The ChatGPT Prompt to Appeal a Denied Insurance Claim (And the One Thing It Can't Know)

Which reason code is on your denial letter? That one line decides whether the advice you are about to follow was written for your situation, and most of it was not.

KFF, analyzing HealthCare.gov transparency data for plan year 2024 and publishing on March 24, 2026, found that insurers cited medical necessity in 5% of denied in-network claims. The unspecified "other" category took 36% and administrative reasons another 25%. So roughly 61% of denials are coding, routing and paperwork, and 5% are the clinical fight that every "use ChatGPT to appeal your denial" article on the first page of Google is built around.

That inversion is the whole reason this piece exists. Arguing medical necessity with an eloquent letter is the right move for one denial in twenty. For the other nineteen the job is closer to what an AI pass over a hospital bill does: put two documents side by side and find the place where they disagree.

What are health insurance claims actually denied for?

Most denials are administrative, not clinical. Here is the 2024 breakdown for in-network claims.

The scale underneath those percentages is worth holding onto. Insurers reporting on HealthCare.gov received about 496 million claims in 2024. Of the 451 million that were in network, 19% were denied, which is roughly 85 million denied claims, and the rate ranged from 3% to 36% across the 157 reporting insurers. Fewer than 1% of those denials were appealed internally, at 262,982 appeals. Insurers upheld 66% of the appeals they received. And at least 5,881 claims reached an independent external reviewer, which works out to somewhere near one in fourteen thousand denials.

Read those numbers in order and the system's actual mechanism shows up. It does not win most disputes. It wins because there is no dispute. Roughly 34% of the people who bothered to file an internal appeal got the denial reversed.

How much of the appeal can AI actually do?

Enough to matter on the paperwork denials, and nothing on the deadlines. I built one denial packet with six problems in it and ran it twice, changing only the prompt.

The packet is three documents, which is what a real person has after a denial. An explanation of benefits with a single claim line, the denial letter itself, and the plan's published prior-authorization list for outpatient advanced imaging.

The denial letter states the reason in prose, names a clinical policy called RAD-104 without attaching it, and gives 60 days to appeal in writing. The prior-authorization list names CPT 70551, 70552, 70553, 73721, 74177 and 74178.

The six planted problems: CPT 72148 is missing from the plan's own prior-auth list that the denial rests on; the radiologist is out of network inside an in-network facility; the allowed amount is $0.00 while the member owes the full billed charge; the 60-day appeal window sits below the federal floor; the letter never describes the rule it relied on or gives the codes and their meanings; and nothing in the packet says whether the plan is self-funded or fully insured.

One caveat on method, since it changes how much the score is worth. I wrote the packet, so I knew what was in it, and this is not a blind test of recall. What it does test is which prompt shape surfaces which problem, and what the model volunteers when nobody asks it to.

What happens when you just ask for the appeal letter

You get a good letter with a bad instruction at the end. This is the first run, using the prompt almost every published guide gives you.

Three of the six problems are in there, argued well. The letter finds the missing CPT code, the network mismatch and the arithmetic, which is exactly the class of thing a language model is good at, because all three are one document disagreeing with another.

Then look at the last line before the signature. It sends you to a state department of insurance. For most readers that agency has no authority over the plan at all, and the letter says it with the same steady confidence as the parts that are right. It also tells you to check your deadline instead of telling you the deadline in the letter is too short, which is the difference between a note and a finding.

The prompt that changes the answer: analyze first, write later

Same documents, same model, one added clause. Do not write a letter yet, and tell me what you cannot determine.

The scoreboard is not the point. The two runs came from one model on one afternoon, and the entire difference is a sentence asking it to name its own blind spot. Left to write, it wrote. Asked what it could not see, it said so and stopped.

I read that as the practical rule for any document a model is about to argue from. Make it inventory the file before it advocates, because advocacy has no natural place to put the words "I do not know."

The one thing an EOB cannot tell you: which rulebook covers you

Your plan type decides who can help you, and your insurance card will not tell you which one you have.

