You’ve probably seen the story. A grieving family fed a $195,000 hospital bill into Claude, and the AI flagged duplicate charges, coding that billed an ER visit as a costlier inpatient stay, and supplies marked up hundreds to thousands of percent over what Medicare pays. It drafted an appeal. The bill came down to about $33,000.
That’s a real case. It’s also not what’s going to happen to you.
Here’s the honest version, the one the viral screenshots skip. Most bills won’t fall 83%. Plenty of bills are simply correct, and the AI will still find “problems” in them — because a chatbot will confidently invent a billing rule that doesn’t exist, misread your insurance plan, and call something a “violation” when it isn’t. In April 2026 the New York Times ran the whole picture under the perfect headline: patients are using chatbots to fight medical bills, with mixed results. AI genuinely helps people spot errors and write letters. It also makes serious mistakes about billing rules and insurance law, in the same calm, sure-of-itself voice.
So the trick isn’t “let AI fight your bill.” It’s narrower and safer than that: use AI to understand, check, and draft — never to conclude. Every step below is built around that one line.
Why this actually works right now
Because the bills really are a mess. Reviews of medical billing keep landing in the same range: somewhere between 75% and 80% of bills contain at least one error. One 2022 study put it at 79%. The average hospital bill over $10,000 carries roughly $1,300 in mistakes. Wrong quantities, duplicate line items, a service you never got, an in-network procedure billed out-of-network.
And when people push back, it works more often than not. Around 74% of patients who disputed an error got it corrected. About 62% who negotiated got the amount reduced. But here’s the number that explains why most people overpay: only about 61.5% ever contact the billing office at all. The bill arrives, it’s a wall of codes, it’s exhausting, and it gets paid.
That wall of codes is the exact thing AI is good at. Those CPT and HCPCS codes — five-digit strings like 99285 or J1200 — have plain-English meanings, and a chatbot can translate a whole itemized bill in seconds. Not judge it. Translate it. That alone turns “I have no idea what I’m looking at” into “wait, why am I being charged for that twice?”
The 4-step method
Step 1: Get the itemized bill (not the summary)
The thing that shows up in the mail is usually a summary — “Hospital services: $8,400.” You can’t dispute that. You need the itemized bill, the one that lists every charge with its code. Call the billing number and ask for “a fully itemized bill with CPT and HCPCS codes.” You’re entitled to it, and asking for it is normal — billing departments field that request constantly.
While you’re at it, dig out your Explanation of Benefits from your insurer. That’s the document that shows what was billed, what your plan allowed, and what it actually paid. It’s your reference sheet for Step 3, and it’s confusing enough that it deserves its own decoder.
Step 2: Strip your identity, then have AI decode each line
This is the step people skip, and it’s the one that matters most. A consumer chatbot is not your hospital and not your insurer. Nothing you paste into it is protected by medical-privacy law — assume it can be stored and used. So before a single character goes into that text box, take the person out and leave the numbers in.
Before you paste anything, remove all of this:
- Your name and any family member’s name
- Your member / subscriber ID and group number
- Your date of birth
- Your address
- Your account number and claim number
- Anything else that points to a specific human being
Give the AI the charges and the codes. Not the patient. Here’s a prompt that bakes the redaction reminder and the safety leash right in:
Act as a plain-language medical-billing explainer — not a lawyer and not
a certified medical coder. I've already removed my name, member ID, date
of birth, address, and claim number from the text below. Don't ask for them.
Here is an itemized bill (codes and charges only):
[paste the line items — no names, no IDs]
Please:
1. Translate each line into plain English — what was this charge for?
2. Flag any line that looks like a possible duplicate, a likely quantity
error, or a charge that seems unusually high for what it is.
3. For each flag, tell me exactly what to check against my Explanation of
Benefits and plan documents to confirm it — don't assume it's wrong.
Do not tell me what I owe. Do not tell me a rule was "violated." Do not
tell me to pay or not pay. Give me things to verify, not verdicts.
Notice what that prompt refuses to let the AI do. It can’t declare a violation. It can’t hand you a number. It turns every flag into a question you go check — which is exactly the handoff to the next step.
Step 3: Verify every flag against your EOB and plan — this is the safety valve
The AI just gave you a list of “problems.” Treat every one of them as a maybe, not a fact.
This is where the confident-tone trap lives. The chatbot will say something like “line 7 is a duplicate charge and a clear billing violation” with total conviction — and it might be a duplicate, or line 7 and line 12 might be two legitimately separate things that just read alike. The only way to know is to match each flag against your own paperwork: the itemized bill, the EOB, and your plan’s summary of benefits. Does the EOB show that charge was already adjusted or paid? Does your plan actually cover it the way the AI assumed? Was that “duplicate” really billed twice, or did you get the same test on two different days?
A chatbot’s certainty is not evidence. Your EOB is. If you can’t confirm a flag against a real document, it doesn’t go in the letter.
Step 4: Have AI draft the dispute letter — then you read, edit, and send
Once you’ve got two or three flags you’ve actually verified, this is where AI earns its keep again. Drafting a clear, calm, specific dispute letter is genuinely hard when you’re stressed and angry, and it’s the exact kind of writing a chatbot does well.
Help me draft a short, professional letter disputing specific charges on
a medical bill. Keep it factual and calm. For each item I list, state the
charge, why I'm questioning it, and request a corrected itemized bill.
