By HANA KAHLEOVA

“Will this only get worse?”

Patients with type 2 diabetes have asked me some version of that question for my entire career. For most of it, the honest answer was “probably.” We treated diabetes as a one-way street: glucose climbs, medications pile up, complications follow. Good care slowed the slope. It didn’t reverse it.

A year ago today, that assumption stopped being true, at least on paper. On October 1, 2025, CMS added E11.A, “type 2 diabetes mellitus in remission,” to the ICD-10-CM code set. To use it, a clinician has to document an HbA1c below 6.5% for at least three months after a patient has stopped all glucose-lowering medication. That’s the definition the American Diabetes Association, the Endocrine Society, the European Association for the Study of Diabetes, and Diabetes UK agreed on in 2021. “Well controlled” doesn’t count. The code exists to separate remission from management.

That’s a real concession. When CMS creates a code for an outcome, it is saying the outcome is real, measurable, and common enough to track.

The evidence behind it isn’t new. In DiRECT, a cluster-randomized trial run through ordinary UK primary care practices, 46% of participants in a structured weight-management program were in remission after a year, compared with 4% on usual care. About a third were still in remission at two years. These weren’t patients at an elite research center. They came from routine general practices. That’s what makes the result a policy question, not a curiosity.

So here’s the policy question. One year in, what has the code changed?

As far as payment and quality measurement go, almost nothing.

A code with nowhere to go

E11.A gives clinicians a precise way to document remission. It does nothing to reward pursuing it, measure it, or pay for the care that produces it. Three systems decide what actually happens in a clinic and a health system’s budget. None has caught up.

Payment. DiRECT was intensive: several months of total diet replacement, then structured food reintroduction and regular follow-up for two years. Compare that with what Medicare covers. Medical nutrition therapy for diabetes is capped at three hours in the first year and two hours a year after that. Medicare’s one structured lifestyle program, the Diabetes Prevention Program, is open only to people who don’t have diabetes yet. Once you’re diagnosed, the benefit built for behavior change is gone.

Meanwhile, the money is flowing somewhere else. GLP-1 drugs are effective and the right choice for many patients, and the global market for this class is projected to reach roughly $100 billion a year by 2030. That isn’t an argument against the drugs. It’s a map of where the system is willing to spend, and where it isn’t.

Quality measurement. HEDIS, Medicare Advantage Star Ratings, and MIPS reward hitting glycemic and blood pressure targets. None of them give a plan or practice credit for a patient who no longer meets the definition of diabetes. That patient just drops out of the denominator.

And the bar keeps moving the other way. Last month, CMS released draft cutpoints for the 2027 Star Ratings, and roughly half became harder to hit than in 2026. The kidney health evaluation measure for patients with diabetes tightened by 7 to 10 points, depending on the star level. Plans are being pushed to manage diabetes more precisely than ever, on a scorecard with no line for the best outcome a patient can have.

Incentives. Put those two together and the business case writes itself, in the wrong direction. A health system that invests in getting patients into remission pays for an intensive program out of pocket and gets nothing on its quality scores for the result. A system that manages the disease indefinitely gets paid for every visit and every refill, and scored on every target.

Picture a patient two years past diagnosis, on two medications, with an HbA1c of 7.2%. If her health system enrolls her in an intensive program and she reaches remission, it has spent money it won’t recover, and its diabetes measures look exactly as they would if she had moved away. If it adds a third medication instead, it gets paid, and she counts toward the target. Every incentive points to the prescription.

The infrastructure already exists

E11.A was created to solve a documentation problem, so coders would stop guessing when a physician wrote “remission” in a chart. But a system that can identify remission that precisely can also reward it. Nothing downstream uses it yet.

Three changes would close the gap. None requires new clinical evidence.

First, build E11.A into quality measurement. NCQA and CMS could add a remission measure to HEDIS, Star Ratings, and MIPS using the code and the 2021 consensus definition. A health system that helps a patient leave the diabetic population should get credit for it, not just lose that patient from the denominator.

Second, create a payment pathway for remission programs. Cover time-limited, structured weight-management and dietary programs modeled on DiRECT for patients in the first several years after diagnosis, when remission is most achievable. Pay for engagement and glycemic outcomes, not visit counts. CMS’s Innovation Center is the natural place to test it.

Third, fund the maintenance phase. Relapse is common without follow-up. The lower-intensity support over several years that separates durable remission from a temporary dip in HbA1c is the least-funded part of the whole pathway, and it’s where programs fail.

Who this is for, and who it isn’t

Remission isn’t realistic for everyone with type 2 diabetes, especially later in the disease, after substantial loss of insulin-producing capacity. For many patients, long-term medication remains the right, evidence-based goal.

The argument is narrower. For patients who could realistically get there, whether anyone offers them the chance shouldn’t depend on whether their health system’s scorecard happens to notice an outcome CMS created a code to describe.

More than 38 million Americans have diabetes, the vast majority of them type 2. CMS has done the hard part: it built a precise, clinically rigorous way to say a patient’s diabetes is gone. With final 2027 Star Ratings expected this month, now is the time to ask why the scorecard still can’t see it.

Hana Kahleova, MD, PhD, is a physician board-certified in internal medicine, endocrinology, and lifestyle medicine, and director of clinical research at the Physicians Committee for Responsible Medicine, a nonprofit that advocates plant-based nutrition. PCRM had no role in the DiRECT trial or in the CMS policies discussed here.

By STEPHEN FARBER

Part one argued that patient access was never the finish line; it was the foundation. This is about what we build now that we are standing on it.

Every good game show has a second round, and the second round is always harder. The first round was about getting health information out of the buildings that held it. Regulation, standards, and a decade of unglamorous plumbing have brought us to an important threshold. Individuals can increasingly pull records from multiple institutions into one place. The board paid out. The audience cheered.

