We let billing define the medical record. AI gives us a chance to take it back.

A recent essay in JAMA by Benjamin Chin-Yee, MD, mourns the death of the dictated consult note; think the loss of narrative and the fading of clinical reasoning as a practiced art. It’s an elegant piece, and the instinct for grief is understandable. There was something real in the discipline that dictation required: the pause, the synthesis, the moment of committing to what you actually thought was going on with a patient. 

But before we reach for the eulogy, it’s worth asking whether the real loss happened much earlier, not when ambient AI arrived, but when we had a genuinely revolutionary idea about clinical documentation and quietly built something else instead. 

This is a story about a vision, a misreading, and a decision point we’re arriving at again. 

What Weed actually intended 

Larry Weed, MD, is remembered today as the father of the SOAP note, a reductive epitaph for a genuinely radical thinker. What Weed actually proposed, beginning with his landmark 1968 paper in the New England Journal of Medicine, was an evidence-based argument: the individual physician’s memory and judgment, operating alone, is an insufficient foundation for complex medical decision-making. This wasn’t an insult to clinicians. It was a systems argument, the same one that led aviation to embrace checklists without concluding that pilots were incompetent. 

His Problem-Oriented Medical Record (POMR) was designed to offload knowledge retrieval to structured external tools, and later, to what he called “knowledge couplers,” software that would match patient data against the medical literature systematically so that clinicians could focus on what only humans do well: synthesis, judgment, and the patient in front of them. Highly standardized at the front end, he argued, so that the back end could be genuinely individualized. The structure existed to make reasoning explicit and auditable, not to replace it. 

Weed was not anti-cognition. He was anti-unaided cognition. He famously didn’t want to see a clinician’s final conclusion; he wanted the data that produced it. In a 2010 interview, he put it plainly: “I need a problem list. Not the doctor’s impression after he did the initial workup. I’m not interested in somebody’s impression.” The distinction between those two positions is everything, because what the healthcare industry built in the decades that followed ignored it entirely. 

What we built instead 

The EHR era borrowed Weed’s vocabulary (e.g., structured data, problem lists, standardization) and used it to construct something he wouldn’t have recognized as his vision. Billing requirements colonized the record. The SOAP note, designed to make clinical reasoning explicit, became a template engineered to make billing defensible. The problem list, intended to reflect what a clinician understood and didn’t yet understand about a patient, became an auto-populated registry of every diagnosis code ever attached to a chart. 

This is the specific betrayal worth naming: a framework designed to impose discipline on clinical reasoning was repurposed to impose compliance on clinical documentation. Those are not the same activity, and conflating them has had consequences. I’ve written recently about the downstream result, what Bryan Vartabedian, MD aptly named medical slop: documentation that is technically complete, clinically thin, and produced by good physicians acting rationally within a broken system. The slop problem is real. But the origin story matters too, because you cannot fix a design failure without understanding what was designed, and why. 

What we designed, over thirty years and largely by default, was a record optimized for billing fidelity, legal defensibility, and regulatory compliance. We did not design it for clinical reasoning. Nobody held a meeting and made that choice. It accumulated — through auditor feedback, through EHR defaults, through the rational adaptations of overextended clinicians — until the record became what it is today: a document dense with information and thin on insight. 

The friction we eliminated without asking what it was doing 

Chin-Yee’s observation about dictation is right, but the narrative note had real problems. It was idiosyncratic, inaccessible, and dependent on individual habits and verbal fluency. It was not a golden age, but it had one feature the templated note lacks almost entirely: it required the clinician to synthesize. Dictation enforced a pause. A moment of genuine commitment to what you thought was happening with this patient, spoken aloud in your own words. 

Weed wanted that synthesis supported by better tools. What we delivered was synthesis replaced by structure. The dot phrase, the copy-forward exam, the auto-populated review of systems. These do not assist cognition; they simulate its output. The note looks like thinking happened. It increasingly does not require that it did. 

This is not primarily a documentation quality complaint, though it is that too. It is a design values problem. We never explicitly decided what the clinical record was for. We let billing requirements, legal departments, and EHR configuration defaults answer that question for us, over decades, by accumulation. The record became a billing instrument because billing drove the incentives. It was not inevitable. It was a choice made incrementally, by omission, without anyone quite noticing what was being given up. 

The decision point we’re arriving at. Again. 

Ambient AI scribes are being deployed at scale right now. Health system leaders are making purchasing and implementation decisions often based on one metric: clinician time saved. That is a legitimate goal: documentation burden is real, burnout is real, and returning meaningful time to clinicians matters. But the time-savings frame is incomplete in a way that should concern anyone paying attention to the story above. 

For the first time in the history of electronic documentation, ambient AI creates a structural possibility that hasn’t existed before: the administrative record and the clinical record can be separated. The machine can own the billing note, capturing the elements required for the E&M code, the medication reconciliation, the review of systems. Fine. Let it.  

But that separation only creates value if someone explicitly decides what the clinical note should do in its absence. What should the Assessment and Plan communicate, and to whom? What does the covering hospitalist at 2 AM actually need? What would Weed’s vision of explicit, auditable clinical reasoning look like in a modern record, freed from the obligation to simultaneously justify a billing level? 

Most health systems are not asking these questions before signing the ambient AI contract. They are letting the vendor’s output template answer it for them, which I would argue is precisely how we arrived at the current state of documentation three decades ago. 

My call to action is direct: before your next ambient AI implementation, convene the clinical leaders who should own this question and ask it explicitly. Not “what does the note need to contain?” That question has been answered, badly, by a thousand auditors and a thousand compliance checklists. The question is: what is the clinical record for?  

The answer will shape documentation governance for the next decade. Weed spent the final decades of his life arguing that medicine was in denial, not about disease, but about its own cognitive limits, and about its failure to build systems that honestly accounted for them. The ambient AI moment is a genuine opportunity to take that argument seriously at last. 

It would be a shame to let a vendor’s output template make the decision for us. Again.