Decided once

Despite the rhetoric, years and many dollars have been spent building transparency between payers and providers, so that both sides can see coverage, price, and payment at the point of care. Visibility has turned out not to be enough though. Even when both sides see the same answer at the same moment, each still reviews it, because neither can—or will—take the other’s answer on faith.

Reduced to its parts, our healthcare system consists of care, a risk pool, and a decision. A person gets sick. A clinician knows how to help. Care costs money and sickness is unpredictable, so insurers pool money in advance, and someone must then decide what the pool pays for. Every complicated thing in healthcare elaborates on one of those three parts. Each layer answered a real question once, and then the layers began answering each other.

In the next version of our healthcare system, spending can converge toward the cost of care plus the true price of risk—i.e., what it costs to protect people against sickness no one can predict. Two changes make that possible: (i) AI makes clinical attention plentiful, so delivering care becomes cheaper; and (ii) neutral certification makes trust affordable, so the paperwork built around distrust can be retired. We call the second change “decided once”: a decision made one time, certified, and honored everywhere by everyone who relies on it.

Two shortages

Two shortages have driven healthcare costs higher for decades. The first is attention. Medicine is made of human attention, and an hour of a clinician’s attention cannot be made more productive the way an hour in a factory can, so, until now, the price of that hour has risen steadily with wages. The second is trust. The participants in our healthcare system cannot (or will not) rely on one another’s judgment, so they surround every transaction with layers of review: eligibility, contracts, authorizations, appeals. The United States spends nearly a trillion dollars a year on nonclinical administration, and much of it is spent reviewing transactions.¹

AI can supply the first missing input directly: i.e., a system can attend to every patient every day and call in human judgment when it is needed. To date, the doctor’s appointment has existed to ration scarce attention; but what happens when attention is no longer scarce is the subject of our next Record essay. This one is about the second shortage: trust, and it is the harder of the two.

Mature industries separate production from settlement. Banks built clearing systems and card networks because the thing they settle is money, and money is easy to verify. Healthcare never made this separation, because the thing being settled is judgment. A clinical decision is uncertain, slow to prove right, and expensive to standardize at scale. So the industry with the most decisions and the highest cost of error reviews everything, with every participant verifying every other participant’s work. AI changes this. A machine’s decision can be preserved together with the exact decision model that made it, including the policy it applied, the evidence it considered, and the conclusion it reached. Exact replay is not automatic; software and hardware changes can shift results,² so replay has to be engineered. A system built the right way can replay a decision years later, as it was made, and measure it against what actually happened. Structured clinical rules have existed for decades. What is new is the ability to inspect every decision completely, at scale, and cheaply—and that is why a shared trust layer can be built now.

Intelligence is not authority

Trust cannot come from an AI system by itself. These systems already demonstrate intelligence beyond most human capacity in bounded tasks, and they will still not be trusted on their own, because intelligence is not authority. By authority we mean the power to make a decision that others will honor without reopening the reasoning. No model has it. An institution can earn it—slowly, from evidence—and lose it quickly when performance slips.

“Decided Once” separates three roles that today are tangled inside every payer and provider. A decision supplier produces a decision and the evidence behind it; that is what a model does. A certifier examines that evidence at a scale that has not been possible before and certifies the decisions that hold up—the ones that correctly applied the policy in force to the facts presented. Clinicians sit inside the certifier in two capacities: as the reviewers who measure it against outcomes, and as the people who take the cases a model should not decide. And a backer—the party that insures the certificate—stands behind each certified decision with money, so that whoever relies on it is protected if it later proves wrong. The backer’s money is what turns a good record into something a stranger can rely on today, before the record is long enough to speak for itself. Over time the certified decision earns authority of its own: used and measured for years, its record becomes the reason to accept it, and the backer’s role shrinks toward the rare miss. At the end of that chain, participants learn to trust the institution, and through the institution, the machine. Intelligence alone creates no obligation that anyone must honor. But trust can be built.

AI has also made argument between the participants in our system free.⁶ A patient with an AI agent can produce a polished appeal letter in seconds, and providers and payers can respond in kind without limit. When every participant can argue endlessly at no cost, the most valuable thing becomes an answer that is final. A certificate carries that answer from one party to the next, so whoever receives it can accept it without repeating the work, and the backer pays if it proves wrong. Each payer still writes its own coverage policy. The certifier does not decide what a plan covers. It certifies that a decision correctly applied the policy in force to the evidence presented, so the certificate holds for anyone who accepts that policy and that evidence.