Then a quieter and stranger question arrives. You are holding years of visits, laboratory results, medications, and discharge summaries, and you realize that having your history in one place is not the same as understanding it. Assembly was the first hard problem, but it was not the last.

The record is finally readable

This is where artificial intelligence walks onto the stage, and the promise is real. Few clinicians have time to read a patient’s entire history every time a decision must be made. Appropriately safeguarded AI can summarize years of encounters, support medication reconciliation, explain a laboratory trend in ordinary language, and surface a pattern buried across multiple providers. Large consumer platforms have already begun connecting AI assistants to medical records; Amazon’s Health AI, for example, draws on a person’s One Medical history to provide personalized explanations and guidance. Once the record can be assembled, a capable assistant can finally make more of it legible.

Legibility, however, is not reliability. The farther AI moves from explaining a record toward acting on it, the thinner the evidence becomes. A 2026 scoping review of agentic AI in healthcare found that most of the available studies were exploratory and lacked robust clinical validation, and that only one trial involved patients. An assistant working from an incomplete, duplicated, or subtly wrong record can produce a fluent and plausible conclusion that is still wrong. AI does not make data quality less important. It makes the consequences of poor data quality harder to ignore.

AI cannot read what was never written

There is a more fundamental limitation, and it has less to do with the model than with the record itself. A longitudinal record can tell us what happened, but it rarely tells us enough about why. It preserves the encounter, the diagnosis, the result, and the prescription while losing much of the reasoning that surrounded them. Why was one treatment chosen over another? What was the person hoping to preserve at that point in life? Who else was helping? What changed after the visit, and what did a family member notice before anyone thought to schedule another one?

Suppose a daughter has spent the past year helping her mother manage a handful of chronic conditions across three practices. She has done everything the first round asked of her. She has assembled the record, and she can see every visit, every result, and every prescription in one place. What she cannot find anywhere in it is the moment, sometime in the spring, when she noticed her mother was no longer quite herself: the skipped meal, the walk that got shorter, the pill organizer that stopped emptying on schedule. None of that was an event, and none of it generated a claim. By the time the record catches up, it will document the consequence, a fall or an admission, without the months of evolving context that would have made the consequence understandable and might have created an earlier opportunity to respond.

Those signals are not peripheral to health. They often determine whether a plan is realistic, whether a change is meaningful, and whether someone follows through, and they explain why a clinically complete record can remain an incomplete picture of a person. AI can analyze what it can reach, but it cannot recover context that no system collected, decide whose version of the story should govern, or determine who may act on what it finds. Those questions belong to continuity, governance, and accountability rather than to the model.

There is a further gap that I will only note here. The daughter is not the only one standing outside the record. The aide who comes three mornings a week, the care manager at the Area Agency on Aging, the pharmacist who noticed the refills slipping: most of the people who see the mother most often were never on the patient access rails at all, and their systems, where they have systems, are still in the first round. That deserves its own treatment. For now, it is enough to say that the record the daughter assembled is incomplete not only in what it captures but in who was ever able to write to it.

The real contest is who gets to act

Put these developments together and the shape of the next few years comes into focus. The assembled record is now valuable enough that something will act on it. The open questions are what will act on it, for whom, and under whose authority. The same missed refill could prompt a reminder from a daughter, an outreach call from a care team, an intervention from a health plan, or a commercial prediction by a platform.

Value invites capture. The parties that gain access to the assembled record, particularly those whose business models depend on it, will influence what that record becomes: a resource that helps an individual, family, and care team make better decisions, or an input optimized primarily for a model, marketer, underwriter, or institution. The infrastructure to assemble a record is increasingly available. The infrastructure to ensure that the record continues to work for the person whose life it describes is not.

Concentration also creates exposure, and the protections around the record can change as it moves. The largest healthcare data breach reported in 2025 involved Conduent, a business associate whose back-office services extended across healthcare organizations and government agencies; the reported number of people affected was later revised to more than 62 million. Even when records remain technically distributed, a common interface, credential, or intermediary becomes a concentrated point of access and therefore a valuable target. And when an individual directs health information to an app that is neither a HIPAA-covered entity nor a business associate, HHS has explained that the data the app receives is generally no longer protected by HIPAA. Other protections may apply, including the FTC’s Health Breach Notification Rule and state consumer health privacy laws, but the result is a different and more fragmented legal regime. The point is not to stop assembling the record but to make the tradeoffs visible, and to treat privacy, security, and governance as part of the product rather than language buried in its terms.

What comes next is continuity infrastructure

I am not a neutral observer. HealthHive, the company I co-founded, is built around this problem. But the architectural question exists regardless of which organizations ultimately answer it, and the daughter makes it concrete.

What she needs is not a better pile. She needs a place where the clinical record is preserved as the authoritative account of the care that was delivered; where the context that only she and her mother possess can be added without silently rewriting the clinician’s account; and where the right understanding reaches the right participant at the moment a decision is being made. I think of this as continuity infrastructure: a governed, longitudinal layer of understanding that sits alongside the clinical record rather than replacing it.

Two properties of that layer are especially important. The first is that it distinguishes source from interpretation. The daughter’s observation that her mother’s walks have shortened is not a clinical finding, and it should not masquerade as one; but neither should it vanish because it lacks a billing code. Keeping the two visibly distinct, with provenance attached to each, is what allows a clinician to trust the record and the family to trust that their contribution was heard. The second is that permissions and changes are visible. The daughter’s authority to contribute and to see information should rest on an explicit role or authorization rather than on the relationship alone, and those roles change. A sibling joins the effort, a home health aide arrives, a new specialist enters the picture, and at some point, the mother herself may become less able to direct her own care. A layer that cannot make those transitions legible will either exclude the people who matter or admit everyone with no accountability at all.

