Start Where Doctors Think
The pre-charting advantage, and why it beats scribes on the path to clinical reasoning
Recently, we were talking to Dominik, an investor at EF, who asked us:
“Both you and scribe companies are building toward AI-assisted clinical decision-making. You are just starting from different places. Why does your starting point give you an advantage? How are you better positioned?”
We fumbled the answer. We talked around it, but we did not nail it. So I decided to write this down properly, to think through the question carefully and articulate the answer with the clarity it deserves.
But as I started writing, I realized the answer is bigger than I initially thought. The question is not just about pre-charting versus scribes. It is about what becomes possible if clinical reasoning can be captured computationally, and why our starting point may be better positioned for that capture.
Let me work through this from the beginning.
What pre-charting actually is
Most people outside healthcare do not realize that a large share of a medical note’s content is determined before the doctor sees the patient.
A typical outpatient note includes the history of present illness, past medical history, medications, allergies, social history, family history, review of systems, vital signs, physical exam findings, relevant lab and imaging results, assessment, and plan. Of these, only a few are generated during the visit itself: today’s vital signs, what the patient reports about their current symptoms, the physical exam findings, and the doctor’s assessment and plan. Everything else comes from the existing medical record, and the doctor’s job before walking into the room is to pull all of this together, understand what is going on with this patient, and prepare the note accordingly.
This preparation is called pre-charting, and it is the foundation of the entire encounter.
Let me make this concrete with a simple example. A 67-year-old man with diabetes, hypertension, and chronic kidney disease is coming in for a follow-up. His chart contains three years of visit notes from multiple providers, dozens of lab results from different dates, cardiology and nephrology consult notes, multiple echocardiogram reports, a hospitalization discharge summary from eight months ago, and an emergency room visit note from two months ago.
Before the doctor sees this patient, someone needs to figure out what medications he is actually taking, which is harder than it sounds. The cardiologist started him on empagliflozin six months ago, but to know that, you have to open the cardiology consult note (buried somewhere in a 412 page faxed document), scroll past the chief complaint and history sections, and find this on third page of that five-page note where it says “I am initiating empagliflozin 10mg daily for HFpEF.” The nephrologist reduced his metformin dose three months ago, but that detail is a single line in a dense eight-page note: “Given eGFR decline to 38, I recommend reducing metformin to 500mg twice daily and avoiding NSAIDs.” During his hospitalization, he was on a higher dose of furosemide for fluid overload, but the discharge summary came in as a part of the 412-page fax, and so on. The doctor needs to piece together the current medication list from all of these sources and understand why each change was made.
They need to compile the relevant lab trends, his hemoglobin A1C trajectory showing improving diabetes control, his creatinine trajectory showing declining kidney function, his borderline elevated potassium that matters because he is on an ACE inhibitor and has compromised kidneys. These lab values are not displayed together anywhere, so the doctor has to open the lab portal, scroll through dozens of results, mentally pick out the relevant ones, and track the trajectory in their head. Some labs came from the hospital system, some from a commercial lab, some from the specialist’s office, and they are all in different formats in different places.
They need to know the patient’s social history, which for this patient is not just background information but directly relevant to his clinical management. The hospitalization eight months ago was triggered by dietary indiscretion, he ate a lot of salty food at a family event and retained fluid. So understanding his current dietary situation matters: is he adhering to the low-sodium diet, does he have someone at home who helps with meal preparation, or does he live alone and rely on convenience foods? But to find this, you have to dig through a social work consult note buried in the hospitalization paperwork, where it mentions he lives alone and his daughter visits twice a week to help with groceries. His smoking status is documented as “former smoker, quit 2019” in an intake form from three years ago, but there is no recent confirmation, and the doctor needs to know whether that is still accurate because smoking accelerates both his cardiovascular disease and kidney disease.
They need to confirm his allergies and medication intolerances, because this patient is on a complex regimen and any adjustments require knowing what he cannot take. The allergy list in the EHR says “penicillin - rash,” but the faxed hospital records list “penicillin, sulfa, lisinopril (cough).” The lisinopril cough explains why he was switched to his current ACE inhibitor years ago, and the sulfa allergy matters because if his furosemide needs to be changed, certain alternatives would be off the table. The doctor needs to reconcile these discrepancies before making any prescribing decisions.
They need to understand what the specialists have said, reconcile conflicting recommendations, know about the hospitalization eight months ago and what triggered it, and identify care gaps like overdue screening tests.
All of this gets synthesized into a coherent picture of what is happening with this patient, and the note is already half-written before the patient walks through the door.
High-quality encounters depend on this preparatory synthesis.
