Research · Navigation and cost

The only market with no price.

Healthcare is the one large market in America where you cannot learn the price before you buy, cannot predict what you will need, and cannot read the contract you signed. Washington ordered every price published. The prices arrived: terabyte files where as many as nine entries in ten are rates nobody will ever pay, policed by 27 fines in four years. This paper is about what the transparency rules actually produced.
Keel LabsMay 19, 2026No. 0215 min read
Figure 1 · One operation, every hospitalpause
Each dot is a hospital's negotiated commercial rate for the same operation: coronary bypass without major complications (MS-DRG 236), plotted on a log price scale. Rates run $27,683 to $247,902 around a median of $68,194. Trilliant Health analysis of insurers' Transparency in Coverage files, Aug 2025.

Imagine buying a car with no sticker price, where the cost depends on which dealer you walked into, who arranged the loan, and how much of your budget you already spent this year, and you cannot find out any of it until weeks after you drive away.

That is healthcare, and I do not think most people inside this industry register how strange it is anymore. It is the only large market where the buyer routinely cannot learn the price before the purchase, cannot predict what they will need, and cannot understand the contract they signed. Price, need, and language. Three blindnesses, and this paper argues they are one problem. It also tells a newer story, the one the outrage pieces skip: what happened when the federal government ordered every price published, and the prices arrived in a form no human being can read.

The fair objection should come first, because healthcare was never going to work like a car lot. You cannot comparison-shop from the back of an ambulance. A third party pays most of the bill, so the sticker was always going to matter less here than elsewhere. Quality is hard to judge, which makes a low price read as a warning instead of a deal. All of that is true, and this paper takes it seriously. None of it explains Tufts Medical Center, where the same coronary bypass carries a negotiated price of $95,989 when Aetna did the negotiating and $144,204 when UnitedHealthcare did. Same operating room, same surgeons, same diagnosis code, $48,215 apart. Emergencies do not produce that number. Negotiation does.

No price

A market where you cannot see the price

The clearest statement of the problem is the title of a federal report: meaningful price information is difficult for consumers to obtain prior to receiving care. When the Government Accountability Office tested that proposition, its investigators called 20 physician offices and asked what a routine diabetes screening would cost. Several offices said the patient would need to be seen by a physician before anyone could say which tests they would even need. Of the 18 offices that responded, 4 could produce a complete estimate. The hospitals GAO called about knee replacements quoted facility charges with the physician and lab fees missing, because those bill separately. No other consumer market asks the buyer to solve that equation blind.

For years you could wave the variation away as an artifact of fake numbers, since hospital "charges" were fictions nobody actually paid. That excuse died in August 2025, when Trilliant Health analyzed the real negotiated rates insurers are now required to publish, covering 2,659 hospitals and 3,491 surgery centers. Across six common inpatient procedures, negotiated prices varied 9.1-fold on average across the country. A coronary bypass without major complications ran from $27,683 to $247,902 against a median of $68,194. The spread on one heart operation, $220,219, is the price of a house. And the strangest finding points away from the usual villain: the $48,215 gap at Tufts sits inside a single building, for a single procedure code. The only variable left is who sat across the table from the hospital's contracting office.

Spread is half the picture. The level is the other half. RAND's employer-led transparency study, the only public accounting of what employer plans actually pay hospitals, found that in 2022 those plans paid 254 percent of what Medicare would have paid for the same services at the same facilities, and 279 percent for hospital outpatient care. The state averages run from under 170 percent in Arkansas to over 300 percent in eight states. Note what RAND had to do to learn this: recruit employers willing to contribute their own claims data, employer by employer, because nobody financing the system could simply look the prices up.

254%
What employer health plans paid hospitals in 2022, relative to Medicare, for the same services at the same facilities. Outpatient care ran 279 percent. RAND round 5.1.
The price of care depends on the service, the provider, the insurer, and how much of your deductible you have already met. You are asked to solve that with none of the numbers.
The library

The prices were published. The files cannot be read.

