Standard of one

Fourteen medicines

Call her Ellen. She’s 72, lives on her own and intends to keep it that way. She has six chronic conditions: asthma, depression, diabetes, high cholesterol, high blood pressure and arthritis, and each condition has its own guidelines. Follow them all, and she ends up with fourteen prescriptions taken at five different times of day.

Each prescription has evidence behind it and, read one at a time, each looks right. Read together, they may not be. Even two prescriptions written for different conditions can interact in ways neither prescriber intended, let alone fourteen. Our healthcare system judges care against the standard of care: what a reasonable clinician would do for a patient like Ellen, usually one condition and one guideline at a time. A standard of one asks what’s right for Ellen across all six. Ellen’s case is her medicines, but the same question should come up when a clinician decides whether to replace a knee or when to start dialysis. Each time, a guideline written for a population has to be applied to one person.

For patients like Ellen, the harm of following each guideline on its own rarely gets traced back. A fall or a hospital stay can follow from two medicines nobody weighed together and, while the chart records the event, it does not report the cause. Each fall or admission costs a patient some independence. Get the whole regimen right, and more patients keep moving and, in turn, live better.

Ellen is hypothetical. In a 2020 review, researchers in Singapore built her regimen by following the guidelines then in force for each of her six conditions, despite none of those six guidelines contemplating patients living with several conditions at once.¹ Her situation is not unusual. Among Americans aged 62 to 85, more than a third take five or more prescription medicines at once, and more than one in five are still on regimens that risk a major interaction.²

Ellen’s doctor has little time for her visit and a long list of items to cover inside it. In that time, six guidelines, each right on its own terms, don’t add up to the right care for one person.

The appointment is the clock

In Decided Once, we described two shortages that have driven healthcare costs higher for decades: attention and trust. Our first essay took on trust. This one is about attention, and why making it plentiful leads straight back to trust.

An effective review of Ellen’s prescription list means knowing why each medicine was started, whether it still helps, what it risks, how it interacts with the others and what she wants from her years ahead. That is skilled, slow work, and on its own, her chart often can’t even answer the first question. Asked why medicines that may no longer fit are continually renewed, clinicians generally name the same barriers: lack of time, the pull of leaving things as they are, care split across many prescribers and uncertainty about what happens when a drug stops.³ Her doctor can make this decision. What’s scarce is the time to make it properly, and the opportunity to make it again when things change.

The medical appointment remains the place where a clinician examines a patient and makes judgments that no record can replace. But because attention is so scarce, the medical appointment has also become the clock that decides when a care regimen gets a real review. That clock should run on the patient’s risk; but, today, it runs on the clinic’s schedule.

A standard of one

Clinical guidelines tell us what the body of evidence supports: what is true on average for the populations the studies enrolled. Clinical guidelines, on their own, cannot tell us what to do for the patient in front of us. Translating guidelines into decisions is the work the current system is forced to ration: human judgment, applied at a standard of one. Good geriatricians already practice this way, and their research societies have written down how.⁴ The question is whether the standard of one can reach every patient who needs it, and then be reviewed when it does.

Ellen doesn’t have six bodies, one for each condition; her fourteen prescriptions meet inside one person. Imagine that the gentler medication options stop controlling her arthritis pain, e.g., the anti-inflammatory gel her arthritis guideline prefers for a knee no longer works. The same arthritis pain guideline’s next step is an oral anti-inflammatory medication.⁵ But those drugs can push Ellen’s blood pressure up,⁶ and about 7% of adults with asthma like Ellen’s react to these medications with worse breathing,⁷ and then geriatric guidelines also advise against the long-term use of oral anti-inflammatories unless alternatives haven’t worked and her stomach can be protected from ulcers.⁸ Whether Ellen has these reactions may already be in her history, if anyone has asked. And none of these guidelines is wrong by itself. But, likewise, none of them can settle how much pain, risk and lost mobility Ellen should accept.

In her life, less pain keeps her walking, and walking keeps her independent. Beyond today’s pain, the decision of what to do next weighs disease progression, side effects and the decline that could end her ability to live alone. Whether each prescription matches its own guideline is the easy question. The decision that matters most concerns the whole medication regimen, and decisions like this one are routine in the clinical life of a 72-year-old. So we return to the toughest problem: most of the judgment healthcare runs on is ordinary, and almost none of it is recorded in a form anyone can review.