What happens during the visit, and what scribes actually see
The doctor enters the room with the pre-charted note in front of them. They already know this patient’s story, the lab trends, what the cardiologist and nephrologist have said, and they already have a mental model of what to focus on today.
The visit itself is about confirmation, updating, and decision-making. The doctor confirms the medication list with the patient, asks about symptoms, does a physical exam, synthesizes everything together with what they already prepared, makes a plan, and explains it to the patient. The visit adds today’s vital signs, the physical exam findings, the patient’s symptom report, and the assessment and plan. Everything else was already prepared.
Now think about what a scribe sees. A scribe sits in the room or listens to a recording of the visit. They hear the doctor ask about symptoms, the patient’s answers, the doctor describing physical exam findings, and the doctor explaining the assessment and plan. The scribe writes this down.
Even when scribes have EHR access, and some do, the feedback they receive is concentrated in post-visit edits to drafted prose. Did I say metformin or metoprolol? Did I say 150mg or 50mg? Should the physical exam be formatted differently? Was the assessment section too verbose? The supervision is about transcription and style, not about which pieces of raw patient data should have been surfaced or excluded.
The scribe does not see the three years of prior notes being filtered for relevance, the lab trends being extracted and interpreted, the cardiology and nephrology consults being reconciled, or the 43-page faxed discharge summary being distilled to its essential points.
By the time the doctor walks in, the hard cognitive work is finished, and the scribe documents the output of that work rather than the work itself.
The two types of cognitive work
The way I think about this is that clinical medicine involves two fundamentally different kinds of thinking.
The first is synthesis, where you start with a mess of fragmented information scattered across multiple sources in inconsistent formats (faxes, scanned documents, typed notes, lab results from different portals, imaging reports, medication records) and your job is to figure out what is relevant, extract it, understand how the pieces connect, and construct a coherent picture of what is happening with this patient. This is the hard part of medicine, this is where expertise shows up, and this is where things get missed.
The second is the visit itself, where you talk to the patient, examine them, refine your understanding based on what they tell you and what you observe, make a plan, and explain it to them. This obviously requires skill, and I do not want to oversimplify, doctors do think during the visit too, they notice things during the physical exam, and they hear things from the patient that shift their hypothesis. The split between “synthesis before” and “communication during” is not perfectly clean, and real clinical work is messier than any neat two-phase model.
But here is what matters for our purposes: even when doctors think during the visit, they do not verbalize everything. A doctor managing a complicated multi-system patient does not turn to the patient and say: “I am weighing the cardiology recommendation against the nephrology recommendation and deciding to prioritize kidney function because the creatinine trend concerns me more than the ejection fraction trend.” They just make the call, and the scribe hears the conclusion while the reasoning stays inside the doctor’s head.
So when I say pre-charting captures synthesis and the visit captures conclusions, what I mean is this: pre-charting is the moment when doctors reveal what they think matters through their actions rather than their words, by what they keep, delete, add, and emphasize, while the visit is where they announce what they decided. Scribes capture the announcement, while pre-charting captures the thinking that led to it.
If the goal is to build AI that can think like a doctor, you need training data on reasoning, because knowing how to transcribe conclusions does not teach you how to arrive at them.
Clinical reasoning as tacit knowledge
This brings me to what I think is the core point.
Medicine today has two kinds of knowledge. There is formal knowledge: basic science, clinical trials, published research, guidelines. This is rigorous, peer-reviewed, the gold standard. But it has huge gaps. Trials are expensive, so only commercially viable questions get studied. Trials take years, so knowledge always lags reality. Trials enroll narrow populations, so they do not represent real patients. Then there is the informal knowledge: what doctors learn through years of practice that never gets written down anywhere.
Think about a cardiologist who has seen ten thousand patients with chest pain. They have built pattern recognition that exists only in their head. They know which findings to worry about and which to dismiss, they know that a certain combination of risk factors in a certain type of patient warrants concern even when each individual factor looks borderline, and they have seen a subtle change in an EKG waveform that turned out to be the early sign of something serious, so now they notice that pattern when others miss it. They cannot fully articulate all of this knowledge, if you asked them to write it all down, they could not. It is tacit, baked into how they see patients.
