How one family cut a $195,628 hospital bill to $32,500 using AI
AI helped one family cut a $195,628 hospital bill to $32,500. Here's exactly how.
The American Hospital Association sent a bulletin to its member hospitals earlier this year warning them about a new threat to their billing operations: patients armed with AI chatbots. 1 Not lawyers. Not patient advocates. Patients — often grieving, often uninsured, often staring at bills that represent years of their salary — uploading their hospital statements to Claude and ChatGPT and asking those tools to explain what they're actually being charged for.
Two documented cases from early 2026 illustrate why the AHA is paying attention. In one, a marketing consultant named Matt Rosenberg used Claude to cut a $195,628 hospital bill to approximately $32,500 — an 83% reduction achieved through three emails and a $20/month subscription. 2 In another, a nonprofit worker named Walter Kerr used Claude to help his partner Jackie Davalos eliminate a $22,604 emergency room bill entirely. 1
Both cases were published in early 2026, outside this publication's standard weekly collection window. They're covered here because the methods they describe — and the limits those methods hit — are directly relevant to anyone holding a US hospital bill right now.
How a $195,628 bill became $32,500 — and what Claude actually did
Matt Rosenberg is a marketing consultant in Dobbs Ferry, New York. In June 2025, his 62-year-old brother-in-law finished a 5K run in Ventura, California and immediately went into cardiac arrest. 2 Community Memorial Hospital worked on him for roughly four hours in the emergency room and operating suite. He didn't survive.
About a month after the cremation, a bill arrived: $195,628. Eight line items. Each one five figures, each described in terms vague enough to be useless — "cardiology," "pharmacy," "medical supplies." The brother-in-law had been uninsured. He had let his old plan lapse while comparing cheaper options and never purchased the new one. 2
Rosenberg told his sister-in-law not to pay anything yet.
Getting the real bill
He called the hospital billing department and asked for an itemized statement — every charge by CPT code, service date, quantity, and unit price. The hospital initially sent only internal codes, not standard billing codes. Two more calls later, he received the UB-04 form, the structured billing document that lists each service as a standard five-digit CPT code. 2
That form was the key. Without it, the bill was unassailable. With it, every charge became a specific, publicly defined medical service — and every publicly defined medical service has a published Medicare reimbursement rate.
What Rosenberg asked Claude
He uploaded the CPT codes and gave Claude this prompt: 2
"Make a spreadsheet with these CPT codes and research what Medicare pays for each one. Flag anything that needs further research."
Claude returned a breakdown. What it found in the Rosenberg bill was not a collection of prices that were merely high. Several items violated Medicare billing rules in ways that should have disqualified them from being billed separately at all: 2
- The hospital billed $30,767 for the main cardiac intervention (code 92941RC under Medicare's Comprehensive Ambulatory Payment Classification). Under C-APC rules, that code bundles in catheters, guide wires, and related supplies. The hospital then separately charged roughly $20,000 for catheters, $3,565 for guide wires, and $77,400 for "medical supplies" — over $100,000 in items that Medicare would not pay separately because they're already included in the procedure's bundled rate.
- The hospital charged for a coronary bypass procedure. That's an inpatient-only surgery. The patient never left the ER. The bypass was never performed.
- The hospital charged for ventilation management. Medicare prohibits separate billing for ventilation management when another critical care code is already present on the same claim.
"It was as if a restaurant charged you for the pizza, then added separate charges for the dough, the sauce, and each pepperoni," Rosenberg later wrote. 3
Verifying the analysis
Rosenberg did not take Claude's output at face value. He ran the same analysis through ChatGPT — asking it explicitly to check Claude's work for errors — and then spent roughly 20 minutes spot-checking the most significant findings against actual CMS documentation online. All three paths confirmed the same violations. 2
Claude estimated Medicare would have paid approximately $28,675 for the same services. Rosenberg used that as his settlement anchor. 2
The letter and the outcome
He drafted a six-page letter — with Claude's help structuring the argument — that detailed each billing violation and its specific Medicare rule, then offered to settle the entire balance for $28,675. The logic was explicit: the hospital's billed amounts contradicted the billing rules that govern Medicare reimbursement, and the offer represented the amount Medicare would have considered appropriate. 2
The hospital responded within a week. Counter-offer: $36,356. Three emails later, both parties settled at approximately $32,500. 2
Total saved: roughly $163,000. Total tool cost: $20.