AI has a real role inside this architecture. It can help the daughter prepare for her mother’s appointment without pretending she is the physician, and it can help the clinician see what changed at home without asking the family to reconstruct six months from memory. An organization can use it to recognize risk before the next encounter rather than after the next claim. What it should not do is silently become the owner of the story or the final authority over what happens next.

A continuity layer is not exempt from this danger. By making information easier to reach, it can become another concentrated and attractive point of access. It should therefore be judged not only by what it can assemble, but by what it refuses to collect, how clearly it preserves provenance, whether access is visible and auditable, and whether permissions can be withdrawn or changed without dismantling the person’s history. Otherwise, continuity becomes another name for concentration.

Why hasn’t this already happened?

A fair objection is that pieces of this layer already exist. Patient portals accept messages, patient-reported outcomes capture experience, remote monitoring brings information from the home into clinical workflows, and record aggregators can add context. Perhaps the incumbents simply absorb this function, governance and all.

That possibility cannot be ruled out. But these tools are generally governed by the institution, program, or vendor that deployed them. The information they collect does not reliably travel with the person when they change doctors, move between plans, or enter a new stage of life. Nor were most designed to preserve the family’s account alongside institutional records without confusing the two.

Context can be collected today. The test is not whether an incumbent can add these features. It is whether continuity survives when the incumbent’s relationship with the person ends. An institution can sponsor the layer, but if changing doctors, plans, employers, or vendors also changes who controls the accumulated story, the system is not longitudinal in the way that matters. A continuity layer earns the name infrastructure only if identity, permissions, provenance, and portability persist independently of any single sponsor.

Who pays is not who governs

The person whose life the record describes will not always be the paying customer, nor must they be. Health systems, health plans, employers, community organizations, and other institutions may have sound reasons to fund continuity because they benefit when people and the people supporting them make better decisions. Payment and governance are not the same thing. Sponsoring the infrastructure should not confer the right to appropriate the person’s story or quietly optimize it for a purpose the person does not understand. The economic bargain should be explicit: the sponsor is purchasing a defined service and a measurable benefit, and the person’s longitudinal story is not part of the purchase.

That boundary belongs in the product’s permissions, contracts, and technical architecture. If it depends only on the sponsor’s goodwill, “person-centered” is branding rather than governance.

The decisive question will not be which system knows the most. It will be which system remains answerable to the person after it knows.

The second round

Patient access created the possibility of something healthcare has never had: an understanding of the person that can persist across encounters, institutions, and changing relationships without belonging exclusively to any one of them. Part one called patient access a foundation rather than a finish line, and a foundation is an invitation to build, but it does not determine what goes on top. We can build intelligent tools that help individuals and the people they trust understand and steward health over a lifetime. Or we can allow the pile we worked so hard to assemble to become the most efficient input ever created for systems that were never designed to serve the person.

The second round is subtler than the first because the Whammy no longer looks like losing your data. It looks like winning it and then handing it to whoever asks first.

Big bucks. No Whammies. This time, watch the board more carefully.

Stephen Farber is Co-Founder and CEO of HealthHive, and the Vice Chair on the Board of Directors of the American Society on Aging.

By JUSTIN MCKAY

When I’m brought back into the diagnostic area, one of the first things I see is a lead apron, well-worn and folded into itself, about ready to slide off the table where it was absentmindedly placed. For many people, this sight would not be startling, but I know firsthand the life-threatening dangers associated with garments that are not cleaned, stored, and maintained properly.

It is time that patients and staff understand this danger, and regulators take action.

My co-founder and I contracted ringworm in two different hospital networks from wearing unclean garments just like these, handled by hardworking, decent individuals in the imaging and surgery departments we routinely visited.

Healthcare professionals should be more mindful of the cleanliness and the condition of these garments, but those efforts alone are not enough to prevent hospital acquired infections. What my experience and the experiences of so many others indicate is that what is not measured, is not managed. To protect patients and medical professionals alike, there needs to be a clear regulatory protocol for cleaning and maintaining high utilization equipment and garments.

Study after study shows the need for this structure. Research shows that when regimented cleaning and maintenance protocols don’t exist, bacteria grows. In fact, Association of periOperative Registered Nurses (AORN) 2025 guidelines spell out this danger pretty clearly: aprons, vests, skirts, and thyroid shields have contamination rates between 62.5% and 100%. This finding is especially worrisome since, as the National Institutes of Health has found, 80% of chronic infections and 65% of microbial infections are associated with biofilms (thin, sticky layers of microorganisms that form on surfaces).

The question I keep coming back to is: why don’t the standards reflect what we already know? And, trust me, on this issue we know a lot about the danger.

A 2014 study found 84% of tested interventional radiology aprons were colonized with S. aureus and Tinea, and 12% with MRSA while a 2019 report detected bacteria on 80% of operating room apron swabs, with an 80% recurrence rate within six hours of disinfectant wipe-down (wipes alone are ineffective). The dearth of effective cleaning practices are also well-documented. A study released in 2025 found 51.7% positivity across 120 swabs from 20 aprons, only dropping to 41.7% after cleaning while another report issued the same year documented 80.7% contamination across 88 garments. Despite these threats, only two-thirds of departments had any cleaning protocol, or internal compliance records, none had a defined frequency for cleaning.

Clearly these garments are pathogen reservoirs.

When facilities only use wipes and don’t completely remove the bioburden, they’re fostering conditions for surviving bacteria. When I ask imaging departments and hospital administrators who is responsible for making sure these garments are clean, after a delay, a sigh, or a confused look, the answer generally is, everybody.

There is a name for this phenomenon: diffusion of responsibility, a principle that explains that individuals are less likely to feel personal accountability in a group setting because they assume others will act.

Some organizations have tried to outline care for garments like lead aprons.