On paper, this is solved. Since January 2021, every hospital must post its negotiated rates. Since July 2022, every insurer must publish machine-readable files listing each price it has negotiated with each provider, for every plan it sells. The disclosure happened. What arrived is the most honest measurement in this paper.

A single carrier's monthly file can reach one terabyte, the entire drive of a decent laptop holding one company's price list. When the Health Care Cost Institute pulled the files for eight payers in one state, Pennsylvania, it received more than 11,000 raw files containing 30 billion rows, 1.6 terabytes. Trilliant Health downloaded more than 20 terabytes to study just two national carriers. Nobody prints these. Nobody scrolls them. The data is public the way the bottom of the ocean is public.

Worse than the size is what the rows say. Analysts who parse the raw files find that up to 90 percent of entries are "zombie rates": negotiated prices for clinician-and-service pairings that will never occur in practice. A cardiologist with a contracted price for a spinal adjustment. The rows exist because carriers cross-list entire fee schedules against every provider under contract, which satisfies the regulation and says nothing about reality. Each row is technically compliant. Almost none of it is information. The government now uses the zombie vocabulary itself: the December 2025 proposed amendments to the insurer rules would, for the first time, order plans to delete rates a provider "would be unlikely to be reimbursed" for. The regulator is proposing to make the disclosure mean something five years after mandating it.

Figure 2 · Reading one payer filepause
A scan moves through the rows of a raw negotiated-rate file. Watch the counters: most entries price services the listed clinician would never bill. Audits of raw Transparency in Coverage files find implausible records can reach 90% of a file. HCCI; CRS R48570.

Even the compliance rate is unreadable. PatientRightsAdvocate.org, the strictest auditor, scored 21.1 percent of hospitals fully compliant in its November 2024 report, down from 36 percent earlier that year. Turquoise Health, reading the same postings with different criteria, finds roughly half of the files it can parse carry the data needed for full compliance. CMS itself has pointed to outside estimates running from about 5 percent to nearly 85 percent depending on methodology. Sit with that: four years into a federal transparency rule, the public cannot learn the price of care, and cannot even learn what fraction of hospitals is disclosing it correctly. The methodology war is not a footnote to this story. It is the diagnosis: a disclosure regime specified so loosely that compliance itself became a matter of interpretation.

90%
The share of entries in a raw payer rate file that can be implausible "zombie rates," per audits of the Transparency in Coverage files. The disclosure is measured in terabytes. The information is not.
The vacuum

Eighty-five warnings for every fine

Enforcement was supposed to be the backstop, so here is the complete federal ledger for the hospital rule, from 2022 through mid-2025: 2,291 warning notices, 1,331 corrective-action requests, and 27 fines. One fine for roughly every 85 warnings. Of the 27 hospitals fined, 21 appealed the penalty.

The GAO explained why the ledger looks like this. In an October 2024 audit, CMS officials told the GAO that the agency does not routinely review whether the prices hospitals post are complete or accurate; it checks accuracy only in response to "credible complaints alleging potentially egregious issues." Follow the loop: a complaint requires a reader, the files are built so that no human can read them, so the complaints do not come. The enforcement regime assumes the existence of the exact audience the data format excludes. HHS agreed with the GAO's recommendation to fix this, which is what agencies say when a fix has no date.

Figure 3 · The enforcement ledgerpause
Every federal enforcement action under the hospital price transparency rule, Jan 2022 through Jun 2025, drawn to scale. The bottom bar is the fines. CMS enforcement data, tallied in Health Affairs Forefront, Oct 2025.