Tools do exist to help. The American Geriatrics Society maintains a list of medications that may be inappropriate for older adults, and by its own terms that list exists to support shared clinical decisions, not to replace them.⁸ But a flag on a list is just a starting point and, by itself, it doesn’t tell anyone what Ellen should do when her anti-inflammatory gel stops working.

Ellen needs a qualified decision. When we say “qualified decision” we mean a decision that combines known evidence with her chart, her claims, her behavioral health, her circumstances, her risks and her own account of what she wants from her care. A qualified decision shows its reasoning and says what the expected outcome is, by her own measure, so what actually happens can be checked against it. It makes clear who pays for what and what would make the decision worth revisiting. And it holds up when an expert reviews it. Qualified is a higher bar than personalized, because a personalized decision can still be wrong. A qualified decision is also not a single right answer. In many instances, several different decisions would hold up. The qualified bar is there to rule out the ones that don’t hold up, and Ellen’s values choose among the rest. Her answer might be to keep all fourteen medications; to change several; to add a treatment; to watch more closely or simply to leave things alone. A standard of one cuts unneeded care as readily as it finds missed care, and sometimes the right call is to do nothing. Ellen’s agreement can’t be assumed either. She should understand why the plan fits her and what would change it. AI can explain the plan plainly enough for her to take part in the decision. If she declines something based on that understanding, honoring that decision is a key part of her standard of one.

The missing record

One obvious path forward is to point a capable model at the evidence and have experts review the model’s work. The expert review is where this gets hard. One review is an audit. The reviews you keep become a record, something a system can improve against and a reason to trust it with the next decision long after the model that made it is gone. Models will keep changing, and what’s been learned about which decisions hold up for which patients shouldn’t have to start over each time. But our healthcare system doesn’t keep such a record: the decisions themselves, why they were made and how they held up, kept apart from the medical record.

We have thousands of guidelines and almost no records of judgment. Our system records what was done, and sometimes why, but rarely whether a decision was right for a particular person. The reasoning that settles a conflict like Ellen’s arthritis question lives mostly in one clinician’s head. And if her doctor handles it well, still nothing records that it was handled well.

The decisions that do get recorded vary, and not always for the right reasons. Some variation comes from the patient, whose values should change decisions. Some variation comes from evidence that is too thin to settle the choice. And some variation comes from nothing more than which doctor Ellen happened to see. Among emergency physicians in the same hospitals, we see that opioid prescribing varies widely even after accounting for differences in their patients, and the risk of long-term use followed the prescriber based on their tendency to overprescribe.⁹ The first two kinds of variation should be kept and made visible. The third is what a standard of one exists to shrink. It rules out the decisions that wouldn’t hold up, so the variable becomes Ellen’s values, and not her clinician’s habits.

Omissions are the hardest decisions to see. A medicine renewed after careful thought looks identical to one renewed out of habit; a decision not to treat and a failure to notice can read the same in the record. Without a record of the reasons, considered restraint and an oversight look alike, and what isn’t recorded can’t be measured, rewarded or priced.

Medical charts also do not fill this gap, as they hold good judgment and bad habits side by side without reference to which is which. And telling them apart means judging each decision against what was knowable when it was made and, since a good decision can still end badly, using outcomes over time to test the judgment. We need to know what was decided for one person, why it was appropriate, and how it held up.

AI can already do half of this. A system can record the inputs and reasoning behind each decision it drafts, cheaply and every time, so the decision can be reviewed. It can’t make decisions reliably on its own though. In early studies on records like Ellen’s, 2024-era models handled the structured work well (matching diagnoses to medicines and spotting which prescribing criteria to apply). But at the decision itself, experts judged 84 of one model’s 202 proposed changes incorrect, though it also suggested 19 valid changes the clinicians missed; some suggestions could have caused harm, and the models were deemed to be poor judges of their own uncertainty.¹⁰ Newer models will do better. The reviewing problem will remain either way, because a model can make any answer sound reasoned. The review has to judge the decision itself rather than how well it is explained.

Expert time belongs at this stage of the review, and the fairest check is an expert deciding the same case as the model does, independently, with the two compared afterward based on objective criteria. Decided Once sets the rule for decisions that can be checked against a written policy: no party grades its own work, and neither the decider nor the grader can share an owner with whoever pays for or delivers the care. A clinical decision is harder to grade though, because several answers can be right, and so the review needs to ask whether the system’s answer is among the ones that hold up, not whether it matches the expert’s exactly. Reviewing every decision forever would only move the cost into review. As a class of decision proves itself, review can drop to random audits. Each kind of decision starts with expert review, and what compounds is the comparison between what each decision expected and what actually happened to the patient, and that comparison is what earns the system more autonomy.

