must make medical necessity determinations “based on the circumstances of each specific individual, including the patient’s medical history, physician recommendations, and clinical notes.”
The memo is addressed to plans and does not regulate hospitals. It does tell a hospital what the plan’s reviewer is required to look at. A diagnosis and a procedure code alone may not give the reviewer the facts about this patient that the request depends on. An AI draft built from only those codes can fill the gap with general information about the condition.
Where the legal exposure is
The larger risk is a letter that states something the chart does not support: a symptom, a test result or a failed course of treatment that the patient never had.
When a letter supports a claim for Medicare, Medicaid or another federal program, the False Claims Act can apply. The Act reaches claims presented to the federal government. It also reaches claims made to a contractor or grantee, such as a Medicare Advantage plan or a state Medicaid program, when the money is to be spent on the government’s behalf or to advance a government program, and the government provides or will reimburse any portion of it. § 3729(b)(2). The Act makes liable any person who “knowingly makes, uses, or causes to be made or used, a false record or statement material to a false or fraudulent claim.” 31 U.S.C. § 3729(a)(1)(B). “Knowingly” includes acting “in reckless disregard of the truth or falsity of the information,” and the Act requires “no proof of specific intent to defraud.” § 3729(b)(1). “Material” means “having a natural tendency to influence, or be capable of influencing, the payment or receipt of money or property.” § 3729(b)(4).
Each of those elements turns on the facts. Materiality is judged for the particular false statement: an invented failed treatment that meets a coverage requirement raises a different question from a wrong date. Sending an AI-drafted letter without checking its statements about the patient could bear on whether someone acted in reckless disregard, depending on what the people involved knew. A written review rule, and a record that the review was done, is evidence of what the hospital did, and it would be weighed with everything else.
Letters to commercial payers raise contract and state-law questions that this post does not cover.
What to put in place
- A written rule. Name the AI tools approved for drafting payer correspondence, who may use them, and for which letters: prior authorization requests, letters of medical necessity, appeals.
- A check of every statement about the patient. Diagnoses, test results, dates, prior treatments and functional status should each match the record. A statement the record does not support is removed, or the responsible clinician confirms it and documents it before the letter goes out. Where the record conflicts with itself, the clinician resolves the conflict before the letter repeats either version.
- A separate check of citations, codes and coverage criteria. Confirm that each cited study exists and says what the letter says it does, that the diagnosis and procedure codes fit this patient, and that the letter addresses the payer’s own criteria. A check against the chart will not catch a study cited for the wrong drug.
- A record of the review. Note who reviewed the final version and when, and keep that version with the claim file, along with the records it relies on.
- Approved tools for patient information. When drafting involves patient information, use only a tool approved for that purpose. Approval should include a privacy and security review and a business associate agreement with the vendor where HIPAA requires one.
- The same rule for billing staff and vendors, and a way to correct mistakes. Letters drafted by revenue cycle staff or a denial-management vendor need the same review. Name who is told, and who decides what to do, when an error is found after a letter is sent.
About the survey
Doximity publishes the survey and also develops and offers AI tools for physicians; its report calls the company “the leading clinical AI company.” The report says the survey was run through SurveyMonkey among Doximity members, that its sample is not population-based, and that its findings may not apply to all physicians. The figures above describe the physicians who answered.
Before a hospital sends an AI-assisted letter, it should be able to say who checked the final version and which records support each statement about the patient.
Sources
- Doximity, Doctor-Driven AI: 2026 Report (survey of 3,379 U.S. physicians, fielded September 2026; published November 2026).
- Centers for Medicare & Medicaid Services, memorandum to All Medicare Advantage Organizations and Medicare-Medicaid Plans, “Frequently Asked Questions related to Coverage Criteria and Utilization Management Requirements in CMS Final Rule (CMS-4201-F),” February 6, 2024.
- False Claims Act, 31 U.S.C. § 3729.
- Wendell Potter, “How AI Helped a Woman Overcome Health Insurance Company Gridlock for Lifesaving Arthritis Treatment,” HEALTH CARE un-covered (Substack), November 12, 2024.
- Aiumtrakul N, Thongprayoon C, Kookanok C, Poochanasri M, Phichedwanichskul K, Cheungpasitporn W. “Quality assessment of large language model–generated prior authorization letters in nephrology.” Frontiers in Digital Health 8:1767648 (published March 3, 2026). doi:10.3389/fdgth.2026.1767648.
- U.S. Department of Health and Human Services, Office for Civil Rights, “Guidance on HIPAA & Cloud Computing.”
- Meritain Health, “Frequently Asked Questions” (member FAQ, “Who is Meritain Health?”).