The Record That Answers Back

A vision for the next phase of healthcare technology.

When Iron Man reached theaters in 2008, the memorable spectacle was a flying suit of armor. The durable idea was Jarvis, the intelligence Tony Stark simply talked to, a presence in his environment that could pull any file, run any analysis, and carry out any instruction while its user kept his attention on the work in front of him. It played as disposable science fiction at the time. Nearly two decades later it reads instead like an early sketch of what clinical software is becoming, and a measure of how far the software running in most clinics still is from it.

The electronic health record digitized the paper chart and preserved its passivity in the process. It stores everything and understands nothing, which means the intelligence in the room has remained entirely human. The clinician serves as the system’s search engine, locating, assembling, and interpreting whatever the database holds but cannot surface, and the cost of that arrangement now appears reliably in the burnout literature. These systems were shaped as much by billing requirements and regulatory incentive as by clinical need, so their inability to think was never an engineering accident. Thinking was simply not what they were built to do.

Some of that cost turns out to be recoverable, and recently it was measured. When software that listens to the visit and drafts the note was studied across five academic medical centers, it returned roughly thirteen minutes of record time to each clinician per day, with measurable improvements in well-being behind it, findings that ran in JAMA last year. Thirteen minutes is the yield of automating one task inside a system that otherwise went unchanged. The more interesting question is what becomes possible when the system itself is redesigned, and over the past two years, piece by piece, that question has been getting answered.

What changed

Start with what a clinician can now do that was impossible five years ago. Preparing for a complex admission, a physician can ask the chart directly whether the patient has ever reacted badly to contrast, when the last colonoscopy happened, how kidney function has trended across five years, and get an answer drawn from the patient’s actual record rather than from a model’s general knowledge. Work that used to mean twenty minutes of digging through scanned documents and buried flowsheets becomes a question, asked and answered. This is not a concept video. It has been running in the hands of practicing clinicians at Stanford’s medical center, built directly into Epic under the name ChatEHR, which is precisely what makes it hard to dismiss.

The reason this matters goes deeper than the minutes it saves. Medicine has always run on language. The case presentation, the handoff between colleagues, the consult note, the history taken at a bedside are all clinical work performed in speech and prose, refined over generations into forms precise enough to carry a life’s worth of information between two people in ninety seconds. Software never had a good reason to abandon that medium. It abandoned it because early databases could only be operated through forms and fields, and two decades of habit hardened the workaround into something that felt like a law of nature. What the Stanford work quietly demonstrates is that the workaround can end, and that the record can meet clinicians in the medium their profession already mastered.

But give clinicians the ability to ask the record anything and you discover, almost immediately, what the record actually is. The answers come back incomplete, and not because the interface failed. They come back incomplete because the thing being asked was never whole to begin with. A patient’s story is written by many hands, the physician’s notes in one module, nursing assessments in another, physical therapy evaluations, occupational therapy goals, and speech pathology reports scattered across systems that were never designed to read one another. Every discipline documents faithfully, and the sum still fails to add up to a person. The conversational interface does not create this fragmentation. It exposes it, the way a good question exposes a witness who has only rehearsed part of the story.

So the next problem is assembly. The corrective idea is easy to state and genuinely hard to build, the patient’s history treated as one continuous object, a single longitudinal fabric that the entire care team writes into and reasons from, where the nurse’s observation and the therapist’s assessment sit in the same weave as the physician’s diagnosis rather than in parallel archives. The principle underneath it deserves to be said plainly. A machine cannot reason about a life that has not been assembled. That idea moved from principle to practice last year, with federal money behind it, when an ARPA-H funded effort at the University of Illinois Chicago put clinicians and technologists in front of real longitudinal records spanning all of those disciplines at once and had them build on the record as a whole.

Yet assemble the record completely and something curious happens. The questions clinicians most want to ask of it change character. A whole record can say with perfect fidelity what has happened, every value, every visit, every intervention, and the questions that fill a clinical day are not about what happened. Will this patient be readmitted within the month. Will this kidney keep declining. Is this chest pain the one that ends in a catheterization lab. The more faithful the memory becomes, the more clearly it reveals its own limit, which is that memory faces backward while medicine faces forward. And for the entire history of the profession, that forward-facing question has been answered by individual judgment, meaning by whatever pattern a given clinician happened to have personally witnessed, a sample of a few thousand patients across even a long career.

