By STEPHEN FARBER
Part one argued that affected person entry was by no means the end line; it was the inspiration. That is about what we construct now that we’re standing on it.
Each good recreation present has a second spherical, and the second spherical is at all times more durable. The primary spherical was about getting well being info out of the buildings that held it. Regulation, requirements, and a decade of unglamorous plumbing have introduced us to an essential threshold. People can more and more pull information from a number of establishments into one place. The board paid out. The viewers cheered.
Then a quieter and stranger query arrives. You’re holding years of visits, laboratory outcomes, medicines, and discharge summaries, and also you understand that having your historical past in a single place will not be the identical as understanding it. Meeting was the primary arduous drawback, nevertheless it was not the final.
The file is lastly readable
That is the place synthetic intelligence walks onto the stage, and the promise is actual. Few clinicians have time to learn a affected person’s complete historical past each time a call should be made. Appropriately safeguarded AI can summarize years of encounters, assist treatment reconciliation, clarify a laboratory pattern in unusual language, and floor a sample buried throughout a number of suppliers. Giant client platforms have already begun connecting AI assistants to medical information; Amazon’s Health AI, for instance, attracts on an individual’s One Medical historical past to supply personalised explanations and steerage. As soon as the file could be assembled, a succesful assistant can lastly make extra of it legible.
Legibility, nonetheless, will not be reliability. The farther AI strikes from explaining a file towards appearing on it, the thinner the proof turns into. A 2026 scoping review of agentic AI in healthcare discovered that many of the accessible research had been exploratory and lacked sturdy scientific validation, and that just one trial concerned sufferers. An assistant working from an incomplete, duplicated, or subtly mistaken file can produce a fluent and believable conclusion that’s nonetheless mistaken. AI doesn’t make knowledge high quality much less essential. It makes the results of poor knowledge high quality more durable to disregard.
AI can’t learn what was by no means written
There’s a extra elementary limitation, and it has much less to do with the mannequin than with the file itself. A longitudinal file can inform us what occurred, nevertheless it hardly ever tells us sufficient about why. It preserves the encounter, the prognosis, the outcome, and the prescription whereas shedding a lot of the reasoning that surrounded them. Why was one remedy chosen over one other? What was the individual hoping to protect at that time in life? Who else was serving to? What modified after the go to, and what did a member of the family discover earlier than anybody thought to schedule one other one?
Suppose a daughter has spent the previous 12 months serving to her mom handle a handful of power situations throughout three practices. She has carried out every thing the primary spherical requested of her. She has assembled the file, and she will see each go to, each outcome, and each prescription in a single place. What she can’t discover anyplace in it’s the second, someday within the spring, when she seen her mom was not fairly herself: the skipped meal, the stroll that bought shorter, the capsule organizer that stopped emptying on schedule. None of that was an occasion, and none of it generated a declare. By the point the file catches up, it is going to doc the consequence, a fall or an admission, with out the months of evolving context that may have made the consequence comprehensible and might need created an earlier alternative to reply.
These alerts should not peripheral to well being. They typically decide whether or not a plan is real looking, whether or not a change is significant, and whether or not somebody follows via, and so they clarify why a clinically full file can stay an incomplete image of an individual. AI can analyze what it could possibly attain, nevertheless it can’t get well context that no system collected, determine whose model of the story ought to govern, or decide who could act on what it finds. These questions belong to continuity, governance, and accountability relatively than to the mannequin.
There’s a additional hole that I’ll solely word right here. The daughter will not be the one one standing exterior the file. The aide who comes three mornings per week, the care supervisor on the Space Company on Growing old, the pharmacist who seen the refills slipping: most people who see the mom most frequently had been by no means on the affected person entry rails in any respect, and their programs, the place they’ve programs, are nonetheless within the first spherical. That deserves its personal remedy. For now, it is sufficient to say that the file the daughter assembled is incomplete not solely in what it captures however in who was ever capable of write to it.
The actual contest is who will get to behave
Put these developments collectively and the form of the subsequent few years comes into focus. The assembled file is now helpful sufficient that one thing will act on it. The open questions are what’s going to act on it, for whom, and beneath whose authority. The identical missed refill might immediate a reminder from a daughter, an outreach name from a care staff, an intervention from a well being plan, or a business prediction by a platform.
Worth invitations seize. The events that achieve entry to the assembled file, notably these whose enterprise fashions depend upon it, will affect what that file turns into: a useful resource that helps a person, household, and care staff make higher selections, or an enter optimized primarily for a mannequin, marketer, underwriter, or establishment. The infrastructure to assemble a file is more and more accessible. The infrastructure to make sure that the file continues to work for the individual whose life it describes will not be.
Focus additionally creates publicity, and the protections across the file can change because it strikes. The largest healthcare data breach reported in 2025 concerned Conduent, a enterprise affiliate whose back-office companies prolonged throughout healthcare organizations and authorities businesses; the reported variety of folks affected was later revised to greater than 62 million. Even when information stay technically distributed, a typical interface, credential, or middleman turns into a concentrated level of entry and subsequently a helpful goal. And when a person directs well being info to an app that’s neither a HIPAA-covered entity nor a enterprise affiliate, HHS has explained that the data the app receives is generally no longer protected by HIPAA. Different protections could apply, together with the FTC’s Health Breach Notification Rule and state client well being privateness legal guidelines, however the result’s a unique and extra fragmented authorized regime. The purpose is to not cease assembling the file however to make the tradeoffs seen, and to deal with privateness, safety, and governance as a part of the product relatively than language buried in its phrases.