The rules are now being rewritten, and the direction is right even where the pace is not. A February 2025 executive order sent the agencies back to work. The hospital rule finalized that November requires, from January 2026, that hospitals stop posting algorithm-based estimates and instead encode the actual median, 10th, and 90th percentile amounts computed from their own remittance data, with a named senior executive attesting inside the file that the data is true, accurate, and complete. The same rule contains a tell: a hospital that waives its right to a hearing gets its fine cut by 35 percent. An agency confident in its penalties does not discount them for going uncontested. Enforcement of the new requirements begins in April 2026, and the proposed insurer amendments, zombie purge included, would not bind until plan years starting in 2027. The fix is real, scheduled, and unproven.

A complaint requires a reader. The files are built so no human can read them. The complaints do not come.
Nobody shops

Almost no one shops, even when they want to

So people do not shop. Part of that is the concession from the opening: a real share of spending is emergencies and complex hospital care nobody can shop for. But the shoppable part, the scans and labs and planned procedures, goes unshopped too. In a nationally representative survey of 2,996 adults, only 13 percent of people with out-of-pocket spending at their last encounter had sought any cost information before care, and only 3 percent had compared prices across providers. Three in a hundred. The same survey kills the apathy explanation: most respondents said price shopping matters and did not believe the expensive providers were better.

Massachusetts makes the diagnosis cleaner. Seven in ten consumers there said they want prices before care. Only 31 percent knew their insurer runs a website that estimates those prices, and only one in five had ever tried to look a price up anywhere. The appetite exists. The access does not.

The deepest evidence that this is blindness rather than indifference comes from the one natural experiment large enough to settle it. When a big firm moved more than 150,000 employees from free care into a high-deductible plan, spending fell 11.8 to 13.8 percent, and the researchers could attribute none of the drop, zero, to people finding cheaper providers. Two years in, nobody had learned to shop. People simply consumed less of everything, valuable care included, and the sickest employees, who were certain to blow through their deductible anyway, cut back in response to the price in front of them that day. Handing a blind buyer a bigger share of the bill does not teach them to see. It teaches them to flinch.

Figure 4 · The shopping funnelpause
Out of everyone who pays out of pocket, a sliver ever looks up a price, and almost no one compares two. Mehrotra et al., Health Affairs 2017.
3%
of people with out-of-pocket spending ever compared prices across providers before getting care. Only 13 percent looked up a single price at all.
The cost of blindness

Cost-blindness gets paid in two currencies

The first currency is money: 41 percent of US adults, roughly 100 million people, carry debt from medical or dental bills, more people than live in California, Texas, and Florida combined. The second is behavior. Adults carrying that debt are more than twice as likely to postpone or skip care they need, and about half of them say cost stopped a test or treatment their doctor recommended within the past year. People learn to ration their own care out of fear of a bill they were never allowed to see. The fear is doing exactly what an unpriced market trains it to do.

100M
Americans carry medical or dental debt. About half of adults with that debt say cost stopped a doctor-recommended test or treatment in the past year.
The language

The contract was written in a language that was never theirs

The third blindness is the contract itself. When KFF quizzed Americans on ten basic insurance concepts, 4 percent answered all ten correctly, and coinsurance, the term modern plan design leans on hardest, was the least understood on the test: about a third could define it. In a separate academic survey, 14 percent could correctly define deductible, copay, coinsurance, and out-of-pocket maximum together, and 11 percent could compute the cost of a four-day hospital stay from a plan description. This vocabulary gap has a price tag. Bhargava, Loewenstein, and Sydnor watched 23,894 employees at one large firm choose among plans that differed only in cost-sharing, and 61 percent picked a plan that was worse at every possible level of spending. The average overspend was about 24 percent of the chosen premium, roughly $372 a year, handed over for misreading a contract. Workers earning under $35,000 made the error at more than twice the rate of those earning over $100,000.

Figure 5 · Choosing the worse planpause
Two plans, identical except for price. People reliably pick the more expensive one, and pay about a quarter of their premium for the privilege of not understanding it. Bhargava, Loewenstein & Sydnor, QJE 2017.