Don't cite any law, regulation, or dollar figure I haven't given you.
The charges I've verified and want to dispute:
[list only the flags you confirmed against your EOB in Step 3 —
charges and codes, still no names or IDs]
Then read every line before it goes out. Cut anything that overreaches. Put your own name back on it — this is your letter, going out under your name, making claims you can stand behind. The AI wrote a draft. You’re the one sending it.
What this means for you
If you’re staring at a big ER bill: the itemized version is your best friend. ER coding errors — like the inpatient-vs-ER mix-up in that viral case — are common and expensive, and they’re exactly the kind of thing decoding the codes surfaces. Run the four steps. But go in expecting to find one or two real errors, not to watch $195K melt to $33K.
If your bill’s already in collections: don’t panic-pay to make it stop, and don’t ignore it either. You can still dispute a bill in collections, and an itemized bill you never received may reset the clock on what’s actually owed. This is also the point where a human belongs in the loop — more on that below.
If you’re helping a parent: you’ll probably be the one pasting things in, which makes the redaction step your job. Strip their name, their Medicare or member ID, their date of birth — everything — before it touches a chatbot. Decode the bill for them, verify the flags, and bring the questions to the provider together.
If the bill turns out to be correct: this happens, and it’s a real outcome, not a failure. Sometimes the codes check out, the EOB matches, and the number is just genuinely large. AI that’s honest will tell you a charge looks standard. If every line comes back “fine,” believe that as readily as you’d believe a flag — the goal was an accurate picture, not a discount.
The 3 things AI gets dangerously wrong
This is the “mixed results” half of the story, and it’s not optional reading. The New York Times found AI helping people and misleading them in the same breath. Here’s where it misleads.
1. It invents billing rules. Ask a chatbot whether a charge is “allowed” and it will often answer with a confident, specific-sounding rule — a Medicare guideline, a billing regulation, a “standard practice” — that is partly or entirely made up. It’s not reading your insurer’s actual policy. It’s pattern-matching what a rule might sound like. Never quote an AI’s rule in a dispute letter. If a real rule matters, it has to come from your plan documents or a person who knows them.
2. It hallucinates policy language and “violations.” It will tell you your plan covers something, or that a charge broke the law, based on assumptions about a policy it has never seen. Your actual coverage lives in your specific plan’s documents — deductible, coinsurance, in-network status, all of it. The AI is guessing at yours from the average of everyone’s. Sometimes the guess is close. Sometimes it invents a “violation” out of thin air, and if you send that, you’ve weakened a letter that might otherwise have worked.
3. It’s confidently wrong. This is the one that gets people. The tone never wavers. A flat-out fabrication arrives in the same steady, reassuring voice as a solid catch. There’s no flicker of doubt to warn you. That’s why Step 3 isn’t a suggestion — the verify-against-your-own-documents step is the only thing standing between “the AI sounded sure” and “I sent a letter full of errors.”
A chatbot is not your lawyer, not your medical coder, and not your patient advocate. It doesn’t carry a license, and when it’s wrong, no one’s accountable but you.
When the stakes are real, bring in a human
The four-step method is for decoding, checking, and drafting. It is not for high-stakes decisions on its own. When there’s a large balance you can’t verify, a denial you’re appealing, or a bill that’s gone to collections, put a real person in the loop:
- A patient advocate or medical-billing advocate, who does this professionally and knows the rules the AI only pretends to.
- Your hospital’s financial-assistance / charity-care office — nonprofit hospitals are required to have one, and you may qualify to have some or all of the bill reduced.
- Your state’s Consumer Assistance Program or insurance department, which can step in on denials and disputes at no cost.
One note on the difference between sides of this. If you do medical billing for a living — appealing denials on behalf of a provider, working with payers all day — that’s a different job with different rules and real PHI obligations, and it has its own playbook. This post is for the patient on the receiving end of the bill.
The bottom line
The viral $33,000 story is true, and it’s the exception. The everyday win is smaller and more reliable: you understand a bill you couldn’t read, you catch a duplicate or a coding error you verified against your own EOB, and you send a calm letter that gets it corrected. AI does the decoding and the drafting. You do the verifying and the deciding. Strip your name before you paste, treat every flag as a maybe, and never let a confident chatbot talk you into a claim your own paperwork can’t back up.
The underlying skill — asking AI precise questions and auditing its answers instead of trusting them — is the one that pays off everywhere. Understand Your Health Results With AI has a whole lesson on doing exactly this with bills, EOBs, and discharge notes safely. AI Fundamentals builds the habit from zero, and Prompt Engineering sharpens the role-and-leash prompting these steps run on.
And bills aren’t the only confusing paperwork AI can help you read. The same careful approach works on the rest of the pile: start with your Explanation of Benefits, decoded, and if it’s a lab report worrying you, how to read a blood test with AI — down to what RDW actually means.
Sources
- Patients Are Using Chatbots to Fight Medical Bills, With Mixed Results — The New York Times (April 8, 2026)
- Grieving family uses AI chatbot to cut a hospital bill from $195,000 to $33,000 — Tom’s Hardware (2026)
- Up to 80% of medical bills have errors — AI can spot them in minutes — Komando (2026)
- How to read an explanation of benefits — CMS.gov (Medical Bill Rights)
- Know your medical bill rights and the No Surprises Act — CMS.gov
