June 4, 2026 · ~9 minute read

How to Use ChatGPT to Appeal a Medicare Advantage Denial (and Where It Falls Short)

I want to start with something most people writing about AI in healthcare will not say. ChatGPT is genuinely useful for parts of a Medicare Advantage appeal. If you are sitting at your kitchen table with a denial letter in one hand and a laptop in the other, opening ChatGPT is not a bad instinct. I have done it myself, on my own appeals and on appeals I have helped family members with.

The problem is not that ChatGPT is useless. The problem is that the parts it does well are not the parts that decide whether your appeal succeeds. I came at this from a Health Informatics background (M.S., San Diego), and I have spent a lot of time looking at where automation actually moves the needle in payer disputes versus where it just produces nice-looking output that quietly fails.

This post is my attempt to be honest about both. Where ChatGPT helps. Where it falls down. And what to do about the gap if you are appealing on your own.

What ChatGPT actually does well

If you paste a Medicare Advantage denial letter into ChatGPT and ask it to explain what the plan is saying, you will get something useful. Denial letters are written in a kind of compressed insurance-legal dialect that is hard to read even for people who work in healthcare. ChatGPT will translate that into plain English and usually get the gist right.

It is also good at the following:

Drafting the general structure of an appeal letter (introduction, statement of facts, medical necessity argument, request for reversal, signature block).
Suggesting medical necessity arguments based on the clinical situation you describe.
Summarizing how the Medicare Advantage appeals ladder works at a conceptual level (plan reconsideration, then IRE, then ALJ, and so on).
Rephrasing your own words into something less emotional and more clinical, which tends to land better with a reviewer.
Pointing you toward the general areas of CMS regulation that apply to your situation.

That is real value. If a friend told me they were appealing a denial and had no idea where to start, I would not tell them to avoid ChatGPT. I would tell them to use it for the writing and then verify everything else.

The “verify everything else” is where this post lives.

The hallucination problem, specifically for appeals

The biggest practical risk of using ChatGPT for an appeal is not that the writing is bad. The writing is usually fine. The risk is that the model will confidently insert citations, deadlines, addresses, and contractor names that are wrong.

Lawyers have already learned this the hard way. In Mata v. Avianca, an attorney filed a brief containing six entirely fabricated case citations that ChatGPT had generated, complete with fake quotes and fake internal references. The judge sanctioned the lawyers $5,000 and the case became the standard cautionary tale about generative AI in legal work. More recently, in March 2026, OpenAI was sued in a case alleging ChatGPT helped a former claimant file 44 post-settlement documents that included fabricated citations.

An appeal is not a federal court filing, but the underlying failure mode is identical. If you ask ChatGPT for the CFR section governing Medicare Advantage reconsideration timelines, you may get 42 CFR 422 Subpart M, which is correct, or you may get a confident but fake subsection number. If you ask it for the current address for an Aetna appeals PO Box, you may get the right one, or you may get one that was correct three years ago. If you ask it which Independent Review Entity handles Medicare Advantage appeals, it will probably say Maximus, because Maximus held that contract for years and dominates the training data.

That last one matters. As of May 1, 2026, the Part C IRE contract transitioned from Maximus to C2C Innovative Solutions. Appeals received on or before April 30 still go to Maximus; appeals received May 1 or later go to C2C. Most LLM training data does not yet reflect this. I checked, and the major models will still tell you Maximus, often without hedging.

An appeal that cites the wrong regulation or sends paperwork to the wrong contractor is worse than no appeal at all. It signals to the reviewer that you do not know what you are doing, and it gives them an easy procedural reason to deny without engaging the medical question.

What ChatGPT cannot do at all

Beyond hallucination, there are things ChatGPT structurally cannot do, no matter how good the next model release is. These are the things that, in my experience, actually decide outcomes.

