The Denial Letter, Reversed: How Regular People Are Using AI to Fight Their Health Insurance — and Winning

There's a folder on a lot of kitchen counters this spring that didn't used to be there. It usually starts with one letter — a "we have processed your claim" envelope from an insurer with a number on it that doesn't make sense — and within a week it's got three or four more: the EOB, the bill from the hospital, the explanation that quietly says denied, the second bill with a late fee tacked on. For most of the last twenty years, that folder went one of two places. It went into a drawer and got paid resentfully a few months later. Or it went into a longer drawer and turned into a collections call.

This spring, a third thing is starting to happen. Somebody opens a chat window, drops the photos of the letters in, and types a single paragraph: I had a baby in March, this is the bill, the procedure code is here, and this is why they're saying it's not covered. Help me write an appeal. Twenty minutes later there's a three-page letter with cited policy language, a request for the specific reviewer's credentials, and a deadline calendar. Two weeks after that, a surprising number of those folders are getting closed not by a payment, but by a reversal.

It's the quietest, most consequential AI shift we've covered all year.

What actually changed

For most of the history of health insurance in this country, the appeals process was designed around a simple, unkind fact: almost nobody appeals. In 2023, insurers denied about 73 million in-network claims on Affordable Care Act plans, and fewer than one in a hundred were appealed. The system wasn't built to be navigated; it was built to be given up on. The forms were long, the deadlines were short, the language was clinical, and the person on the other end was an adjuster reading hundreds of letters a day.

What changed wasn't the system. It was who could write the letter.

Insurers started using AI to deny claims, and patients started using AI to appeal them. This is the part nobody planned. Over the last two years, several large insurers rolled out machine-learning systems to flag claims for denial — sometimes faster than a human could meaningfully review them. Patient-side tools showed up on the other end of the same pipe. Counterforce Health, a nonprofit funded by the University of Pennsylvania and the NIH, built a free appeal generator that's helped roughly twenty thousand Americans secure coverage and reports a 70–80% success rate on the appeals it drafts. Sheer Health built an app that connects to your insurance account directly and answers questions about your benefits in plain English. A handful of smaller services do versions of the same thing, with different privacy postures and different specialties.

The general-purpose chat assistants got good enough to do the same job. This is the underrated half of the story. You don't need a dedicated tool. The current generation of ChatGPT, Claude, and Gemini is genuinely capable of reading a denial letter, cross-referencing your plan documents, and drafting an appeal that cites the right policy language and the right medical evidence. Mayo Clinic's patient advocates have published step-by-step guides for doing this with the free tier of any of the three. The workflow is not particularly clever. It's that the cost of producing a competent appeal letter dropped from a half-day of confused effort to a twenty-minute conversation — and at that price, people who would have folded started filing instead.

The success rate on appeals was always quietly high. This is the lever underneath the whole shift. For decades, the data has been weirdly consistent: the small fraction of patients who appeal a denial win their appeal a majority of the time. The barrier was never that the appeals don't work. The barrier was that almost nobody files one. Lower the cost of filing, and the math of the whole system flips.

The honest summary: AI didn't invent a better appeals process. It removed the activation energy from the one that already existed. Once the letter was free to write, the wins were already on the table.

What the workflow actually looks like

The marketing pages talk about AI-powered medical advocacy. The real workflow is much smaller and much less dramatic. Here's what an actual session looks like at the kitchen tables we've been hearing from this spring.

The trigger. A bill or denial letter arrives that doesn't add up. The categories that keep showing up are pretty predictable: a routine procedure flagged as not medically necessary, an in-network provider billed as out-of-network because of a paperwork error, a maternity charge that should have been bundled, a prior authorization for a chronic condition denied for the third time, an emergency room visit re-categorized as non-emergent after the fact. In every case the patient's first reaction is the same — this has to be wrong — followed by a familiar resignation about what it would take to prove it.

The first prompt. The patient opens whichever chat tool they already use and types something like: "I have a denied health insurance claim. I'll share the denial letter, the relevant page of my plan documents, and the bill. Walk me through whether this is appealable, and if so, draft a strong appeal letter." Then they paste — or, increasingly, take a photo of and upload — the actual paperwork. The assistant reads it and asks one or two clarifying questions: what state are you in, was the provider in-network at the time of service, do you have the procedure code. These are the questions a good patient advocate would ask, and they're the ones most people don't think to answer in their first frustrated draft.

The diagnosis. The model reads the denial reason against the plan language and tells the patient, in plain English, what they're actually looking at. Sometimes this is the most useful part of the whole conversation. "The denial cites 'not medically necessary,' but the procedure code you have is a covered preventive service under your plan's Section 4.2, which is the section they're appealing under. The appeal has good standing." Or, just as often, the opposite: "This denial is technically correct under your plan's exclusion list — the appeal route here isn't to argue coverage; it's to ask the provider to recode the service." That second sentence saves people from filing a doomed appeal and pushes them at the actual lever.