What changed this past year is the size of the sample. With the de-identified journeys of hundreds of millions of patients gathered in one place, a physician can see how people resembling the one in front of them actually fared, and a model trained on more than one hundred billion patient events can estimate the probability of what has not happened yet, the readmission, the new diagnosis, the deterioration that tomorrow’s labs will only confirm. This is Epic’s own work, in Cosmos and now Comet, and whatever one thinks of any particular vendor, the substance stands on its own. The accumulated experience of medicine can now be aimed forward at the trajectory of a single patient, which means the record has outgrown its oldest job. An archive reports what happened. Medicine has always needed an instrument that helps decide what to do.

A prediction on its own, though, is not yet help. Delivered into today’s workflow it becomes one more alert, and medicine already knows what a system that only warns does to the people working inside it. Every forecast turns into another task on a human list, and the clinician goes back to serving the machine, this time as the hands for its conclusions. For anticipation to be worth anything, the system has to be able to act on it, to schedule the follow-up, adjust the order, route the referral, close the loop it opened. And the moment software begins to act on patients rather than merely advise about them, trust becomes the whole question, because the obvious way to build such a system turns out to be the wrong way. A single generalist model asked to retrieve evidence, reason over it, draft the note, place the orders, and check its own work becomes hard to trust and impossible to audit, a black box precisely where medicine most needs glass. The emerging answer distributes the work instead, one agent retrieving, one reasoning, one documenting, one executing, with an orchestrating layer holding them to a shared plan and a record of who did what. There is nothing exotic in this design, and that is its strength. It is how hospitals already produce reliability. No floor runs on a single infallible mind. Every floor runs on structured cooperation among fallible ones, with roles, checks, and accountability, and the formal case for building clinical software the same way was made last year in Nature Biomedical Engineering by Eric Topol and Pranav Rajpurkar, under the name multi-agent systems for healthcare.

Step back far enough and these four developments stop looking like separate findings. Each one begins where the previous one runs out. Conversation exposes fragmentation, assembly produces a memory, memory demands foresight, and foresight demands a trustworthy way to act. A system that converses in ordinary language, rests on one continuous record, anticipates rather than merely archives, and acts through coordinated agents whose work can be audited is not four systems. It is one system, described from four directions by groups that had no obligation to agree with one another. Each piece has now been demonstrated somewhere. What remains unsettled is how the whole should be built, and on that question the market has mostly taken the easier of two available paths.

The limits of attachment

The commercial response so far has been a strategy of attachment. Scribes are appended to existing records, chat assistants are layered over existing databases, analytics modules are sold beside systems whose internal structure has not changed in a decade. The logic is understandable. The installed base of legacy systems is enormous, hospitals are rightly conservative about their most critical infrastructure, and the relief these tools provide is real, as the scribe numbers show.

The limitation is structural, and it appears the moment the attached intelligence is asked to do more than transcribe. Set a reasoning system on top of data that was never organized to carry clinical meaning and every answer becomes an act of reconstruction. The model must guess which of four medication lists is authoritative, which problem entry went stale years ago, which scanned document holds the operative detail. Its retrieval turns probabilistic exactly where medicine requires reliability, and when a clinician asks why the system said what it said, the trail of evidence is nowhere to be found. However fluent the assistant sounds, it is working inside a structure that was never built for it, and no amount of fluency substitutes for a foundation.

There is a precedent for this dilemma, and it happens to be visible in orbit right now. The International Space Station was assembled module by module over twenty-five years, by many nations, each new piece docking onto whatever was already there, and its interior wears that history openly, cable bundles strapped along every surface, equipment wedged into corners, workarounds layered on workarounds until the workaround became the architecture. China’s Tiangong station flew two decades later and looks nothing like it. The wiring runs behind the panels and the modules read as one design, not because its engineers were better, but because they arrived later, could study everything the first station had learned the hard way, and were free to design the whole before building the parts. The first station proved the thing could be done at all, which is the honor of pioneers. The second shows what those lessons are worth to whoever can start from a clean sheet. Clinical software has reached the same juncture. The incumbent platforms carry their history the way the older station carries its cabling, decades of billing logic, regulatory retrofits, and departmental silos accreted module by module, every new capability docking onto structure that predates it. The weight is not a failing. It is simply weight, and there are moments in the life of a technology when the best way to honor what the pioneers built is to learn everything it has to teach and then begin again clean.