It cannot read your specific denial letter and rebut the plan's actual cited criteria. It can summarize the letter in general terms. It cannot pull out the specific InterQual or MCG criterion the plan referenced and explain why your clinical situation meets it.
It cannot certify itself as the author of record. This sounds like a technicality and it is not. Appeals get weight partly from who signs them. A patient, a treating physician, or a designated representative carries authority. An anonymous chatbot does not, and increasingly the unauthorized-practice-of-law issue is becoming an explicit legal question.
It cannot print, stuff, and physically mail a certified letter to the plan's appeals PO Box with a tracking number. The evidentiary record matters. State Departments of Insurance and the IRE both want to see that something was actually delivered, when, and to whom. A USPS Certified Mail receipt is the cheapest, most boring, most legally durable form of that proof.
It cannot track the 30-day and 60-day clocks for plan response and auto-escalate if the plan misses them. Under CMS rules, plans must respond to a standard pre-service reconsideration within 30 days and a payment reconsideration within 60 days. If they miss, the case auto-forwards to the IRE. Knowing the rule is not the same as actually counting the days and escalating on Day 31.
It cannot file a complaint with your specific state Department of Insurance. Each state DOI has its own form, its own submission portal, and its own preferences for documentation. ChatGPT does not know which state you are in unless you tell it, and even then it does not have a live connection to your state's commissioner office.
It does not reliably know which IRE currently holds the contract. See the C2C versus Maximus issue above.
It cannot pull NCD and LCD numbers from primary CMS sources. It will guess based on what coverage determinations look like in its training data. Sometimes that guess is right. Often it is close but not exact, which is worse, because a wrong NCD number reads as a typo at best and as fabrication at worst.
It cannot avoid hallucinating statute numbers, deadlines, and addresses when you push it on details. It will produce something that looks correct because looking correct is its job.

The “verify every fact” workflow if you are doing this solo

If you are going to use ChatGPT for an appeal anyway, which is a reasonable thing to do, here is the workflow I would actually recommend.

Use ChatGPT to draft the medical necessity argument and to translate the denial letter. This is the part it is good at.
Separately, open the CMS website and pull the actual reconsideration form (CMS-20033 for IRE-level reconsideration, available on CMS.gov). Do not let ChatGPT generate this form. Use the real one.
Cross-check every regulatory citation against eCFR.gov. Medicare Advantage appeals live in 42 CFR Part 422 Subpart M. If a citation ChatGPT gave you does not resolve to a real section, cut it.
Cross-check every deadline against the CMS reconsideration page. The big ones are 60 days from the denial notice to file your reconsideration, 30 days for the plan to respond to a service request, 60 days for a payment request.
Confirm the current IRE before sending anything to it. As of this writing it is C2C Innovative Solutions for cases received May 1, 2026 or later. Check the CMS IRE page for the current contractor and address.
Pull the plan's appeals address from a current member handbook or the plan's website. Do not trust ChatGPT for an address. Insurer PO Boxes change.
For state DOI complaints, go directly to your state insurance commissioner's website. The complaint form will be there. Do not let ChatGPT generate a fake one.
Send the appeal by USPS Certified Mail with Return Receipt. Keep the green card. This is your evidentiary anchor.
Calendar the response deadline the day you mail. If it passes, escalate the same day.

This is a lot of homework. It is the homework that actually wins appeals.

Why DenyBack separates AI language from regulatory data

When I built DenyBack, this is the architectural choice I made that I think matters most. The AI generates persuasive language only. Every regulatory fact, every address, every form number, every IRE name, every state DOI commissioner, every CFR citation comes from hardcoded primary-source databases that I maintain by hand against CMS, eCFR, and the state insurance commissioner sites.

The AI never gets to guess at a regulation. It writes the medical necessity narrative and the framing. Then a separate layer injects the verified regulatory references into the letter. If C2C replaces Maximus, I update one row. If a state changes its DOI complaint URL, I update one row. The model does not learn this and then drift on it.

This is not a technical brag. It is the only way I have found to use a language model in a setting where one fabricated CFR citation can sink the whole appeal.

If you want the AI without doing the verification homework

That is what DenyBack is for. $39 flat, one time, for the full 45-day appeal cycle. The fee covers the AI-drafted appeal letter, the certified mail to the plan on Day 1, automatic escalation by certified mail on Day 14 to the plan's grievance office, Day 30 to the state DOI, and Day 45 to the IRE if the plan has not overturned. All regulatory references come from primary sources, not from the model.

You can absolutely do this yourself with ChatGPT and a couple of hours of homework per escalation step. If you would rather not, that is what we built.

Permission to ignore the pitch

If you have time and patience and you would rather drive this yourself with ChatGPT plus the verification workflow above, that is a completely legitimate path. The appeals system is winnable solo. It just takes the kind of attention to procedural detail that most people in the middle of a medical crisis do not have.

Either way, the most important thing I can leave you with is this. Use ChatGPT for the language. Never for the legal facts. The facts have to come from a primary source you can point to, every time.