The letter. The model produces a draft that, if you've ever seen a successful appeal letter, looks the part. A statement of the facts. The relevant policy citations from the patient's own documents. The relevant clinical evidence (peer-reviewed when it's a treatment question, billing-code documentation when it's a coding question). A specific ask: reverse the denial, cover the claim, refund the patient for any payments already made. A deadline acknowledgment. A request for the credentials of the reviewer if the appeal is denied again — because under most state laws, a medical necessity denial has to be reviewed by a physician in the same specialty, and asking for that record up front matters at the next level.

The polish. The patient reads the draft, fixes anything that's wrong, adds any personal context the model couldn't have known, double-checks the citations against the actual plan documents, and sends it. The whole exchange takes between fifteen minutes and an hour, depending on how complicated the underlying claim is. The historical version of this exercise — same letter, no chat — was the kind of project people put off for a weekend and then put off for another weekend until the deadline passed.

The result. This is the part that surprises most first-time users. A meaningful share of these appeals just work. The denial gets reversed at the first level. The bill gets adjusted. The folder gets closed. The patient who'd already accepted the loss — who'd budgeted for it, who was waiting for the collections call — instead gets a letter back saying upon further review.

It is not magic. It is not always going to work. But the rate at which it works for ordinary people writing letters they wouldn't otherwise have written is high enough that the calculus of should I bother to appeal has fundamentally changed.

What people are actually using it for

The launch story is always the dramatic appeal — the cancer treatment, the surgery, the five-figure save. Those are real. The day-to-day usage is much smaller and much more common.

The "wait, this is wrong" surprise bill. A new parent who got an unexpected $2,000 charge from her insurer two years after giving birth — that one made the news this spring when the chat-drafted appeal reversed it cleanly. Less dramatic versions of the same story are happening every week: the out-of-network anesthesiologist who showed up at the in-network surgery, the urgent care visit billed twice, the lab result re-categorized after the fact. These are the appeals that used to go unfiled because the patient assumed the system was correct. The new default is to ask the assistant first.

The prior authorization wall. This is the use case that's growing fastest, and it's where the patient-side AI is starting to genuinely change outcomes. Prior auth denials are the bureaucratic chokehold on chronic care — the medication that's been working for six months suddenly requires re-justification, the specialist visit that the patient and the doctor both think is obvious gets flagged for review. The appeal letter is procedural and repetitive, which is exactly the shape of writing the models are best at. New federal rules require standard prior authorization decisions inside seven days; the practical effect of that, combined with patients who can produce a strong appeal in twenty minutes, is a meaningful number of approvals that wouldn't have happened a year ago.

The "is this even a real bill" sanity check. Sometimes the appeal isn't an appeal. It's a translation. The patient pastes the bill into the chat and asks the assistant to explain what the EOB actually says, what each line item is, and what the patient is actually being asked to pay versus what the insurer should have paid. A lot of the wins in this category aren't about overturning a denial. They're about catching billing errors before they're paid, or noticing that the same service was billed twice, or realizing that the "amount you owe" line is wrong because it didn't apply a credit. The chat doesn't pay the bill. It just makes the bill legible — and a legible bill is a bill that gets argued.

The chronic-condition fight, on repeat. For people managing chronic conditions, the denial cycle is a kind of seasonal weather pattern. The same medication, the same therapy, the same equipment — denied, appealed, approved, denied again at the next renewal. The chat assistant, especially with memory turned on, has become a quietly powerful tool here. It remembers the patient's diagnosis, the relevant policy language from last year's appeal, the clinical evidence that worked last time. The third appeal in a sequence is much easier to write than the first. For families managing serious illness, that's not a small efficiency win — it's the difference between continuing the treatment and not.

The "I want my doctor on my side" prep. A lot of successful appeals require a letter of medical necessity from the treating physician. Doctors don't have time to write these, and historically the patient either lived without the letter or got a one-paragraph version that didn't help. A new pattern showing up in chronic-illness communities: the patient uses the chat assistant to draft the letter — the long version, with citations — and then sends it to their doctor with a note that says if this matches your view, please print and sign on letterhead. Almost every doctor we've heard about says yes. The doctor was always on the patient's side. The cost of producing the artifact was the obstacle.

The provider-side use the patients don't see. Worth flagging because it changes the dynamic. The same AI tools are being adopted by doctors' offices and hospital billing departments to appeal claims on the patient's behalf. Some of the country's larger health systems now run AI-assisted appeals automatically on denied claims for their patients. The patient often never finds out. The reversal just shows up in the mail. If your bill is coming from a major health system, it's worth asking the billing office whether they've already appealed before you write your own letter — you may be duplicating effort, or you may be the second line of defense that pushes a borderline case over.