The backer’s promise is easy to misread, so it is worth stating exactly. It is a promise about reliance, not about outcomes. It does not mean that every wrong decision gets paid, and it does not mean a payer covers what its policy excludes. It means that when a provider delivers care on the strength of a certificate, or a patient forgoes an appeal because of one, the cost of the certifier’s error lands on the backer—not on the person who relied on the certificate—and the error is counted against the certifier’s record. Title insurance works the same way: it does not change who owns the land, it pays when the search was wrong, and because its money is at risk, it polices the quality of the search. The promise is narrow in three ways. It covers decisions certified before care and relied upon, not decisions in general. It covers the certifier’s error in applying the policy, not false or missing facts; each patient’s facts are still verified, and a party that supplies false evidence gets no protection from a certificate built on it. And it corrects a single case: the policy stays what the payer wrote, and the next decision is certified against it. The backer may be the payer itself, insuring its own reliance, or a third party that insures the certifier; either way, it is the party whose money is at risk. What changes is where the cost of a mistake lands, and therefore who has a reason to prevent it.

One MRI, fewer repeated decisions

Take an MRI of the knee as an illustration. Today, a doctor orders it; an insurance reviewer questions it; the doctor’s office argues back; the scan can wait weeks; and even after it all clears, the bill can enter a second review that runs for months while the patient receives letters no one fully understands. In the certified version, the order arrives with its evidence attached, and coverage and price are settled before the scan is scheduled. Payment follows once delivery is confirmed, because the questions that usually trigger review were answered before the machine was turned on.

Scale that up and the industry has two working ends. One is care, delivered when the patient needs it. The other is payment, with coverage, price, and approval settled before treatment. Together, these turn payment into a formality. The two ends meet only at certificates. Most of what sits between them today exists because answers came late. A claim is a request to settle—after care—a question that could not be settled before it. An appeal is a dispute about that settlement. When the question is settled before the care, most claims and appeals disappear. What remains is the work healthcare’s layers of review were always meant for: errors, new facts, and fraud. Patients keep every right to appeal that the law gives them. They should just need those rights less often.

Certificates turn verification from an operating expense into a capital investment. One careful verification of a decision method serves millions of cases, though each individual patient’s facts are still verified. Every reviewer removed from the system is a permanent saving.

Who can be trusted to certify?

The layer that holds the certificates must be neutral, meaning no participant should be able to improve its own position by weakening the others’ confidence in it. That has to be built into the structure. The company that makes a decision cannot be the one that grades it, and neither of those two can share an owner with the backer, with a payer, or with the providers who deliver the care. The backer must be separate from the certifier for a reason beyond neutrality: the backer is the one whose money is lost the moment quality slips, and that is what makes it the certifier’s most attentive customer.

Credit rating agencies were supposed to be neutral certifiers too. They failed in the years leading up to the financial crisis of 2008 partly because the issuers being rated paid for the ratings. Any certifier faces the same pull, because someone has to pay for certification. The constraints on any workable model are already clear. The fee must be blind to the verdict, so a certified decision costs the same whether it approves or denies. The parties who rely on certificates, not the party being graded, should bear that fee. And the certifier’s errors must be public and must cost it—in the backer’s premium, in its published error rate, and ultimately in whether anyone still accepts its certificates. A certifier paid the same for yes and no, whose mistakes are counted in public and priced by a backer with its own money at risk, has only accuracy to sell.

The first trusted certifier grows stronger with time in two ways. First, within each kind of decision, use is what builds the record. Every certified decision eventually meets its outcome, and a certifier with ten thousand reviewed outcomes can be trusted with more than a certifier with ten. Second, across kinds of decisions, certified answers accept one another, so each new area increases the value of every existing certification. The review problem shrinks from every organization reviewing every other organization to everyone relying on one, with an independent auditor reviewing the certifier.

Adoption then feeds on itself. A patient prefers a practice where care is continuous. A provider prefers a payer that answers before treatment. A payer prefers providers whose certified claims need no review. Each participant that accepts certification shrinks the territory where review still pays, while those that refuse keep running the full review apparatus over a shrinking share of their business. At some point, refusing the certificate costs more than accepting it. That is when a product becomes infrastructure, and it is how card networks and clearinghouses became the default. This network effect is the most valuable thing about the arrangement, and it is not neutral in its effects: participants inside gain, and those outside lose ground. That is one more reason the certifier must be neutral and open to any participant that meets its standard, on published terms.

Not everyone gains at first. For some payers, friction is margin: e.g., a denial that is never appealed, a payment that sits for months, a patient who gives up. Certification removes those levers along with the cost of pulling them. So the first adopters will be the organizations for whom review is pure cost: e.g., groups paid a fixed amount to keep people healthy, plans in value-based arrangements, and providers who carry risk. Their savings arrive immediately. Others follow when the fight costs more than the margin it protects, and when dealing only with counterparties who have not adopted certification becomes too expensive. Free argument makes every denial more expensive to defend, and a certified decision is one the other side cannot profitably contest.