None of this is carelessness. These are people making four-figure decisions in a language nobody taught them, holding a generic benefits guide organized around plans rather than around a life. The behavior follows: in Voya's enrollment research, 91 percent of employees re-select last year's plan, and about half spend under twenty minutes on the choice that sets their family's financial exposure for the year. A market with no visible prices, no readable disclosures, and no comprehensible contracts produces exactly this person: someone who clicks "same as last year" and hopes.

What would change our mind

How this argument could die

A claim this strong should carry its own failure conditions, so here are ours. The proposed insurer amendments published in December 2025 would purge the zombie rates and standardize the files; if that rule is finalized and negotiated prices then begin converging, the no-price thesis weakens with every point of convergence, and we will report it. The hospital attestation rule took effect in January 2026 with enforcement from April; if CMS shifts from warnings to fines at scale and the hospital files become trustworthy, the bottleneck moves elsewhere and this paper needs a successor. And one question the literature cannot yet answer: whether a person who finally sees a real, personal price acts on it. The best evidence says yes when the incentive is attached, since a national insurer's reference-pricing program for imaging cut spending on those scans 8.1 percent, almost entirely by changing where people went. But the shopping studies above tested people handed clumsy tools they had no reason to trust. Nobody has run the experiment with a tool that answers for their plan, their deductible, their town, in their own words. That is the experiment we intend to run.

What Keel Labs is building

The cost problem and the fear problem are the same blindness. One system, holding three facts, closes both.

Transparency failed as a policy because it assumed a human reader, and a terabyte of zombie rates is not addressed to one. It is, however, exactly the kind of thing software reads. The raw material now exists in public: every negotiated rate, every plan document, every fee schedule, buried in formats that defeat people and do not defeat machines. What has never existed is a system that holds one person's plan, their likely needs, and the real price of a service in the same place, and explains the result in words that person understands. This is how we build it.

01

Resolve the plan they actually hold

We read the specific coverage a person elected off the benefits graph, deductible met and all, so every answer starts from their real situation and not a generic plan summary.

02

Estimate the need before it arrives

Individual utilization is hard to predict, so we reason about likely needs from a real life, family, history, and risk, instead of asking someone to forecast a year of healthcare they cannot see coming.

03

Attach the real price, and admit when there isn't one

We pull the actual negotiated rate for a service at a specific site, filtered against the zombie rows, and when the data is missing or implausible the system says so. An honest "we cannot price this yet" beats a confident wrong number every time.

04

Fathom holds all three at once

Plan, likely need, and real price have never sat in the same place for the same person. Our model puts them there, and grounds every answer back to the source, so it can be trusted.

05

Amanda says it in plain language

Confidence follows understanding. A calm voice that explains coverage anonymously, without judgment, at the moment of the decision, is how the fear comes down at the same time as the cost.

The mechanism is almost boring once you see it. You do not lower the cost of care by squeezing a vendor. You lower it by letting the buyer finally see what they are buying, and the evidence says the money is there: an 8.1 percent drop in imaging spending when people could act on price, and the 24 percent of premium recovered when someone stops choosing the dominated plan. Price, need, and language, held together for one person at the moment they decide.