Where it falls apart

The reason to be careful here is the same reason this matters. The stakes are higher than in most of the AI workflows we cover on this site. A wrong answer about your taxes is bad; a wrong answer about your insurance can leave you with the original bill and a botched appeal that locks in the loss. Worth being specific about the failure modes.

• The chatbots will make up citations, confidently. This is the single biggest pitfall, and it's the one Mayo Clinic's patient advocates flag first. Ask ChatGPT or Claude or Gemini to cite a study supporting your appeal, and you may get back a citation that looks completely real — correct journal name, plausible author, persuasive title — and is entirely fabricated. Insurers' appeals reviewers check citations. A fake one in your letter doesn't just fail; it discredits the whole document. The working rule: any cited study, statute, or guideline has to be verified against an actual source — PubMed, the actual plan document, the state insurance commissioner's website — before the letter goes out. The chat is not the source. It's the drafter.
• The model doesn't know your specific plan. The general-purpose assistants know how insurance works in the abstract. They don't know what your plan covers, what your deductible looks like this year, or what your state's specific consumer-protection laws are. If you don't paste in your actual plan documents and your actual denial letter, the appeal will be written against an average plan that isn't yours. Specificity is the whole game.
• The appeals process is full of hard deadlines. Most internal appeals have to be filed inside 180 days of the denial. Expedited appeals — when a delay would seriously harm the patient — have to be filed inside 72 hours and decided inside 72 hours. State external review windows vary. The chat assistant can help you map the deadlines, but it has been wrong about specific numbers, and a missed deadline is fatal to the appeal. Always confirm the actual filing window against the denial letter you received and your state's insurance commissioner page.
• Privacy is a serious tradeoff and you should decide it consciously. Pasting a denial letter into a chat means pasting your name, your insurance ID, your diagnosis, your procedure codes, and often your address into a system run by a private company. The major chat assistants have settings that let you turn off training on your conversations; the dedicated tools like Counterforce Health have privacy postures built specifically for medical information. Two practical moves: (1) before you upload, redact what the appeal doesn't need — full account numbers, family member names that aren't relevant — even if it's a slight pain; (2) check the data-retention settings on whichever tool you use, and pick a tool whose posture you can live with. The AI Safety & Privacy Checklist is the right starting point if you want the general framework.
• It can't tell you when to escalate to a human. Some appeals belong with a licensed patient advocate or a healthcare attorney from day one — large balances, denials involving life-threatening conditions, denials that have already been upheld at internal appeal, anything where the carrier is acting in bad faith. The chat will happily draft a letter in every one of those cases, and the letter will probably even be pretty good. But the patients we've heard from who got the best outcomes in the hardest cases usually had a real human in the loop somewhere — the state insurance commissioner's office, a nonprofit advocacy group, or an attorney working on contingency. The chat is the first draft, not the whole strategy.
• It is not a substitute for talking to your doctor. This needs saying. The appeal is about the coverage of the care. The care itself is a clinical decision, and the clinical decision is between you and your physician. If your insurer denies a treatment your doctor recommends, the appeal might fix the coverage problem. It doesn't change the underlying medicine. Use the chat for the paperwork. Use the doctor for the care.
• The system will keep evolving on both sides. Insurers are already starting to deploy AI-flagging tools that look for patterns in patient-side appeals. The dynamic is going to keep moving. The right posture isn't I have figured out the cheat code. It's I have a tool that lets me participate in a process I used to be locked out of. The wins keep being real even if the specifics keep shifting.

A useful working rule, borrowed from one of the patient advocates we spoke to: use AI to write the appeal you would have written if you had a free weekend and a paralegal. Then verify the parts that matter against the actual sources. That's the whole protocol.

How to try it this week

You don't have to be in the middle of a fight to learn the workflow. The right time to figure out the appeals process is before you need it.