What comes subsequent is continuity infrastructure
I’m not a impartial observer. HealthHive, the corporate I co-founded, is constructed round this drawback. However the architectural query exists no matter which organizations in the end reply it, and the daughter makes it concrete.
What she wants will not be a greater pile. She wants a spot the place the scientific file is preserved because the authoritative account of the care that was delivered; the place the context that solely she and her mom possess could be added with out silently rewriting the clinician’s account; and the place the precise understanding reaches the precise participant in the intervening time a call is being made. I consider this as continuity infrastructure: a ruled, longitudinal layer of understanding that sits alongside the scientific file relatively than changing it.
Two properties of that layer are particularly essential. The primary is that it distinguishes supply from interpretation. The daughter’s statement that her mom’s walks have shortened will not be a scientific discovering, and it mustn’t masquerade as one; however neither ought to it vanish as a result of it lacks a billing code. Maintaining the 2 visibly distinct, with provenance hooked up to every, is what permits a clinician to belief the file and the household to belief that their contribution was heard. The second is that permissions and adjustments are seen. The daughter’s authority to contribute and to see info ought to relaxation on an express position or authorization relatively than on the connection alone, and people roles change. A sibling joins the hassle, a house well being aide arrives, a brand new specialist enters the image, and in some unspecified time in the future, the mom herself could develop into much less capable of direct her personal care. A layer that can’t make these transitions legible will both exclude the individuals who matter or admit everybody with no accountability in any respect.
AI has an actual position inside this structure. It will probably assist the daughter put together for her mom’s appointment with out pretending she is the doctor, and it could possibly assist the clinician see what modified at residence with out asking the household to reconstruct six months from reminiscence. A corporation can use it to acknowledge threat earlier than the subsequent encounter relatively than after the subsequent declare. What it mustn’t do is silently develop into the proprietor of the story or the ultimate authority over what occurs subsequent.
A continuity layer will not be exempt from this hazard. By making info simpler to succeed in, it could possibly develop into one other concentrated and engaging level of entry. It ought to subsequently be judged not solely by what it could possibly assemble, however by what it refuses to gather, how clearly it preserves provenance, whether or not entry is seen and auditable, and whether or not permissions could be withdrawn or modified with out dismantling the individual’s historical past. In any other case, continuity turns into one other title for focus.
Why hasn’t this already occurred?
A good objection is that items of this layer exist already. Affected person portals settle for messages, patient-reported outcomes seize expertise, distant monitoring brings info from the house into scientific workflows, and file aggregators can add context. Maybe the incumbents merely take in this perform, governance and all.
That chance can’t be dominated out. However these instruments are typically ruled by the establishment, program, or vendor that deployed them. The knowledge they acquire doesn’t reliably journey with the individual once they change docs, transfer between plans, or enter a brand new stage of life. Nor had been most designed to protect the household’s account alongside institutional information with out complicated the 2.
Context could be collected right this moment. The take a look at will not be whether or not an incumbent can add these options. It’s whether or not continuity survives when the incumbent’s relationship with the individual ends. An establishment can sponsor the layer, but when altering docs, plans, employers, or distributors additionally adjustments who controls the amassed story, the system will not be longitudinal in the best way that issues. A continuity layer earns the title infrastructure provided that id, permissions, provenance, and portability persist independently of any single sponsor.
Who pays will not be who governs
The individual whose life the file describes won’t at all times be the paying buyer, nor should they be. Well being programs, well being plans, employers, neighborhood organizations, and different establishments could have sound causes to fund continuity as a result of they profit when folks and the folks supporting them make higher selections. Fee and governance should not the identical factor. Sponsoring the infrastructure mustn’t confer the precise to acceptable the individual’s story or quietly optimize it for a goal the individual doesn’t perceive. The financial cut price must be express: the sponsor is buying an outlined service and a measurable profit, and the individual’s longitudinal story will not be a part of the acquisition.
That boundary belongs within the product’s permissions, contracts, and technical structure. If it relies upon solely on the sponsor’s goodwill, “person-centered” is branding relatively than governance.
The decisive query won’t be which system is aware of probably the most. Will probably be which system stays answerable to the individual after it is aware of.
The second spherical
Affected person entry created the potential for one thing healthcare has by no means had: an understanding of the individual that can persist throughout encounters, establishments, and altering relationships with out belonging completely to any certainly one of them. Half one referred to as affected person entry a basis relatively than a end line, and a basis is an invite to construct, nevertheless it doesn’t decide what goes on prime. We are able to construct clever instruments that assist people and the folks they belief perceive and steward well being over a lifetime. Or we are able to permit the pile we labored so arduous to assemble to develop into probably the most environment friendly enter ever created for programs that had been by no means designed to serve the individual.
The second spherical is subtler than the primary as a result of the Whammy not seems like shedding your knowledge. It seems like profitable it after which handing it to whoever asks first.
Large bucks. No Whammies. This time, watch the board extra fastidiously.
Stephen Farber is Co-Founder and CEO of HealthHive, and the Vice Chair on the Board of Administrators of the American Society on Aging.