If your employer self-funds, your employer pays the claims out of its own money and hires a familiar insurance company only to administer them. The card carries that company's logo either way. Self-funded plans are governed by ERISA and sit largely outside state insurance law, so a state insurance department generally has no jurisdiction and the state external review process may not apply to you. KFF's 2025 Employer Health Benefits Survey puts 67% of covered workers in self-funded plans, including 80% at firms with 200 or more workers and 27% at firms with 10 to 199 workers.

That is why the letter run's closing sentence is worse than a wasted stamp. You get a confident, well-written instruction to contact an agency that will write back to say it cannot help, after some of your 180 days have already burned. One phone call to your benefits office answers the question. Ask whether the plan is self-funded or fully insured, and ask for the summary plan description in writing.

What is the appeal sequence, and what are the deadlines?

Two stages, both with hard clocks. The internal appeal goes to your plan, and the external review goes to an independent organization that your plan does not control.

The external stage carries protections worth knowing before you decide it is not worth the trouble. The independent reviewer decides your claim de novo, meaning it is not bound by anything your plan concluded. It has to weigh your medical records, your doctor's recommendation, the plan terms, evidence-based guidelines and the plan's own clinical review criteria. Your plan must contract with at least three reviewers and rotate assignments among them, and may not pay them in a way tied to upholding denials. The decision binds the plan, and the federal process costs you nothing. A grandfathered state process may charge a filing fee capped at $25, refunded if you win.

One clock runs against the plan too. Once a reviewer is assigned, the plan has five business days to hand over the claim file, and if it misses that deadline the reviewer may terminate the review and reverse the denial outright.

Two document requests that beat a well-written letter

Ask for the file and check the notice for defects. Neither move requires a single good sentence, and both do more work than persuasion.

The first is the claim file. Under 29 CFR 2560.503-1(h)(2)(iii) you get reasonable access to, and free copies of, all documents, records and other information relevant to the claim. That is how you obtain the clinical review criteria the plan applied, which is the same document the independent reviewer is required to weigh later. Ask for it in writing, name the date of service, and name the policy the denial letter cited.

The second is a procedural exit almost nobody mentions. If the plan fails to strictly adhere to the internal claims and appeals requirements, you are deemed to have exhausted the internal process and may go straight to external review or to court, under 45 CFR 147.136. There is an exception for de minimis violations that cause no prejudice, and you can demand a written explanation of the violation, which the plan must supply within ten days with the specific basis for claiming the exception. A denial letter that gives you 60 days when the floor is 180, or that never describes the rule it relied on, is not merely sloppy. It is a fact you can use.

This is where AI earns its keep, incidentally. Comparing a notice against a list of required elements is dull, mechanical reading, and it is the same skill that makes it useful on errors in a hospital bill or on a credit report dispute under the FCRA.

Where this approach fails

Four ways, ordered by how much damage each one does.

The record problem is the serious one. In an ERISA plan, courts reviewing a benefits denial generally confine themselves to the administrative record, which is the claim file plus what you submitted during the internal appeal. Content you put there cannot reliably be corrected later, so a fabricated citation is not a bad draft but a permanent fact about your case. With 67% of covered workers in self-funded plans, that is the ordinary case rather than an edge case.

The citation problem is specific and hard to catch. The failure mode that gets people is not an invented rule, it is a correct legal proposition with a fabricated citation attached, because the substance reads as true and only the reference is fake. Open every statute, regulation and plan provision a model cites before any of it goes into a record. My own runs above name provisions I checked at Cornell's reproduction of the CFR afterward, which is the only reason I am willing to print them.

The evidence problem is the quiet one. AI improves the writing and cannot produce your physician's letter of medical necessity, your chart notes, or the plan's clinical criteria. Internal appeals were reversed about 34% of the time in 2024, and the two moves most likely to change your outcome are both document requests rather than arguments.

The clock problem is the one that ends things. A chat window does not track your 180-day internal window or your four-month external window and will not warn you when either is closing. Everything else in an appeal can be retried.