1. Pull the last "explanation of benefits" you got. Not a denial — any EOB. Open the chat tool you already use. Paste the EOB in (redacting your account number and date of birth first) and ask: "Explain this EOB to me line by line, in plain English. What does each charge mean, what did the insurer pay, what am I being asked to pay, and is there anything that looks unusual?" Read the answer slowly. The point of this exercise isn't the answer. It's getting comfortable with the conversation.
2. Find your plan's Summary of Benefits and Coverage document. Every plan has one. It's usually a PDF available through your insurer's member portal. Download it. The next time you have a bill or a denial you want to fight, you'll need this file in front of you. Get it now while there's nothing on fire.
3. If you have a denial sitting in a drawer, dig it out. A lot of appeals are still inside the 180-day window even if you'd given up on them. Check the date on the letter. If you're inside the window, paste the denial letter and the relevant page of your Summary of Benefits into the chat and ask: "Is this appealable, what's the strongest argument, and what would the first-level appeal letter look like?" Read the diagnosis carefully before you decide whether to file.
4. Verify everything before you send. This is the discipline that separates a good appeal from a self-discrediting one. Every policy citation has to be checkable in the actual plan document. Every clinical citation has to be checkable on PubMed or another real source. Every regulatory citation has to be checkable on your state insurance commissioner's website. If the chat invents something, you find it now, not the reviewer.
5. Send the letter on paper, with tracking. Most appeals departments accept submissions through the member portal, and that's fine, but a paper letter with a tracking number, sent to the address on the denial letter, creates a paper trail that's hard to lose. This is one of the few places where the old-fashioned move is still the right one.
6. Calendar the deadline. The insurer typically has 30 to 60 days to respond to a first-level appeal. Put the deadline in your calendar with a 7-day buffer. If they miss the deadline, that's information you can use — in some states it triggers automatic escalation rights — and the chat assistant can help you draft the follow-up.
7. Know where the human help lives, before you need it. Every state has a Department of Insurance with a consumer assistance line. Most states have nonprofit patient advocates who'll help for free or low cost. If the first-level appeal fails and you're in a state with an external review process, that's typically your next move. Bookmark the page for your state now. The AI handles the paperwork. The humans handle the harder cases.

For the longer reference on AI in this corner of the world — what it does well and what it's still bad at across actual healthcare work — the Healthcare career guide is the next stop. For prompts you can adapt for the specific letters above, the Prompt Library has versions you can copy.

The bigger shift

A pattern keeps showing up in the AI shifts we've covered this year. The biggest changes don't come from new models. They come from new shoulders the same model can lift from. The chat lived in the browser, then on a walk, then in your headphones, then on the strip of plastic above your nose, then at the kitchen table on Sunday night, then at the homework table at 7:30. This week, it's sitting on the kitchen counter with the bill folder open.

What's different about this one is the stakes and the politics underneath it. For a long time, the appeals process was a slow asymmetric war that the patient side mostly lost by default. The insurer had specialists writing the denials and AI systems pre-filtering the claims. The patient had a stack of mail, a half-hour after dinner, and a vocabulary they didn't really speak. The thing that's leveling — and it is genuinely leveling, in real numbers — is that the patient now has a competent drafter on their side too. The same kind of system that's flagging their claim is helping them write the letter back. The volume of appeals is up. The success rate of appeals, which was always quietly high, is starting to actually matter at the population level. State legislatures and federal regulators are watching.

The risks are the ones we keep flagging. People who don't know about these tools — who don't know that the chat can read the denial letter, who don't know that the appeal is winnable, who don't know that the deadline is 180 days — keep paying bills they shouldn't pay. That's a gap that compounds: a few thousand dollars in unfought medical bills here, a credit hit there, a payment plan that quietly determines the next five years. The job we've taken on at this site, more than anything else, is to make sure that gap doesn't keep widening for the people who care about this stuff and are just a step behind.

The upside is the one worth ending on. A parent who fought a maternity bill and won. A patient who got her cancer treatment covered because the appeal she wrote in twenty minutes cited the right twelve studies. A retiree who realized — for the first time in a decade of paying surprise charges — that he'd been entitled to argue and didn't know. None of these are headline use cases. They are the kind of small, repeated wins that change a household's trajectory over a decade.

Pick the next bill that doesn't make sense. Don't put it in the drawer. Put it in a chat. Read the answer. Write the letter. If it works, the folder closes. If it doesn't, you've learned more about how the system works than most people learn in a lifetime of paying it. Either way, you stopped being the part of the system that was designed to give up.

A few places on the site that pair naturally with this:

• The Healthcare career guide is the right longer read for how AI fits into actual healthcare work in 2026 — including the provider side of the appeals story.
• The AI Model Comparison covers how the major chat assistants stack up on the kind of document-reading and letter-drafting this workflow depends on.
• The Prompt Library has reusable prompts you can adapt for EOB explanations, denial appeals, and letters of medical necessity.
• The AI Safety & Privacy Checklist is the right read before you upload anything with your name, your diagnosis, or your insurance ID on it.
• For the broader ethical frame, Ethical AI Usage covers the principles underneath the verify-before-you-send discipline this workflow lives or dies by.

info

This content was developed with AI assistance and is regularly reviewed for accuracy. It is informational, not legal or medical advice — for decisions that affect your health, your coverage, or a contested bill, talk to a licensed professional or your state's consumer assistance office.