"Hospitals have armies of billing specialists, coding experts, and lawyers," Rosenberg wrote. "Patients have grief, fear, and confusion." 2 AI made a task that used to require professional expertise — decoding CPT codes against Medicare rules — something one person could do in a few hours at a kitchen table.
コンテンツカードを読み込んでいます…
How to run this play on your own bill
The Rosenberg method has five steps. Steps 1 and 2 require only patience. Steps 3–5 require an AI subscription (Claude or ChatGPT, both offer paid tiers starting at $20/month). The whole process, for a bill with meaningful violations, should take one to two days of focused work.
Step 1: Request the itemized bill with CPT codes. Call the hospital billing department and ask specifically for an itemized statement listing every charge by CPT or HCPCS code, service date, quantity, and unit price. Hospitals are required to provide this on request. If they initially send a summary or internal codes only, call back and ask again for the UB-04 form (for hospital services) or the CMS-1500 form (for physician services). You may need two or three calls; this is normal and expected.
Step 2: Do not pay anything until you have this document. Paying the initial invoice — even partially — signals acceptance of those charges. Hold the bill.
Step 3: Upload the CPT codes to an AI and ask for the Medicare rate comparison. Use Rosenberg's exact prompt as a starting point: "Make a spreadsheet with these CPT codes and research what Medicare pays for each one. Flag anything that needs further research." The AI will look up Medicare's published reimbursement rates for each code and note where your charges deviate significantly. Look specifically for: unbundled charges (components billed separately when the main procedure code bundles them), procedures listed that match your medical records but weren't actually performed, and codes that Medicare rules prohibit billing in combination.
Step 4: Verify independently. Run the AI's output through a second AI tool with a check prompt: "Review this Medicare billing analysis for errors. Examine every detail. Flag any inaccuracies." Then spot-check the two or three most significant findings against the CMS Physician Fee Schedule (available free at cms.gov) or against the Medicare Coverage Database. This takes 20–30 minutes and is not optional — AI billing analysis can contain errors, and you need to be confident in the claims you're putting in writing.
Step 5: Write a settlement offer anchored to Medicare rates. Draft a letter — AI can help structure it — that does three things: (1) names each specific billing violation with the relevant CPT code and Medicare rule, (2) states what Medicare would have paid for each service, and (3) offers to settle the full balance at or near the Medicare rate (or a modest multiple, such as 150%–200%, if the Medicare rate alone seems too low to be accepted). Send it to the hospital's Patient Financial Services department by email or certified mail. Request a written response within 30 days.
A note on what this method is suited for: it works best on large bills from uninsured or underinsured patients, where the hospital used the chargemaster rate — the full list price — as the starting point. If you have insurance and a relatively small remaining balance after your insurer has paid, the billing error dispute route (documented in this publication's May 19 article) is usually the faster path.
What AI gets wrong: the Kerr case
Walter Kerr's story arrived as the more dramatic headline — $22,604 ER bill, completely eliminated — but it also contains the most important warning about relying on AI for medical bill disputes.