This tacit knowledge is not captured anywhere in the healthcare system today. It is not in EHRs, because EHRs store data rather than capturing reasoning. It is not in medical literature either; textbooks like Harrison’s contain deep clinical reasoning, they teach you how to think about diseases, how to approach diagnostic problems, how to weigh different treatment options; but what they cannot capture is the application of that reasoning to real patients in real clinical practice. Harrison’s can teach you everything about managing diabetes, everything about managing chronic kidney disease, everything about managing heart failure, but when you have a specific 67-year-old sitting in front of you with all three conditions plus a hospitalization eight months ago plus conflicting specialist recommendations plus borderline potassium plus a social situation where he lives alone and struggles with dietary adherence, how do you actually apply all of that textbook knowledge? What do you prioritize? What do you weigh more heavily? What do you surface and what do you set aside? That application is what gets built through years of practice, and that application is what lives in doctors’ heads. It is not in clinical trial data either, because trials tell you what works on average for patients who met the inclusion criteria, which is not the same as knowing how an experienced doctor would apply that evidence to your specific patient with all their complexity.
The way to capture tacit knowledge is to watch what doctors actually do when they make decisions, what they pay attention to, what they ignore, what they emphasize, what they de-emphasize, and then correlate those patterns with what actually happens to patients over time.
Pre-charting is the moment where tacit knowledge gets revealed through action, and the learning signal comes from two places. First, we have access to the raw source data, the 412-page faxes, the scattered lab results across different portals, the buried consult notes, the discharge summaries, all of it; and we learn what is relevant from this vast mess of fragmented information. Scribes never see any of this raw data, so they cannot learn the mapping from chaotic inputs to relevant outputs. Second, when a doctor edits our pre-charted note; adds something we missed, deletes something we included that does not matter, moves something to be more prominent, they are showing us what they think matters without explaining it.
So between the raw-data-to-relevance mapping and the edit signal on top of it, we are capturing a training signal for clinical reasoning that is not being captured through the scribe workflow.
The difference in training signal
Let me be specific about what this means in terms of what each system actually learns from.
For scribes, the training signal comes from doctor edits to transcribed notes. The doctor reviews what the scribe wrote and makes corrections that tell us “I said metoprolol, not metformin” or “put the physical exam in a different order” or “I prefer bullet points for the plan” or “add that I discussed diet with the patient.” These are corrections about transcription accuracy and formatting preferences, and they teach the system to be a better transcriptionist rather than teaching clinical reasoning.
For pre-charting, the training signal also comes from doctor edits, but the nature of those edits is completely different. The system presents a synthesis of the patient’s record, and the doctor modifies it in ways like “delete the dermatology consult from two years ago, it is irrelevant” or “add the potassium trend, it matters because he is on an ACE inhibitor with declining kidney function” or “move the procedure summary to the HPI section, that is the most important context” or “you missed that the nephrologist said to avoid NSAIDs, it was buried in that long note but it matters.”
These are corrections about clinical relevance that teach the system what matters and what does not, how to weigh information, and they transfer the pattern recognition that experienced doctors have built over years, the tacit knowledge that is not recorded anywhere else.
The scribe hears the visit. The doctor says: “Your EKG looks fine. Your troponin is negative. I think this is musculoskeletal pain from when you helped your friend move furniture last weekend. Take ibuprofen and follow up if it gets worse.” The scribe transcribes this. From the scribe’s perspective, the job is done: capture what was said, format it correctly.
We see something different. Before the visit, we pulled together the patient’s record: a stress test from two years ago, a history of acid reflux from five years ago, a triage note from three days ago where the patient mentioned helping a friend move, cardiac risk factors scattered across intake forms. We presented this to the doctor. The doctor deleted the acid reflux history. The doctor kept the stress test. The doctor highlighted the triage note about furniture moving, which we had buried. The doctor reorganized the cardiac risk factors to be easier to scan.
Each of those edits teaches us something. Deleting the acid reflux tells us it was not relevant to this presentation. Keeping the stress test tells us prior cardiac workup matters for chest pain. Surfacing the furniture note tells us that recent physical activity context can be diagnostic. The scribe learns how to transcribe conclusions. We learn how doctors decide what matters.
Every one of these edits teaches us something about evaluating chest pain: what history matters, what testing is relevant, what contextual details change the clinical picture, and how to think about risk. This is the training signal for clinical reasoning, and the scribe workflow does not capture any of it.
How the learning compounds differently
Scribe learning compounds toward better documentation, where each visit processed improves transcription accuracy, formatting consistency, and ability to match doctor preferences for note style. This is useful, but the ceiling is lower because the variation in documentation style is narrower than the variation in clinical reasoning. A scribe that has processed ten million visits is only slightly better than one that has processed one million, and the marginal improvement is small. The ceiling is being a very good transcriptionist.
Pre-charting learning compounds toward clinical reasoning, where each patient processed and each doctor edit received teaches patterns of relevance. These patterns are specialty-specific, condition-specific, and context-specific, and there is no ceiling because medicine is vast. Each new pattern learned makes the system meaningfully more capable, and the learning does not asymptote.