Let’s start with Spaulding, the classification system from 1968 on which the entire infection prevention field still relies. This classification system was built for a different era, when no one was thinking about a garment that gets passed around all day and never sees a washing machine. Which is why, under Spaulding, lead aprons fall into the “non-critical” category. Current evidence shows these garments routinely come into contact with patients during each shift (Honigsberg, 2017; AORN, 2017), directly challenging the assumptions underlying their “non-critical” classification.

Next there is The Joint Commission, which requires facilities to inspect protective garments and document those inspections. Those rules sound great until you realize there’s no mandate for cleaning protocols or inspections. No required frequency. No validation that decontamination ever happened. Checking a box that says “apron accounted for” tells you nothing about what is living on it.

AORN has done the hard work of documenting the problem and it recommends cleaning, but the guidelines are voluntary. There is no enforcement mechanism, no audit trail, no accreditation consequence for ignoring them. The evidence is right there in the guidelines. The same organization that admits that garments are contaminated, and come into contact with patients each shift, only directs that you should consider wiping them down after use. What is missing is informed regulation.

Finally, the U.S. Food and Drug Administration (FDA) regulates lead aprons as medical devices, but its authority ends at the point of sale. How a facility cleans, stores, or reprocesses those devices after purchase is not on their radar. That is almost completely unregulated. Instructions for use (IFUs) exist on paper, but compliance is unaudited. The Occupational Safety and Health Administration (OSHA) covers the radiation hazard, but is completely silent on the infection hazard the garment itself may create. The bottom line: while 62-100% of these garments are contaminated, not one regulatory body requires anyone to properly clean them.

We must close this deep gap in oversight.

The defense that “nobody told us we had to clean it” is not going to hold up when the published evidence clearly outlines the threat. These risks are not theoretical. Candida auris cases have risen over tenfold from 2019–2024. Facilities that treat contaminated garments as somebody else’s problem are sitting on real, growing liability exposure. And the longer the literature piles up, the harder it gets to argue ignorance.

Hospitals are already being penalized for infections they can track. The Centers for Medicare and Medicaid Services withholds 2% under value-based purchasing and hits the bottom quartile with an additional 1% penalty through the Hospital-Acquired Condition Reduction Program (HACRP). Studies show a single surgical site infection can have a median absolute cost increase of $24,230. What happens when one of those infections traces back to something nobody thought to look at?

What Should Happen Next

Thousands of perioperative professionals were in the same room in New Orleans for AORN 2026. Attendees should demand action. They demanded action a few years prior regarding surgical smoke concerns, and their voices were heard.

AORN’s recommendations are a start, but they are not forceful enough to change behavior. It is time to tie compliance to accreditation, and make contamination data part of the audit. The Joint Commission still stops at visual inspection; the FDA still classifies these garments alongside items that don’t sit against open surgical fields for hours. Both bodies, and other state and federal regulators, must do more.

I encountered this threat in a very personal way 13 years ago, and now scientists have documented the threat in a way that should be impossible to ignore. It’s time regulators listen and take action.

Justin McKay co-founded RadCare Services after personally contracting ringworm from contaminated hospital-issued lead garments in two separate healthcare networks.

By BRIAN FENGLER

Every night in the emergency department, the electronic health record hums with small demands for attention. A nudge about sepsis. A best-practice advisory. A documentation suggestion that arrives between patients and asks for one more click.

Most of us treat those prompts as background noise. Almost none of us stop to ask whether a given nudge is, in fact, a query, or whether it was built to the same standard a clinical documentation specialist has to meet.

In August, that question got an official answer. And it is not the answer many health IT vendors have been selling.

The American Health Information Management Association and the Association of Clinical Documentation Integrity Specialists released the 2026 update to their Guidelines for Achieving a Compliant Query Practice, shaped by feedback from nearly 200 individuals and groups. The principle is plain: “Whether a query comes from a person or technology, the standard for compliance should remain the same.”

Then comes the sentence that should end a familiar sales claim: “Organizations should not assume that a vendor-supplied or EHR-integrated tool produces compliant queries by default.”

For years, “compliant by default” functioned as an unspoken promise around AI-enabled clinical documentation and utilization review tools.

If a system could draft a query, the thinking went, it could draft a legal one. The 2026 guidelines close that gap. Compliance is no longer something the software is presumed to get right. It is something the health system that buys it has to prove.

I am an emergency physician. I co-founded EvidenceCare in 2016 because I watched clinicians drown in prompts long before generative AI arrived. What the new guidelines make clear is that the prompt problem was never only about alert fatigue. It was also about who is accountable when a machine asks a doctor to document something that may change how a case is coded, scored, or paid.

A query is what it does

Section XI of the guidelines, titled “Query Technology,” brings four categories of tools into scope: computer-assisted physician documentation, computer-assisted coding, large language model and generative AI platforms that draft or deliver query language, and EHR-integrated advisories and alerts.

The guidelines define a query by function, not by label. A communication is a query when it “presents a provider with a specific diagnosis or documentation option for consideration in connection with a specific patient encounter,” and that definition applies “regardless of how the communication is labeled or where it appears within the technology platform.” Prompts, nudges, advisories, and alerts are named explicitly. A tool marketed as passive decision support does not receive an exemption because of what the vendor calls it.

The definition is narrower than it first sounds. Passive clinical decision support that surfaces reference information without steering a provider toward a specific response for a specific patient is not a query. The practical test is this: does the tool point a clinician toward a particular answer for a particular patient? When that is unclear, the guidelines say the communication should be treated as a query and held to every applicable compliance standard.

On a busy shift, that distinction is not academic. “Here are the sepsis criteria” is reference material. “Consider documenting severe sepsis for this encounter” is a query. Calling the second one a nudge does not change what it is.