In mid-2025, Kerr's partner Jackie Davalos, a former Bloomberg reporter then training as a pastry chef, received a collections notice for $22,604 from George Washington University Hospital. The bill traced back to a fall she'd had two years earlier. The hospital had recorded her as uninsured at admission — which the hospital later acknowledged was a clerical error. She had been insured. 1
Kerr uploaded her billing and medical records to Claude. The AI suggested the hospital may have failed to meet legal requirements around debt and insurance. Kerr used Claude's arguments to draft letters to executives at GWU Hospital and its parent company, Universal Health Services. The hospital waived the bill. 1 Hospital spokesperson Susan LaRosa confirmed a clerical error had occurred at admission. Universal Health Services spokesperson Maria English said the debt was resolved "after receiving all information and completing communication with the patient." 1
The result was 100%. But the method had a flaw that only surfaced after an independent legal review.
Ariel Levinson-Waldman, founding chair of Tzedek DC (a Washington D.C. legal aid organization focused on consumer debt and financial exploitation), reviewed Kerr's dispute documentation. She found that some of Claude's legal analysis was correct — and some was not. The chatbot had cited debt and insurance laws that actually govern insurers or third-party debt collectors, not hospitals directly. It also failed to inform Kerr and Davalos of other avenues they could have pursued. 1
"Getting useful answers from the chatbot usually requires knowing how to give it the right instructions — or knowing enough about health insurance to provide the right context," said Andrew Cohen, a staff attorney at Health Law Advocates (a Massachusetts nonprofit providing free legal representation on healthcare access issues), also quoted in the New York Times report on these cases. 1
Kerr's own framing of his experience captures the limits precisely. He described Claude as "a useful adviser, but not a perfect one." And his summary of what AI cannot replace: success "often requires persistence, something AI can't solve for you." 1
The practical takeaway from both cases: AI is reliable at matching CPT codes to Medicare rates, identifying unbundled charges, and helping draft clear letters. It is less reliable at interpreting which specific laws apply to a hospital's billing conduct versus a collector's. When the dispute involves a billing error with a verifiable code-by-code analysis (Rosenberg's method), AI's strengths align well with the task. When it involves broader legal arguments about what the hospital was required to do, the analysis needs independent legal verification before you rely on it.
First 3 moves: what to do within 72 hours of any US medical bill
These steps apply to any bill — whether or not you plan to use AI, regardless of your insurance status or the bill amount. Do them before paying a single dollar.
- Request the itemized bill. Call the billing department and ask for every charge listed by CPT/HCPCS code, service date, quantity, and unit price. This is your legal right. Billing errors — duplicate charges, procedures billed but not performed, unbundled components — are common and only visible at the code level.
- Do not pay the first invoice. The initial statement reflects chargemaster rates — list prices that bear no fixed relationship to Medicare rates, insurer-negotiated rates, or any other market benchmark. Paying it signals acceptance. Wait until you have the itemized bill and have reviewed it.
- Check financial assistance eligibility before assuming you owe anything. Every nonprofit hospital in the United States — the majority of US hospitals — is required under IRS 501(c)(3) rules to offer charity care or financial assistance. 4 Most patients are never told this exists. The nonprofit Dollar For helps uninsured and underinsured patients apply for hospital charity care at no cost. 4 For patients who qualify, financial assistance can eliminate the bill entirely without any negotiation — before the AI playbook, before the letters, before any of it.
Which path fits your bill?
| Situation | Recommended path |
|---|---|
| Large bill, uninsured/underinsured, from a hospital stay or ER | Run the Rosenberg method: itemized bill → CPT code analysis → Medicare anchor offer |
| Billing error visible (wrong network, duplicate charge, wrong code) | Formal written dispute → 30-day warning letter with executive-level recipients (see May 19 playbook) |
| Uninsured with a nonprofit hospital | Dollar For charity care application first, before anything else |
| Active collections notice, bill 2+ years old | Check your state's statute of limitations on medical debt before paying; dispute the debt in writing within 30 days of first contact |
Rosenberg put it directly: "I wouldn't have known what to look for without Claude. But I didn't take the information at face value." 5 AI as the research layer, human judgment as the verification layer — both cases run on that combination.
Cover image: AI-generated illustration.
このコンテンツについて、さらに観点や背景を補足しましょう。