An AI that connects the plan you hold to the care you are likely to need to what it will actually cost, in words you understand, is the first time those three facts have ever been in the same room. That is the loop we are closing.
What this paper does not claimWe are not claiming AI sets prices or repairs the transparency rules, only that it can make the published information usable for one person at one moment. The headline spreads, the 9.1-fold variation and the $48,215 same-hospital gap, come from Trilliant Health, a single analytics firm reading the federal files, and a firm that sells price-transparency analytics has an obvious interest in those files mattering; treat its figures as point-in-time. "Up to 90 percent zombie rates" describes raw machine-readable files as characterized by analyst audits, not a regulator's count, and the share varies by carrier and file. Hospital compliance is genuinely indeterminate, from roughly a fifth to most hospitals depending on whose methodology you trust, and we present that fight rather than resolve it. The enforcement tallies are a Health Affairs Forefront count of CMS's posted actions through June 2025. RAND's percent-of-Medicare figures rest on claims contributed voluntarily by employers and states, so coverage varies by geography. The deductible natural experiment is one large firm. The reference-pricing result is one program, for one category of imaging, at one insurer. The enrollment-behavior figures are a single vendor's surveys, cited as such. The 100 million medical-debt figure is a survey-based national estimate. And the claim that a person who finally sees a real, personal price will act on it remains, today, the hypothesis we are testing rather than a fact we can cite. The load-bearing numbers trace to government, KFF, peer-reviewed, and named-firm sources below.
SourcesGAO-11-791, Health Care Price Transparency: Meaningful Price Information Is Difficult for Consumers to Obtain Prior to Receiving Care (2011; 20 physician offices called on diabetes screening, 4 of 18 respondents gave complete estimates) · Trilliant Health, 2025 Health Plan Price Transparency Report (Aug 2025; TiC negotiated rates, 2,659 hospitals and 3,491 ASCs; six procedures vary 9.1x on average; MS-DRG 236 range $27,683 to $247,902, median $68,194; Tufts Medical Center MS-DRG 235: Aetna $95,989 vs UnitedHealthcare $144,204; 20+ TB downloaded for two carriers) · RAND, Prices Paid to Hospitals by Private Health Plans, Round 5.1 (2024; 2022 claims; 254% of Medicare overall, 279% outpatient; Arkansas under 170%, eight states over 300%) · Health Affairs Forefront, Reforming Hospital Price Transparency Enforcement (Oct 2025; CMS actions 2022 to Jun 2025: 2,291 warnings, 1,331 corrective-action requests, 27 penalties, 21 appealed) · GAO-25-106995 (Oct 2024; CMS does not routinely review completeness or accuracy; complaint-driven review) · CMS, CY 2026 OPPS Final Rule, Hospital Price Transparency fact sheet (Nov 21, 2025; allowed-amount encoding, executive attestation, 35% penalty reduction for waived hearings, enforcement from Apr 2026) · Tri-agency Transparency in Coverage proposed rule, Federal Register (Dec 23, 2025; zombie-rate exclusion, taxonomy file; plan years 2027 if finalized) · Executive Order 14221 (Feb 2025) · CRS R48570 (payer files up to ~1 TB) · Health Care Cost Institute, Leveraging TiC Data (eight Pennsylvania payers: 11,000+ files, 30 billion rows, 1.6 TB; implausible records up to 90% of a raw file) · PatientRightsAdvocate.org semi-annual compliance reports (Nov 2024: 21.1%; earlier 2024: 36%) · Turquoise Health price transparency impact reports · CMS-cited methodology range (~5% to ~85%) per GAO reporting · Mehrotra, Dean, Sinaiko & Sood, Health Affairs 2017 (n=2,996; 13% sought price information, 3% compared) · Pioneer Institute / DAPA Research, Healthcare Price Transparency in Massachusetts (2019; n=500; 70% want prices, 31% aware of estimators, 1 in 5 ever looked) · Brot-Goldberg, Chandra, Handel & Kolstad, QJE 2017 (11.8 to 13.8% spending drop, no price-shopping contribution) · KFF Health Care Debt Survey (2022; 41% of adults, ~100M; 51% with debt skipped recommended care over cost) · KFF, Assessing Americans' Familiarity with Health Insurance Terms (2014; 4% answered all ten) · Loewenstein et al., Journal of Health Economics 2013 (14%; 11%) · Bhargava, Loewenstein & Sydnor, Choose to Lose, QJE 2017 (61% dominated choices, ~$372, 24% of premium) · Voya Financial enrollment research (91% re-select; 49% under 20 minutes) · Sinaiko & Mehrotra, Health Services Research 2020 (reference-based pricing for imaging, 8.1% spending decrease). Figures are point-in-time and directional.

Keep reading: a market with no price tag is how a plan that cannot win stays on the menu.

From the Keel Labs research feed. A paper every week. A field note every day.