The same rules as a handwritten note

An AI-generated query is held to the same standard as one written by hand. It must include relevant clinical indicators sourced from the health record. It must avoid subjective interpretation. Its title, tracking identifier, and subject line must not reference a desired diagnosis, a reimbursement impact, or a quality measure. Answer options must be clinically supported, must include an “other” option or room for an alternate explanation, and must not use formatting, emphasis, or ordering that steers the provider toward a preferred response.

Two provisions are easy to overlook, and they are where a lot of product design quietly fails. Automated templates “should not auto-populate the answer field or suggest diagnoses without provider validation.” And if a technology-generated query does not produce the response an organization wanted, sending a follow-up manual query solely to override the provider’s earlier answer is noncompliant unless new clinical indicators exist.

The guidelines are candid about where the risk originates: tool configuration, training data, and system design, especially when those choices introduce directional bias or put reimbursement ahead of clinical accuracy.

What “compliant by design” would require

The guidelines set the outcome and leave the architecture to the vendor. Meeting that outcome is unlikely to mean running a large language model over a chart and shipping fluent text. It usually means choices made before the first query is generated.

Templates have to constrain the model, and the model has to be tested against those constraints. If the formatting logic cannot bold, highlight, or reorder answer options, the output is shaped before delivery instead of cleaned up afterward. Answer fields cannot pre-select a response. Clinical indicators need to be traceable, so an organization can show how each indicator was sourced and why each answer option was included. And human review has to live somewhere in the workflow. The guidelines allow review before delivery or “ongoing structured auditing of query output,” and they expect regular structured auditing when there is no pre-delivery review.

From the product side, that is the difference between an AI that writes well and an AI that can survive an audit.

The buyer owns the risk

On accountability, the guidelines leave little room for interpretation: “Organizational process owners retain accountability for every query delivered to a provider, including those generated through automated or AI-assisted processes.” Vendor vetting and compliance review of technology query tools sit alongside human review, audit frequency, staff training, and the process for reporting and correcting noncompliant queries. The burden of proof sits with the health system that adopts the tool, not the company that sold it.

Regulators are moving in the same direction. As summarized by Arnold & Porter, the HHS Office of Inspector General listed “AI-generated prompts encouraging unsupported coding” among “potentially abusive risk adjustment practices” in its March 2026 compliance guidance for Medicare Advantage.

Query practice drew enforcement before AI entered the picture. In December 2024, Independent Health agreed to pay up to $98 million to resolve False Claims Act allegations that a subsidiary retrospectively searched medical records and queried physicians for information that would support additional diagnoses and higher risk scores. The allegations were resolved without a determination of liability, and no AI was involved. What the case shows is how closely regulators already examine the query process.

Put that settlement beside the OIG language and the 2026 guidelines, and the pattern is hard to miss. Query practice was already an enforcement target. Artificial intelligence did not invent the risk. It scaled it, then packaged it as a feature.

For leaders evaluating an AI-enabled CDI or utilization review tool, four questions now belong next to the familiar ones about accuracy, integration, and speed:

1. How does the tool prevent a leading query, in specific terms, instead of stating that it does?
2. Which tools in our environment, including EHR alerts, meet the definition of a query?
3. Who reviews queries before delivery, or how often do we audit them afterward?
4. What is our process for correcting a noncompliant query once we find one?

Before the next audit cycle

The central point of the 2026 update is that “compliant by default” was always a marketing claim, and it is now a claim the adopting organization has to test for itself. Query volume, provider burden, and the scale of documentation review make some automation inevitable. Done well, it remains one of the better answers to the pressure CDI teams are under.

But “done well” now has a written definition, and it is more specific than many products on the market were designed to satisfy. Organizations that treat the update as a design requirement, and that review every alert, prompt, and AI-drafted query against Section XI, will be in a stronger position when the next audit cycle arrives. More importantly, they will still deserve the trust that the query process depends on.

Brian Fengler, MD, is the Chief Medical Officer and co-founder of EvidenceCare, a clinical performance improvement company he co-founded in 2016.

By EVAN REID

We reconcile medications at discharge because we decided, correctly, that the handoff out of a hospital is a clinical event with its own failure modes. Then a patient dies, and the same hospital that would not let them leave with an unreconciled med list hands the family a folder and a phone number for the funeral home.

Grief is separate, and the system is at least gesturing at that. What follows the folder is work. Months of it, procedural, deadline-bearing, and owned by nobody on the care team.

I run a consumer estate-settlement publisher, so I see the searches families make in the days after. They search for form MV-39. They search for whether they are allowed to use their father’s car, and for which county holds the hearing. They do unfamiliar administrative work at the worst cognitive moment of their lives, and they do it alone.

First, the objection

The reflexive answer: this is what a trust is for. It is a good objection and it fails on timing.

A trust is a pre-death instrument. The family standing in a corridor at 2am cannot act on it. Whatever should have happened in 2019 is off the table now, and an intervention available only years in advance answers nothing at the moment of discharge. Here is why that matters: the discharge desk is the last moment at which anyone official is still in the room.

The objection also concedes too much. A fully funded trust avoids probate. It does not touch death certificates, account closures, creditor notice, Social Security, final tax filings, or retitling a vehicle. Many trusts are only partly funded, so assets never retitled into the trust go through probate anyway, and families often learn this only after the death.

Take the car. Retitling a dead parent’s car means finding the right state form, the right affidavit, and an answer to whether anyone may use it in the meantime. No trust touches any of that. It is administrative sediment, identical for the well-planned estate and the unplanned one, and it lands on the family in week one.

So a trust changes who handles part of the work. The family still handles the rest.

What the burden actually is

Let’s break it down. Here is what nobody prices at discharge.

Court filing fees are the small part, a median of about $233 across the fifty states, ranging from $30 to $750, according to our 50-state probate cost study. That is the number people expect. The damage comes from attorney fees.