Because acceptance of the institution survives every generation of models, ownership matters more than technology. Authority of this kind takes years to earn and can be destroyed the day the certifier is sold to someone with a stake in the outcome, so the certifier needs a permanent, neutral owner that takes no side. Whether that owner is a company, a member-owned utility, or something else matters less than three constraints. First, it is governed by those who rely on it—payers, providers, and patients—with no single participant able to control it. Second, it publishes how it certifies, how often it is wrong, and what happened afterward, because transparency—not a label—is what makes neutrality verifiable. And, last but not least, it cannot be sold to anyone with a stake in the outcome. The rating agencies and the credit-score companies show that a certifier can be a business; they also show what happens when the certifier’s methods are opaque or its fees come from the party being graded. Participants can build on the institution and compete to supply decisions without owning it.

What must be proven, and why now

Whether AI lowers healthcare spending depends on the setting, because the setting decides who pays for the argument. Where everyone is paid for each service performed, AI arms both sides of the fight: more documentation and billing on one side, more utilization review and post-payment audit on the other. Each side’s tools pay for themselves against the other side, and the cost of the fight rises even when care does not. Where the organization that decides also bears the cost—a group paid a fixed amount to keep people healthy, a plan in a value-based arrangement or a provider carrying risk—the fight is pure overhead, and certified decisions retire it. The fix is not smaller risk pools; it is finality. A decision made once and honored everywhere lowers spending wherever it is adopted, and it will be adopted first where the argument costs the decider directly. The goal is that the cost of each unit of care falls steadily while the total amount of care grows—an easy place to see why is that, today, more than one hundred million Americans live where primary care is too scarce to reach them,³ to say nothing of specialists. It is also important to note that not every cost yields to this solution: e.g., drugs under patent, hands-on caregiving, and buildings will not become cheaper this way. What does become cheaper is deciding, coordinating, and reviewing. Administrative savings recur every year once the work is gone, and the productivity of care will improve as the models improve.

Three things must be shown before “Decided Once” is more than an argument: (i) certified decisions must match or beat expert reviewers on the same cases (including, most importantly, in the eyes of the expert reviewers); (ii) payers must accept certificates without re-reviewing them—which they will do only when a backer makes the cost of a wrong certificate someone else’s problem, so this is a test of the backer as much as of the certifier; and (iii) the work retired must cost more than certification, backing, and monitoring combined. Each premise can be tested on a single decision class, and that testing is already in progress at Kā Labs.

In sum, the old arrangement—everyone reviewing everyone—was tolerable while argument was still expensive. Now that argument is free, we see four possible outcomes. The first is stalemate: i.e., we keep the current arrangement and let machines argue with machines. But free argument will only multiply the number of disputes without settling them, and disputes that do not settle pile up as unpaid claims and delayed care. As an example, the federal system built to arbitrate surprise medical bills expected about twenty-two thousand disputes in its first year and then received more than three hundred thousand.⁴ And many of those disputes were not resolved until the end of 2025, after years of adding dispute-resolution capacity.⁵ The second is private monopoly: i.e., one dominant company imposes final answers. The problem here is that the losing side rarely accepts a competitor’s verdict, and it will fight in court and in legislatures until the arrangement fails. The third is public administration: i.e., the government simply imposes the rules. But we already have public bodies deciding some of these questions, including the surprise-billing arbitration example above, and they tend to rule case by case, after care, and only within the scope the law gives them. The fourth outcome—and the only one that improves with use—is neutral, machine-driven infrastructure, where the adversaries accept a shared certifier and stop paying for the fight.

That is the layer we started Kā Labs to build toward. Kā Labs builds companies that supply those decisions, on a shared substrate that makes a decision certifiable in the first place: i.e., the decision model, the evidence, and the replay are preserved by construction, so a certifier can inspect them and a backer can price them. We want to prove certification in practice with payers and providers, one bounded decision class at a time. We are also standing up the first version of the institution that certifies them with the standard a certified decision has to meet, the evaluation data published in the open, and the first payers and providers convened around a single decision class. We intend to compete to supply decisions, not to own the verdict. That is where the value goes: to decisions good enough to certify, and to everyone who stops paying to review them. The shared institution makes reliance possible. Kā’s companies must earn their place by delivering better care and reducing the work required to pay for it.

For decades, healthcare priced scarce attention and scarce trust into every transaction. AI is already loosening the first. The second will take longer, because trust has to be earned in public, case by case, with money behind it. As both ease, most of the administrative economy loses its reason to exist. What remains is care, a risk pool, and a decision made once, final everywhere.