There is also an asymmetry worth naming. Insurers are the heavier AI users in this exchange. Cigna's PXDX system is alleged to have supported denials at roughly 1.2 seconds per claim according to ProPublica's reporting, and UnitedHealth's nH Predict is the subject of a class action over Medicare Advantage post-acute denials, with a Minnesota magistrate judge ordering disclosure of how the tool works on March 9, 2026. Both of those reach me through trade and general press rather than dockets I opened, and the widely repeated 90% error rate attached to nH Predict comes from the plaintiffs' complaint, not from any finding. Treat it as an allegation.

On privacy, the short version: you cannot violate HIPAA with your own health information, since the law binds providers, plans and their business associates rather than patients. The exposure is retention and discoverability of chat logs, which the hospital bill piece covers at length. Strip your name, member ID and date of birth before pasting. The codes are what the analysis needs.

FAQ

Can ChatGPT write my health insurance appeal letter?

It can write the letter, and the letter is rarely the part that decides the claim. I ran a constructed denial packet twice on August 13, 2026. Asked for a letter, the model produced a clean two-argument appeal in about 350 words and closed by telling me to file a complaint with my state department of insurance. That instruction is wrong for most readers, because 67% of covered workers are in employer self-funded plans under ERISA, which state insurance departments generally do not regulate (KFF 2025 Employer Health Benefits Survey). Asked to analyze the documents first and to name what it could not determine, the same model flagged the same contradictions, added the two the letter run walked past, and said plainly that it could not tell whether the plan was self-funded or fully insured. Same model, same documents, one clause of difference in the prompt. Use the analysis prompt first, then let it draft. And check every statute and section number it hands you before any of it goes into an appeal.

How long do I have to appeal a denied health insurance claim?

At least 180 days from the date you receive the adverse benefit determination, for an internal appeal on a group health plan, under 29 CFR 2560.503-1(h)(3)(i). If your denial letter gives you less than that, the letter is the thing that is wrong, not your calendar. After the internal appeal, you have four months from receipt of the final adverse determination to request external review by an independent organization, under 45 CFR 147.136. The plan then has five business days for a preliminary eligibility check and one business day to tell you the result, you get ten business days to send the reviewer anything extra, and the reviewer issues a written decision within 45 days. Urgent cases run on a 72-hour clock at both stages. These are the numbers a chat window will not track for you, and a missed deadline is the one part of an appeal that cannot be redone.

What is external review, and is it free?

External review is a fresh look at your denial by an independent review organization that has no stake in the outcome, and under the federal process it costs you nothing. The reviewer decides the claim de novo, meaning it is not bound by anything your plan concluded internally, and its decision binds the plan. Your plan must contract with at least three review organizations and rotate assignments among them, and it may not pay a reviewer in a way tied to how often denials are upheld. The reviewer has to consider your medical records, your doctor's recommendation, the plan terms, evidence-based guidelines and the plan's own clinical review criteria. A grandfathered state process may charge a nominal filing fee capped at $25, refunded if the denial is reversed. All of that sits in 45 CFR 147.136. Almost nobody uses it: KFF counted at least 5,881 external appeals filed in 2024 against roughly 85 million denied in-network claims.

Why does it matter whether my health plan is self-funded?

Because it decides who has authority over your appeal, and you usually cannot tell from your insurance card. In a self-funded plan your employer pays the claims and hires a familiar insurer only to administer them, so the logo on the card is the same either way. Self-funded plans are governed by ERISA and are generally outside state insurance law, which means a state insurance department has no jurisdiction and the state external review process may not apply. KFF found in its 2025 Employer Health Benefits Survey that 67% of covered workers are in self-funded plans, including 80% at firms with 200 or more workers and 27% at firms with 10 to 199. There is a second reason to care. Courts reviewing an ERISA denial generally look only at the administrative record, which is the claim file and what you submitted during the internal appeal, so a wrong fact in an AI-drafted appeal is not a draft. Ask your benefits office which type you have before you write anything.