Only seven states set ordinary attorney fees by a statutory or presumed schedule. In the other 43, the price of settling an estate stays hard to know until the work is done. Lawyer billing rates have a median of about $312 an hour across the states, per Clio’s Legal Trends data. If a straightforward estate takes ten to twenty-five hours, which is our working estimate, that is roughly $3,100 to $7,800 before anything goes wrong.

Then there is the asset nobody can pause. Holding an inherited home costs a national median of $6,312 a year, or $526 every month, in property tax and insurance alone, before a mortgage payment, before utilities, before a roof. That figure ranges from $2,974 in West Virginia to $12,907 in Florida, per our carrying-cost study. The meter starts the day of the death and it does not care whether the family has letters yet.

The rules also change at the state line in ways nobody would guess. Nearly every state offers some small-estate shortcut around full probate, with thresholds from about $15,000 to $400,000 in Wyoming, and many of them cover personal property only. The same estate can be a short affidavit in one state and a court process in the state next door. Families do not know this. Why would they?

Why the hospital and not someone else

The usual reply: this is not clinical, so it is not ours.

We already accepted that discharge carries non-clinical obligations with clinical consequences. Medication reconciliation is not a treatment. Neither is confirming a patient has a ride home, or that the prescription is affordable. We do those because we recognized the handoff as the point where preventable harm concentrates.

Death is a discharge. It is the one discharge where the hospital holds complete information, a captive moment, and a documented point of contact, and it is the only one where we hand over nothing but a folder.

The counter-argument I take seriously is scope creep. Hospitals should not own every downstream social problem, and staff are not estate administrators. Agreed. So the ask should be small.

Next steps: the proposal

The ask is one checklist item.

At the point of death certification, the discharging hospital should hand over one state-specific sheet answering four questions. How many certified death certificates this state’s agencies will require, and where to order them. Whether this estate plausibly clears the state’s small-estate threshold. Which county court has jurisdiction. And the first filing deadline that exists in this state.

That is all. Four facts, one page, generated from the decedent’s state of residence, using information the hospital already holds. It is not legal advice, it creates no liability the hospital does not already carry, and it occupies the same slot in the workflow that medication reconciliation does.

The reason to do it: these four facts are cheap to supply and expensive to discover, and the cost of discovering them late falls on people we just finished caring for, at $526 a month, in a process where a missed deadline in an unfamiliar county court is an easy mistake to make.

We built a whole discipline around the idea that handoffs fail. Then we drew its boundary at the exact moment the patient stops being one.

Evan Reid is the founder of Settled Estate, a consumer estate-settlement resource covering probate procedure in 47 states. It is a publisher, not a law firm.

By JOHN SAMARAS

Most employers still won’t cover GLP-1s for weight loss, and some of them now tell the employee where to buy them without the plan. The International Foundation of Employee Benefit Plans’ 2026 survey, published in July, found that 60 percent of employers cover the drugs for diabetes only and 36 percent for both diabetes and weight loss. Among the employers that exclude weight loss, 27 percent steer employees to a direct-to-consumer platform and 21 percent to their FSA, HSA or integrated HRA.

The same survey shows why. For its respondents, GLP-1s went from 6.9 percent of annual claims in 2023 to 11.4 percent in 2026. In a Business Group on Health survey reported by Reuters in June, 10 percent of large employers that cover them plan to drop them in 2027. Cigna stopped covering them for its own employees in July.

What the employee walks into

I record the published cash price of every US GLP-1 telehealth program every Monday, and this is the market those employees are sent to. At the maintenance dose, Novo Nordisk sells its Wegovy pen direct for $349 a month and its Wegovy pill for $299. The pill’s $149 price covers only the starter doses. The cheapest month on my chart, checked September 28, 2026, is $149 with auto-refill for compounded semaglutide, a pharmacy-made version that is not FDA-approved and is not Wegovy or Ozempic.

The tax treatment is one sentence in IRS Publication 502: you can include “amounts you pay to lose weight if it is a treatment for a specific disease diagnosed by a physician (such as obesity, hypertension, or heart disease).” Prescribed drugs qualify. The 2026 limit on a health FSA is $3,400.

Run through an FSA, a $149 month costs an employee in the 22 percent bracket about $105, because the dollars skip federal income tax and the 7.65 percent payroll tax. The employer skips its own 7.65 percent on the same dollars. The federal government gives up that revenue on a drug the plan declined to cover.

The tax break has no formulary

A covered drug passes through a formulary, a prior authorization and a pharmacy network. The cash benefit passes through none of them. Publication 502 turns on whether a drug was prescribed, and a compounded product qualifies on the same terms as a branded pen.

Screening the seller becomes the employee’s job. On September 29 the Justice Department announced the extradition of one of two men charged with selling counterfeit Ozempic to US distributors “at deeply discounted prices,” with fake boxes, inserts, labels and needles. The FDA’s own advice for buying medicine online lists “deep discounts or prices that seem too good to be true” as a warning sign.

The record goes too

On the plan, a fill leaves a claim with the drug, the dose and the date. On the cash route that record sits in a telehealth app or in the FSA administrator’s receipts, and it doesn’t reach the medical claims.

Stopping is the event a plan would most want to see. In the STEP 1 trial extension, people had regained about two-thirds of the weight they lost within a year of stopping semaglutide. Some of the prices employees budget against also carry an end date. Walgreens’ weight-management page, read September 26, lists four drugmaker offers, three of them with an end date: Foundayo pills from $149 a month, the first two months of Wegovy injections from $199, and Zepbound KwikPens from $299. All three say “Through December 31, 2026,” and the Foundayo and Zepbound offers list higher prices for the doses that follow.

Open enrollment asks employees to size their 2027 FSA in November, against prices that expire on December 31.

