I used Claude to negotiate $163,000 off a hospital bill. In a complex healthcare system, AI is giving patients power.

• Matt Rosenberg, a NY-based marketing consultant, asked Claude to analyze a $195,000 hospital bill.
• Claude helped him navigate complex billing codes and compare the charges to Medicare prices.
• He negotiated a $163,000 discount with the hospital and says AI is changing the power patients have.

The hospital bill arrived a month after the cremation. Eight lines of vague descriptions — cardiology, pharmacy, medical supplies — each with a dollar amount, many in five figures. The American health system's casual itemization of grief.

My 62-year-old brother-in-law had suffered a massive heart attack in June 2025 after running a 5K, barely an effort for a lifelong triathlete on most days, but this day, the trigger for a life-ending coronary.

Telling his wife he was having difficulty breathing, he was rushed first to urgent care and then to the ER at Community Memorial Hospital in Ventura, CA.

Six weeks later, the bill for the hospital's four hours of failed effort trying to save him: $195,628.

My sister-in-law's first instinct was to pay it. The life insurance would more than cover it. I told her to wait — to email me everything and sign nothing until we understood what we were looking at. The American healthcare system counts on people in grief making quick decisions. We weren't going to be those people.

My brother-in-law hadn't been uninsured by circumstance but by miscalculation — he'd let his policy lapse while shopping for something cheaper. When you regularly bike 10 miles before breakfast, re-upping health insurance feels less urgent than it should. The heart attack had come before he'd set up a new policy.

First, I had to learn a new language

The first lesson in medical billing is that hospitals speak in tongues. That initial bill was too vague — categories rather than procedures.

When I requested an itemized version, what arrived was a longer list of internal codes with prices — codes that existed only within the hospital's accounting system, opaque as an inside joke.

As a former ad agency guy, I've been trained in negotiation. The first task is to define the playing field, a reasonable starting point in response to their $195K gambit.

My strategy was to figure out what an insurer would've paid for identical services. Medicare is the country's largest healthcare payer, so I'd start there.

After two requests via phone, the hospital sent a UB-04 form — the same coded document they'd submit to Medicare or Blue Cross. I prepared to do what seemed straightforward: look up Medicare payments for each code, create a spreadsheet, and build our negotiating position on the radical proposition that the uninsured shouldn't pay more than the insured.

I was too lazy to Google it all, so…

There were a lot of codes. I took a shortcut and went to my AI assistant, Anthropic's Claude, which I typically use for research and to understand how AI intersects with our lives, for good or ill.

"Make a spreadsheet with these CPT codes and research what Medicare pays for each one," I typed. "Flag anything that needs further research."

Claude responded with questions: Which insurance type? Which geographic location? Which year? I'd thought Medicare rates were just Medicare rates. Lesson no. 2.

Their price was not right

Within a couple of minutes, Claude produced a spreadsheet. But it showed zeros for many of the codes instead of the dollar amounts I expected. In the notes column for these, it said, "See 92941RC, C-APC comprehensive payment."

Code 92941RC was a cardiac intervention priced at $30,767. I asked Claude to explain.

Claude responded with the flat precision of a tax attorney: Medicare doesn't pay à la carte for complex procedures. When doctors place a cardiac stent, Medicare makes one bundled payment covering everything — the procedure, the catheters, the guide wires, the contrast dye. It's called a Comprehensive Ambulatory Payment Classification, designed to prevent exactly the kind of bill I was looking at.

The hospital had unbundled the procedure. After charging $30,767 for the main intervention, they'd added separate lines for catheters (around $20,000), guide wires ($3,565), and medical supplies ($77,400) — over $100,000 for items Medicare would've paid nothing for because they're already included. (A gift with purchase!) It was as if a restaurant charged you for the pizza, then added separate charges for the dough, the sauce, and each pepperoni.

Claude found more ways that the bill my sister-in-law got differed from what Medicare would've received. The hospital had included a charge for a bypass, an in-patient-only procedure — you don't have a heart bypass in the morning and go home that same day. But my brother in-law hadn't had one; he'd never even made it past the ER/OR. They also billed for ventilation management, though Medicare forbids charging for ventilation when there is another critical care code.

Within an hour of back-and-forth conversation over details, Claude calculated Medicare would've paid approximately $28,675 instead of $195,628.

I knew I had to check Claude's work — so I used AI again

Before explaining our position to the hospital, I needed to verify Claude's analysis. Large language models are known to hallucinate. I've seen them make things up in service of giving me the answer it could tell would please me. I couldn't take the information at face value.

So I fed Claude's work to ChatGPT. "Check this for accuracy. Examine every detail. Flag any errors."

My logic: While each AI might hallucinate, two competing systems seemed unlikely to share the same delusion. And giving ChatGPT facts to check, I felt, would be less prone to sycophantic pattern-matching than asking it to run the same exercise — where there was clearly an outcome that would have pleased me.

ChatGPT confirmed the analysis. I then spent 20 minutes spot-checking on Google, reading the actual Medicare rules about bundling and inpatient charges. It was all there, but dense reading. I wouldn't have known what to look for without Claude.

Then it was time to negotiate

I drafted a six-page letter detailing each way the bill violated Medicare billing rules — the improper unbundling, the mutually exclusive service billing, the inpatient procedure on an outpatient claim. We offered to pay $28,675 promptly in exchange for a zero balance.

Within a week, the hospital countered at $36,356 without defending their initial billing. We split the difference at roughly $32,500. Three emails, a settlement agreement Claude helped me draft, and done. (I should note that as a former agency new business guy, I've spent a lot of time with lawyers and contracts and speak that language well — I don't recommend using AI as your lawyer.)

I learned to be an educated health care consumer

The collaboration with Claude and ChatGPT hadn't just saved $163,000. It had revealed the byzantine architecture of American medical billing — a system built on the assumption that patients won't understand what they're being charged.

Hospitals maintain "chargemaster" prices — fantasy numbers that bear the same relationship to reality as airport sandwich prices. No insurer pays those rates. They exist largely so insurers can claim massive negotiated discounts and so hospitals can report charity to maintain tax-exempt status. The only people who don't play this game are the uninsured. They get the unreal bill presented as real. Claude helped me call "not it."

The opacity isn't impenetrable; it's just tedious. The rules exist, published in federal registers, available to anyone willing to decode them. The system counts on people not having the time, energy, or knowledge to understand it.

The power difference between hospitals and patients has always been enormous

Hospitals have armies of billing specialists, coding experts, and lawyers. Patients have grief, fear, and confusion.

AI changes the equilibrium. Not by being infallible — it isn't — but by making the tedious manageable. What would have been days of regulatory research became a couple hours of conversation, collaboration, and brainstorming. Where I had questions or found errors, Claude informed or corrected. Together, our strengths and weaknesses produced a winning strategy.

Not everyone has the experience — but as these tools become more common, those who don't know how to use them will probably know someone who does.

My sister-in-law could've paid in full. She would've been among countless Americans who pay because fighting feels impossible, or go bankrupt because they can't pay and don't know they can fight. But she had someone to call, and that someone had Claude, and Claude could parse regulations hospitals hope patients are too intimidated to read.

The hospital that treated my brother-in-law wasn't uniquely unethical. They're caught in their own trap. But for the first time, the trap has a map — and the map is finally readable.

Matt Rosenberg is a marketing consultant for technology, travel, entertainment, and advertising companies. A former television writer, he lives in Dobbs Ferry, NY.

Do you have a story to share about using AI in your daily life? Contact this editor, Debbie Strong, at dstrong@businessinsider.com.