Disclaimer

This article is educational and is not legal advice, medical advice, or a recommendation of any insurer, plan, provider or service. The denial packet is constructed for this test and describes no real claim, no real provider and no real plan. The AI output above is a real run on Claude Opus 5 on August 13, 2026, and a different model, prompt or day will produce different output. Regulatory text was read at Cornell's Legal Information Institute reproduction of 29 CFR 2560.503-1 and 45 CFR 147.136 on August 13, 2026, because eCFR blocked automated access that day; confirm each provision at eCFR before relying on it. If your plan is employer self-funded and therefore governed by ERISA, what you submit during the internal appeal may become the permanent record a court later reviews, so have a benefits attorney read the appeal before you send it on any claim large enough to matter. Deadlines and plan terms vary, and only your plan documents govern your claim.

Before the denial there is the bill itself, and the same grading method run over a hospital bill is where this workflow starts. If the unpaid balance has already reached your credit file, disputing it under the FCRA is the next document. For the same method applied to a lease rather than a claim, see the nine lease clauses courts refuse to enforce and the prompt for reading a lease before signing. On the insurance side of the ledger, checking a car insurance renewal increase and pricing renters insurance against what you actually own use the same read-the-document habit. And if the balance is already sitting there earning interest, an audit of what your bank charges and a payoff order for what you owe are the two places that money usually needs to go.

Sources

• KFF, "Claims Denials and Appeals in ACA Marketplace Plans in 2024," published March 24, 2026 (496 million claims; 19% in-network denial rate; 37% out of network; 3% to 36% range across 157 insurers; denial reasons: other 36%, administrative 25%, excluded service 13%, prior authorization 9%, medical necessity 5%; 262,982 internal appeals, under 1% of denials; 66% upheld; at least 5,881 external appeals): https://www.kff.org/patient-consumer-protections/claims-denials-and-appeals-in-aca-marketplace-plans-in-2024/
• KFF, "Claims Denials and Appeals in ACA Marketplace Plans in 2023," for the prior-year comparison (56% of internal appeals upheld, so 44% overturned): https://www.kff.org/private-insurance/claims-denials-and-appeals-in-aca-marketplace-plans-in-2023/
• KFF, 2025 Employer Health Benefits Survey (67% of covered workers in self-funded plans, including 27% at firms with 10 to 199 workers and 80% at larger firms; 37% of covered workers at firms with 10 to 199 workers in level-funded plans), figure confirmed at the primary page on August 13, 2026: https://www.kff.org/health-costs/2025-employer-health-benefits-survey/
• 29 CFR 2560.503-1, ERISA claims procedure (at least 180 days to appeal, (h)(3)(i); 72-hour urgent decisions, (f)(2)(i) and (i)(2)(i); 15-day pre-service and 30-day post-service initial decisions with one 15-day extension; free copies of all relevant documents, (h)(2)(iii); fresh reviewer with no deference to the initial determination, (h)(3)(ii)), read at Cornell LII on August 13, 2026: https://www.law.cornell.edu/cfr/text/29/2560.503-1
• 45 CFR 147.136, external review (four-month filing window; 5 business day preliminary review and 1 business day notice; 48-hour cure; 10 business days for claimant submissions; 5 business days for the plan to send the file, with reversal available if it misses; 45-day decision; 72-hour expedited; $25 nominal fee cap in a grandfathered state process; at least three IROs with rotation; de novo and binding; deemed exhaustion for failure to strictly adhere, with the de minimis exception and 10-day written explanation; diagnosis and treatment codes with meanings on request, and that such a request is not itself an appeal), read at Cornell LII on August 13, 2026: https://www.law.cornell.edu/cfr/text/45/147.136
• eCFR was unreachable to automated access on August 13, 2026 (302 redirect to unblock.federalregister.gov), which is why the regulatory text above was read at Cornell LII: https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-B/part-147/section-147.136
• Healthcare Finance News, on the UnitedHealth nH Predict class action and the March 9, 2026 disclosure order (secondary; the 90% error rate is an allegation from the complaint, not a finding): https://www.healthcarefinancenews.com/news/class-action-lawsuit-against-unitedhealths-ai-claim-denials-advances
• CBS News, on the same litigation (secondary): https://www.cbsnews.com/news/unitedhealth-lawsuit-ai-deny-claims-medicare-advantage-health-insurance-denials/