Three things a plan sponsor can do

Keep the record. A plan that sends employees to cash can still count the cash fill toward the deductible in exchange for the receipt, which names the drug, the dose, the date and the pharmacy.

Print the maintenance price in the benefits guide, not the starter price, so the FSA election covers the dose people stay on.

Give employees the FDA’s four signs of a legitimate online pharmacy: it requires a prescription, lists a US street address and phone number, has a licensed pharmacist to answer questions, and holds a state board of pharmacy license.

John Samaras is the founder of GLP Chart, which verifies the cash price of every US GLP-1 telehealth program every Monday.

The price data above is free to reuse under CC BY 4.0, with a link and no registration. The October release of my GLP-1 Cash Price Index, covering September, publishes today at https://glpchart.com/price-index/ with the CSV and the JSON beside it. If you want a different cut, by state, by drug or by program, ask and you will have it inside a day.

By ADAM LEE

Reflecting on a Private Equity (PE) internship in New York City reveals a major takeaway: financial engineering can be incredibly empowering, but applying it to public infrastructure often catches vulnerable communities in the crossfire.

The PE industry typically views the world through balance sheets, debt structures, cash flow optimization, and lucrative acquisitions. While these principles generate immense profit on paper, applying them to local healthcare—specifically safety-net hospitals—creates severe real-world tradeoffs.

Here are the major takeaways regarding the intersection of healthcare and business.

1. How Hospitals Get Saddled with an LBO

In the PE world, a Leveraged Buyout (LBO) is the primary acquisition method. An LBO pushes the debt onto the acquired asset, similar to convincing a bank that a newly purchased home will pay off its own mortgage.

For hospitals, this process typically follows a specific pattern:

• The PE firm borrows huge sums to purchase a hospital and its infrastructure.
• That borrowed debt lands directly on the hospital’s balance sheet.
• The firm sells the hospital’s land and buildings to a real estate company for quick cash.
• The hospital is left paying rent (with built-in escalators) just to operate inside its own building via a sale-leaseback arrangement.

The clearest example is Steward Health Care.

In 2010, Cerberus Capital Management bought the struggling Caritas Christi Catholic hospital system outside Boston and rebranded it as Steward. Six years later, Steward sold its hospital real estate to Medical Properties Trust for $1.25 billion. Cerberus exited with roughly $800 million in profit, leaving the hospitals to cover the rent, as documented by the Harvard T.H. Chan School of Public Health. Interest and rent are paid first, leaving whatever remains for patient care and equipment upgrades.

2. “Optimization” Means Cutting Unprofitable Care

When financial pressure mounts, hospitals must find money quickly through “operational efficiency.” In healthcare, this means evaluating the earnings of each service line. Emergency departments and maternity wards often lose money due to 24/7 staffing and lower Medicaid reimbursement rates, whereas specialty procedures like orthopedics and cardiac care reimburse at a premium.

Cuts are driven by margin rather than community need:

• Maternal and pediatric care: Consolidated or shut down due to limited profitability.
• ER staffing: Thinned out to save payroll, resulting in longer public wait times.
• Specialty surgery: Protected and expanded because it generates cash.

These financial sorting criteria actively impact patient outcomes. A 2023 JAMA study summarized by the NIH found that hospital-acquired conditions—driven mostly by falls and bloodstream infections—rose about 25% after private equity acquisition compared to non-acquired hospitals, despite the acquired hospitals placing fewer central lines.

3. Bankruptcies and Community Impact

PE-owned companies accounted for nearly half of the largest healthcare bankruptcies in 2025, according to the Private Equity Stakeholder Project (PESP) bankruptcy tracker.

Steward filed for Chapter 11 in May 2024. While most of its hospitals found buyers, Carney Hospital in Dorchester (serving low-income, majority Black and Hispanic neighborhoods) and Nashoba Valley Medical Center in Ayer (serving 17 towns in central Massachusetts) closed permanently due to a lack of qualified bids, creating medical deserts.

These closures disproportionately impact at-risk populations:

• Marginalized and minority communities: Residents must travel further for basic or life-saving emergency care, increasing critical ambulance transit times.
• Economic collateral damage: Hospitals are major local employers; their closures eliminate local jobs and devastate surrounding micro-economies.

The Complicating Factor

The narrative that “PE is inherently bad” overlooks critical context. Systems like Caritas Christi were already failing due to underfunded pensions, aging infrastructure, and a total lack of capital. PE firms provide capital when no one else will, keeping hospitals open for years.

The core issue is the misalignment of risk and return. In the Steward case, Cerberus profited and exited in 2016, but the 15-year rent obligation and ultimate failure in 2024 were borne by local patients and healthcare staff, not the investors who had already cashed out.

The Takeaway

These outcomes result from executing rules that reward yield optimization, not inherent cruelty. Every step of the Steward deal was legal, disclosed, and financially rational. The model works on paper but fails to price in the destruction of community assets, like local emergency rooms, which were never on the balance sheet to begin with.

This highlights a policy gap rather than a moral failing. Healthcare requires oversight as a public necessity rather than a standard asset. Potential fixes include:

• State review of hospital sale-leasebacks before closure.
• Transparent disclosure of dividends paid to PE firms out of a provider.
• Financial exposure for sponsors if a facility collapses shortly after their exit.

In January 2025, Massachusetts implemented a version of this by banning new acute care hospitals from operating on main campuses leased from REITs and adding PE deals to the state’s review process, as noted by Georgetown CHIR. Until other states implement similar policies, local neighborhoods will continue bearing the cost of Wall Street’s bottom line.

Adam Lee is an undergraduate student at UVA focusing on health policy. His newsletter is adamnalysis

By KIM BELLARD

The Blue Cross Blue Shield Association is crying wolf about AI. Its new analysis claims that “hospital systems are increasingly billing patient hospital stays as more medically complex,” and that these upcodings cost BCBS plans almost $1b from 2023 to 2025. That’s just claims from hospital systems, just to BCBS plans, so the total number could be much higher.