A living plan

A qualified decision is made with the facts available today, and that means every decision that comes later must be a new decision. Months later, Ellen spends four days in the hospital and comes home with two new prescriptions. As before, each medication may meet its own guideline, and someone will likely reconcile her medication list. What matters most for Ellen though is whether her whole regimen still fits her, and that is a new decision. Her conditions stay with her while her doctors, her insurance and her care settings change, and the living plan is what carries through them.

The qualified decision answers what’s appropriate for Ellen now. The living plan sets out when that answer needs another look. Plentiful attention has to cover both: attention to the patient will identify that something has changed and another decision is required; and attention to the decision applies the evidence and judgment to settle it well. Noticing that a hospital stay has occurred is cheap. A discharge alert can do it. Knowing the stay calls for a fresh decision, and then reaching one that would hold up, is what consumes scarce expert time today.

We already know that flags alone have low impact. In one large European study, software reviewed the medicines of older hospital patients and sent advice directly to their doctors, who acted on the advice about 15% of the time; adverse drug reaction rates also did not improve.¹¹ A recommendation no one has reviewed gets the same reception whether it’s right or wrong. That’s the trust shortage again, seen inside the exam room. A system that only notices or sends flags is just another alert, and one that only decides still has to wait to be asked. AI is already making attention to the patient cheap, as Decided Once argued. Whether it can do the same for judgment that someone will stand behind is what has to be proven, and the record is the evidence.

That’s what we mean by autonomous healthcare. An assistive system finds the next step and leaves it in someone’s inbox. An autonomous one carries out the parts of a clinical care plan it has been trusted with, leaves the rest to the pharmacists, nurse practitioners and doctors authorized for it, and brings in a clinician when a case exceeds its authority or looks unlike the cases it has already proven itself on. It acts within boundaries Ellen’s clinicians have authorized, and with her informed agreement. After Ellen’s hospital stay, it would reconcile her two new prescriptions with the rest of her plan, draft a simpler schedule for her pharmacist to approve and book the follow-up labs the plan calls for. If a new medicine clashed with her blood pressure or her asthma, or she didn’t want it, it would bring in her doctor. Whoever operates the system answers for what it does inside those boundaries, and no record of good performance will allow it to act beyond what the law allows. Intelligence alone carries no authority. Authority is earned one class of decision at a time, and every class starts in shadow: recorded, compared with the clinician’s, changing nothing.

What has to be proven

Three tests decide whether this path is real. The first is clinical quality: can AI make decisions that hold up when an expert reviews them (including decisions not to act) and know when to defer back to the expert? The bar is high for a reason. A clinician’s mistake reaches their own patients, but a system’s mistake reaches everyone it serves. The second is productivity: can qualified decisions be made, kept current and carried out with far less human labor, including the work of building, evaluating and maintaining the record itself? The comparison is the best practical alternative, whether it is a mailed prompt, a pharmacist-led medication review or a capable model whose suggestions a clinician checks. If total human work is not reduced, all that is being built is another layer of review. The third is outcomes and total cost: is Ellen as well or better off by her own measure, at a lower total cost of care?

Passing one test doesn’t prove the next. A system that pays close attention may find more to do, or less. The count of prescriptions proves nothing on its own. What counts is whether Ellen is as well off or better, and whether her care as a whole costs less over a clinically sensible period. The first to want that answer are the organizations already being paid a fixed amount to keep her well, for whom every avoidable review and admission is pure cost. They can also pay now for what keeps her walking, because a plan that states its expected outcomes lets them see the payoff coming, and the record tells them whether to believe it. Employers who pay their own claims can make the same bet on younger patients, years before the decline starts. That’s prevention, and today it mostly goes unfunded.

One idea from Decided Once carries forward. Ellen’s plan should be agreed once, before care, by her, her clinicians and whoever pays, so no one relitigates it afterward. It reopens only when something new happens, like her hospital stay or a change in what she wants, or when someone has a real reason to appeal.

One plan

Ellen still has six conditions. She may still need many of her medications. What changes is that the decisions about them are made together, against what matters in her life, and made again when her life changes.

For years, each of her conditions has had a plan, and none of them was written for a woman who means to keep living on her own.

Kā Labs is testing this thesis. The Record will report what holds up and what doesn’t, with the evaluation data published in the open. Once care is planned for one person, another question follows: what should that person’s care cost, and who will commit to paying that price? That’s the next essay.