But edits are not the only training signal, the second signal is outcomes. Pre-charting builds a longitudinal patient record, which means we see what information was surfaced, what decisions were made, and then what happened to the patient over the following months and years. This lets us correlate patterns with real-world outcomes, learning that patients with this lab trajectory who were managed this way had better outcomes than patients who were managed that way, that this pattern that doctors often overlook actually predicts hospitalization six months later, and that when this information was surfaced and acted on, patients did better than when it was missed.
Edits bootstrap the learning, and outcomes compound it. By the time the system is good enough that doctors barely need to edit, the outcome signal will have taken over.
The honest competitive picture
But let me step back and be honest about something, because there is a counterargument worth taking seriously.
The market is not dumb. Scribes won the first wave for real reasons. The pain point was obvious and immediate: doctors hate documentation. The technology was ready: speech-to-text improved dramatically over the past few years. The value proposition was simple to measure: time saved on notes. It did not require deep EHR integration to get started. It was easy to sell: “we will write your notes for you.”
So scribes are not a mistake. They picked a real problem with a clear solution that was technically feasible and easy to sell. The market correctly identified documentation burden as a pain point and built products to address it.
But does that mean the first wave of products in a space is always best positioned for where the market ultimately goes?
I do not think so. Consider what happened with mobile devices. Blackberry dominated the early smartphone market because they solved the obvious problem with the technology that was available at the time: business people needed email on the go, and Blackberry delivered that with physical keyboards and push email infrastructure that worked reliably. They were not wrong. They built exactly what the market needed, and they won. But when the underlying technology shifted, when touchscreens and mobile processors and app ecosystems became viable, the end state turned out not to be “better mobile email” but “mobile computing.” Blackberry had built everything around the first problem, and a different starting point turned out to be better positioned for where things actually went. The first wave solved the obvious problem brilliantly, but the first wave was not the end state.
I think something similar may be happening in clinical AI. Scribes solved the obvious problem: doctors hate writing notes, so let AI write the notes. That was the right product to build when speech-to-text matured. Both scribes and pre-charting companies now claim to be building toward AI that can help doctors think. But does starting from transcription versus starting from synthesis lead to the same place? I think you’ve understood this by now.
There is also a ground truth that does not match the hype around scribes. In roughly one hundred qualitative interviews we conducted with US doctors, only five were actively using ambient scribes, and about forty reported trying them and dropping them. The complaints were consistent: too verbose, wrong details, wrong formatting, bad at figuring out who is speaking when multiple people are in the room, and most importantly, terrible at the assessment and plan section. Even doctors who stick with scribes report spending significant time after visits correcting the output, which undercuts the time-saving value proposition. And there is a subtler problem: doctors told us they find themselves speaking differently during visits, in a more robotic and performative way, so the scribe picks up everything correctly. They are optimizing their communication for the transcription tool rather than for the patient. That is not how doctor-patient conversation worked before scribes, and it is not clear that is a good trade. Scribes are not as entrenched as the funding headlines would have you believe.
The other thing worth considering is that pre-charting was not really buildable until recently. Processing long faxed documents, extracting structured information from heterogeneous sources, synthesizing across scattered data: this required advances in AI that are only now becoming possible. But it was not just AI. The regulatory environment also changed. The 21st Century Cures Act led to the CMS Interoperability and Patient Access final rule in 2020, which mandated FHIR-based APIs by July 2021. For the first time, health data became programmatically accessible in a standardized way. Before this, getting data out of EHRs was a nightmare of proprietary formats and closed systems. Now there is at least a foundation for building products that can ingest and synthesize patient information at scale.
So it is not that the market evaluated scribes versus pre-charting and chose scribes. It is that scribes were the thing you could build when speech-to-text matured, and pre-charting is becoming buildable now as AI capabilities and regulatory infrastructure catch up. The timing is not a coincidence.
And here is where I need to be even more honest: scribe companies are already moving into pre-charting. DeepScribe, for example, has launched a pre-charting feature. Their CEO, Matthew Ko, talks about this. (https://www.deepscribe.ai/resources/matthew-ko-ceo-deepscribe-ai-pre-charting)
They have funding, customers, and EHR integrations already in place. So what is the real differentiation?
I think the honest answer is that I do not know yet, and that this is an early market and no one has won yet. The space is underexplored and pre-charting as a product category is just emerging. Freed wasn’t the first scribe company, but it still won in the market of independents.
The advantage would be the learning that would compound, and will build over time. This will come from being in the market and executing, not from some structural moat that exists on day one.