Beginning clean means inverting the order of construction. The data itself is organized for machine reasoning first, with conditions, medications, results, and histories structured so that an agent can traverse them deterministically and trace every answer to its source. Only then are conversation, documentation, orders, and analytics built on top, all drawing from that single substrate. A system built in this order has no chat feature in any meaningful sense, because conversation is simply how the system is used, the way Jarvis was never an application that had to be opened. Whether the established platforms can rebuild their foundations beneath decades of accumulated structure, or whether that inversion belongs to systems designed for it from the start, is now the central architectural question in clinical software, and it will not be settled by marketing language.

When the record reasons

Follow a single clinical day and the difference becomes concrete. Preparation for a visit starts from a synthesized chart, the relevant history already assembled and the open questions already flagged, so the clinician reads instead of excavating. Documentation happens while doctor and patient speak, because the conversation itself is the input. Orders, referrals, and follow-up arrangements execute during the encounter as coordinated tasks rather than accumulating into the administrative debt currently paid down at kitchen tables near midnight.

The other half of the encounter has been waiting even longer. A patient is the only person present at every appointment of their own life, the one continuous thread running through every hospital, specialist, pharmacy, and lab they will ever touch, and yet the systems built around them have treated the patient as a visitor to their own story, granted a portal that works like a lobby, a place to view fragments and leave messages into silence. The research on what happens when that changes is older than the current wave and just as clear. When patients were simply allowed to read their clinicians’ notes, in the open notes studies that began over a decade ago, they understood their care better, remembered their plans, and took their medications more reliably. Understanding, it turns out, is itself therapeutic. Regulation has since conceded the principle, writing the patient’s right to their own data into American law and prohibiting the blocking of it, but a right to download files is not yet a record that travels. In a system designed for reasoning, portability stops being a compliance checkbox. The record moves when the person moves, answers them in their own language at whatever depth they need, and holds the thread on the three hundred and sixty days a year when no clinician is in the room, so that the connection between patient and doctor becomes a continuous channel through one shared, living record rather than a message left in a box.

Beyond any single person, clinician or patient, lies a question most record systems cannot answer at all. Anyone who cares for a panel has wanted to know what is trending across it, which patients are quietly drifting toward trouble, what connects the last thirty presentations of a symptom in a particular season or zip code. The information exists, distributed across individual charts, but the systems holding it were built to describe patients one at a time, and population insight was exiled to reporting modules and quality dashboards far from the point of care. A record designed for reasoning removes that distance. When asking is the interface, the panel becomes as easy to interrogate as the chart, and the epidemiology of a clinician’s own practice, always present in the data and always absent from view, becomes part of ordinary clinical awareness.

This capacity matters most where resources are thinnest. The distance between well-funded and under-resourced medicine, whether between countries or between a flagship academic center and a rural practice, has always included an information gap layered on top of the material one, and of the two, information is far cheaper to close. A clinic that cannot afford another physician or another imaging suite can still afford to know more about its patients, its patterns, and its risks. A record that reasons is leverage, and the practices that stand to gain the most from it are precisely the ones furthest from the institutions where new technology gets piloted first.

The vocabulary for all of this remains unsettled, and each candidate term names a layer rather than the whole. Agentic describes the execution, conversational describes the interface, AI-native describes a build philosophy. That no single word has stuck is itself evidence that the thing is new, and history suggests the name will arrive only after the artifact makes it unavoidable, the way the smartphone was named by the device that made every previous phone feel insufficient. What can be said now is that the research phase of this transition is over and the construction phase has begun. A conversing, reasoning, task-executing intelligence was a special effect in 2008 and is a buildable production system in 2026. The standard by which the coming systems will be judged is simple. A clinician who has worked with a record that answers back will not volunteer to return to one that does not.