Well, you might say, a billion here, a billion there – that’s just healthcare these days. Everything is going up. But BCBSA is less sanguine; it thinks the coding is part of a strategy, and it thinks AI is driving that strategy.

“If patients are truly sicker, we’d expect to see more treatment,” said Luke Chalker, BCBSA’s senior vice president of product and data science. “For example, we’re seeing significantly more anemia diagnoses at these hospitals without a corresponding increase in transfusions. The disconnect between diagnoses and treatment suggests that AI is identifying more billable conditions, not sicker patients.”

The analysis indicated that most (70%) of the increased costs came from secondary diagnoses that shifted payment into higher reimbursement categories, so the hypothesis is plausible.

The American Hospital Association, of course, is having none of it. An August 2026 ”fact sheet” on AI and coding intensity called claims of AI driving coding intensity “unsubstantiated,” citing numerous factors that drive coding intensity. It blamed insurers for having AI-supported “downcoding programs,” the purpose of which are “to arbitrarily and inappropriately reduce provider reimbursement for medically necessary care that has already been delivered.”

AHA also snarkily but accurately noted how many Medicare Advantage plans have been caught upcoding members to boost their Medicare payments, so it depends on where one is sitting.

It’s also worth noting that this year the Trump Administration rolled out a Medicare program — Wasteful and Inappropriate Service Reduction (WISeR) – to use AI to help identify “fraud, waste and abuse” – a.k.a., deny claims. It has not, to say the least, gone well.

Both sides are using AI, to each other’s detriment. Which is to say, our detriment.

The BCBSA press release notes that more than 60% of hospital systems are using AI technologies, and, indeed, the 2026 Oliver Wyman Healthcare RCM Survey found that 63% of health systems are using AI in their revenue cycle management. But it’s not quite the smoking gun BCBSA would have us believe. Oliver Wyman found that “roughly 20% to 40% of organizations we surveyed report broad or enterprise-wide use of AI-enabled tools,” but this included ambient listening (33% in broad use or enterprise-wide standard), clinical decision support (38%), clinical document integrity (41%), electronic prior authorization (41%), and coding/billing automation (43%).

On the other hand, Oliver Wyman also found 70 to 90% of decision makers expected to increase AI-related RCM spending over the next three years, and you have to wonder why that number isn’t 100%.

To BCBSA’s point, the report did find that health systems “are seeing meaningful gains with targeted uses of AI improving how clinical complexity is captured, with some studies showing accuracy hit 90% or higher in specific clinical domains,” with the result that “reimbursement levels are increasing, contributing to measurable shifts in cost-of-care trends.”

Accordingly, the report suggested “payers need to accelerate investments in analytics, payment integrity capabilities, auditing and validation.” Which will drive greater investments from health systems, and so on.

And, in case you were wondering, it is not just health systems and payors using AI. The American Medical Association’s “2026 Physician Survey on Augmented Intelligence” found that more than 80% of physicians are using AI professionally, double what was reported in 2023. Like health systems, they’re using it for lots of reasons, but “diagnostic accuracy” is listed as one of the greatest expected advantages. Whether “accuracy” means more precision or upcoding depends on which side of the ledger one is on.

Yep, we’re in an AI war about healthcare claims, although Mr. Chalker says: “It’s not a war. It’s a completely one-sided blood bath.”

In a New York Times article on the BCBSA report, Reed Abelson and Teddy Rosenbluth quote Dr. Shiv Rao, CEO and founder of healthcare AI firm Abridge: “That’s the danger, where it’s bots fighting bots, agents fighting agents, a horrible dystopic future nobody wants to live in.” They also cited a recent panel with leaders from Cigna and Emory Health, during which Dr. Joon S. Lee, the chief executive of Emory Healthcare, said: “My nightmare is that we have bots on Emory Healthcare’s side, bots on Cigna’s side that keep talking to each other. A whole data center could be occupied just doing that.”

It should be everyone’s nightmare. Bots on bots doesn’t make patients better. Bots on bots doesn’t make healthcare more affordable. Bots on bots just exacerbate the problems we already have in healthcare.

Ms. Abelson and Ms. Rosenbluth point to a Health Affairs article earlier this year by Dr. David Brailer, a former Obama official who helped oversee the adaption of electronic health records (EHRs). Its title — Why AI Will Accelerate Health Care Inflation – is daunting. EHRs, he points out, led to higher hospital payments, and we should expect the same with AI. “The extraction economy of American health care does not resist useful technology,” he said. “It recruits it. AI is being recruited the same way, and AI’s scale and speed is different from anything electronic records achieved.”

 In other words, if we’re not very careful and very thoughtful, AI will make all the things we hate about our healthcare system much worse, much faster. And we don’t have much track record in being very cautious or very thoughtful.

At Health Datapalooza, Amy Gleason, the Administrator of the U.S. DOGE Service, had a great quote: “We have normalized the absurd in healthcare.” I couldn’t agree more. If AI serves to further normalize what everyone agrees are the absurdities of our healthcare system, then maybe it is time for AI to get rid of us.

I’ve long argued that all of our problems in healthcare stem from the fact that we don’t really know what “quality” is for health care. Because of that, we also don’t really know who is providing it, or who is not, either intentionally or unintentionally. Focusing AI on that would benefit both those who provide care and those who pay for it, not to mention those who receive it.

Yes, AI should help us make our healthcare system more efficient, reducing the ridiculous administrative tasks we have allowed to accumulate, but if that just results in “bots on bots” fighting over diagnosis codes, we’ll have wasted our best chance to actually fix our healthcare system.

Kim is a former emarketing exec at a major Blues plan, editor of the late & lamented Tincture.io, and now regular THCB contributor