Someone might argue that scribes could capture pre-charting by having the doctor dictate the relevant information before the visit, but this does not work. When a doctor dictates, the scribe hears only what the doctor chose to say and does not see what the doctor looked at and rejected. The training signal from pre-charting is not just “here is what mattered” but “here is everything, and here is what mattered versus what did not”, that contrast is where the learning happens, and dictation gives you only the positive examples. You never see the negative examples, and you cannot train a system to filter if you only ever show it the output of filtering rather than the input. Additionally, tacit decisions do not get verbalized, doctors do not say “I am ignoring the dermatology consult from 2019 because it has no bearing on today’s visit,” they just skip it.
The other thing worth acknowledging is that this will likely be an oligopolistic market rather than a winner-take-all situation. There will be multiple winners, and the question is whether we can be one of them by moving fast, building the right feedback loops, and compounding our learning before the market matures.
That is a less dramatic story than “we have a structural advantage that no one can overcome.” But I think it is the true story, and it is still a compelling opportunity.
What this actually helps us build
Beyond the general learning about clinical reasoning, pre-charting accumulates specific assets that compound over time.
The first is a computable patient record. EHRs store data, but that data is not usable by algorithms because a fax is a PDF image, an external hospital record is unstructured text, and lab results come in different formats from different systems. Pre-charting requires ingesting all of this, extracting structured information, and normalizing it to a standard representation, and the byproduct is a longitudinal patient record that a machine can actually query and reason over.
The second is specialty-specific attention patterns. Through doctor edits, we learn what cardiologists focus on versus what nephrologists focus on versus what endocrinologists focus on. When a primary care doctor sees a complex patient with heart failure, kidney disease, and diabetes, we will be able to surface what each relevant specialist would notice.
The third is risk stratification from real patient trajectories. Over their course of treatment, some patients will deteriorate, some will be hospitalized, some will develop complications, and some will get better. These outcomes are often predictable from patterns in the patient’s history, but no one sees these patterns in time because the information is scattered and the signals are subtle. We have the computable patient record, we see the full longitudinal history in structured form, and we see what happens next because every subsequent encounter updates the record. This allows us to correlate patterns with what actually happened to those patients and identify which signals predict deterioration before it becomes clinically obvious.
The fourth is practice-derived evidence that complements clinical guidelines. Clinical decisions are supposed to be guided by evidence, and the gold standard for evidence is the randomized controlled trial. Trials produce findings, findings get synthesized into guidelines, and guidelines tell clinicians what to do. But this system has a fundamental limitation: trials are expensive, they take years, and they enroll narrow populations who meet strict criteria. The result is a guideline that says: for patients like those in the trial, this intervention worked on average. The famous statistic is that it takes 17 years for research findings to reach clinical practice. But most real patients are not like those in the trial; they may be older, they may have multiple concurrent conditions, they may be taking multiple medications, and their situation is more complicated than anything that was studied. Most clinical questions will never get a trial. For patients on drug A who also have condition B, should we add drug C or switch to drug D? What is the optimal sequence of treatments? Which patients are most likely to benefit? What happens when you use this treatment in elderly patients who were excluded from the original trial? Trials may systematically exclude the elderly, patients with multiple comorbidities, pregnant women; the patients who most need treatment are often the least studied. And trials test drugs in isolation, but real patients are on five, ten, fifteen medications, and no trial will ever systematically test all those combinations; just because there are too many. But we will see what thousands of clinicians actually do for similar patients, the choices they make, and what happens next in the longitudinal record. Over time, we learn which approaches lead to better outcomes for which types of patients. This is not a causal claim, we treat these associations as signals for prospective validation, with appropriate risk adjustment to account for confounding. But even as signals, they are more than anyone else has.
Back to Dominik’s question
So let me come back to where we started, Dominik’s question about why our starting point gives us an advantage over scribes.
The answer is that scribes capture conclusions while pre-charting captures reasoning. Scribes hear what a doctor decided after they had already figured out what to do. Pre-charting sees how they got there, revealed through what they keep, delete, and emphasize when reviewing raw patient data.
I do not think this gives us a moat that no one can overcome. Scribes are already building pre-charting products, the market is early, and no one has won. But the learning compounds over time, and we have been capturing clinical reasoning from day one. The model is not the moat; the learning is, and that learning accumulates through being in the right part of the workflow to capture it.
But the real answer is bigger than scribes versus pre-charting. Clinical reasoning today is private and perishable. Each doctor builds it over decades, cannot fully transfer it to others, and loses it when they retire. What we are building is a way to capture that reasoning, accumulate it across thousands of doctors, and make it shared infrastructure that the next generation can build on rather than starting from scratch. That is what it would mean for medicine to become a learning system instead of a static one.
That is what we should have told Dominik.