AI Scribes in Healthcare: Hype, Reality, and the Fight for Clinician Time — The AI-Ready Doctor

From Epic's Art to Oracle's Voice, EHR to Doximity Scribe, everyone's suddenly in the scribe game. But let's cut through the noise. Are AI scribes just a shiny feature, or are they the frontline revolution in clinical workflows? Dr. B, are you ready for it?

Yes, let's go ahead and get started. I'm ready.

Amazing. So you've called Scribes a native product. What exactly do you mean by that?

You know, when we meet in a room, whether it is in the hospital or in the clinic, that dyad of patient and doctor, when they have an interaction, it is captured in a medical note. That medical note is taught to us since the first day in medical school, how it is structured, how do you filter the information in order for it to be organized in that note. If you go anywhere in the world, while there might be a few cosmetic differences, for the most part, a medical note is part of the DNA of clinical medicine. So if you have ambient scribes, they usually have as an output a medical note. You want to make sure that because it is in the interface between doctor and patient, that it is respecting the medical notes, how they are captured. What does that mean? For example, if I was to speak to a patient about a social history. Where have you been? Where do you live? Where did you move from? This is part of knowing the epidemiology or any risk factors, but AI, the earlier versions were trained to filter out anything that was non-medical and that was non-medical. So it's not until the later versions that we started to see the AmbientScribes leave some of that social history there because it is important to us. It's not just talk. So when I say that it is native, what I mean by that is it is part of the DNA of medicine. Additionally, as we see electronic medical records moving into the space of ambient scribes, they use the word native to describe that they have created an ambient scribe within the their environment, which means that it is native to that electronic medical records, by virtue of which they say it's going to not worry about HIPAA because it is within a cybersecurity environment. You don't have to worry about medical legal. You don't have to worry about voice or a conversation being recorded. But at the same time, the flip side to that definition of native is the fact that Unfortunately, you have a simplified version of what an ambient scribe, whereas somebody who dedicates their whole business model to ambient scribe is constantly improving it and modifying it. We do know now that Open Evidence has its own scribe that had something that went to the pain point of physicians, which is giving you evidence. So as I'm talking to my patient, when it made my note in the end, it also inserted certified studies and articles, we do find that it is a bit on the basic side. Still, no matter, it does give us evidence. So it's almost choice as to, as a physician, what kind of Ambienscribe do you wish to use? Do you want to use one that is standalone, much more sophisticated, constantly being upgraded, improving like in the example I mentioned about the social history, or do you want to use one native to the electronic medical record, both of whom though have to respect the medical note, which is the end output, the end product. If you're changing how notes are captured, you're changing the DNA of clinical medicine.

So does Doximity Scribe respect the— HIPAA?

Yes. All the ones that I use, and I have used a few. For example, I've used Freed AI, I've used Nabla, I've used Scribe, ScribeD, I've used Doximity Scribe, Open Evidence, and even the ones that are native to the EMR, such as Copilot and DAX. Those are all names of scribes, and there's more coming. That do help with scribing. What I liked about Doximity Scribe, first of all, disclaimer, I am a Doximity fellow, digital health fellow, where we beta test a lot of their digital products before they hit the market. Almost 1,000 physicians worked on that product. And I like the fact that there are templates that those of us in various specialties subdivided into groups, created those templates and added some of the things that we do natively in our specialty. In my specialty, I needed to add the oxygen saturation. I needed to add the risk factor of asbestos exposure. I needed to add, in the critical care, we needed to add the ventilator settings. We needed to add the vasopressors. So those are things that each template had its own, and I was able to create, even after the templates, my own. For example, I created a template where I gave it a command that every time it gave me a medical note, at the end, it gives me the ICD-10 codes, and it gives me a small summary that I could copy and paste into the sign-out sheet. So it saved me a lot of time beyond just the actual note itself and the transcription.

The open evidence scribes, like, I Would you choose that over Doximity? How does it make it better that you have, that you have the, these studies at the end?

Good question. I'll tell you two things. One of the tips that I would ask the user is Doximity is for free. Open Evidence is for free for physicians. Some others, they have limited number for free, and then after that you got to pay. What I did was I invested for a month or two in this one and that one, and I ran them all at the same time. And in the end, I compared the notes and I compared my ability to change them to the way I like. And that's how I narrowed it down to the ones I'm using today. Open evidence. When I am looking at a patient who came in, for example, yesterday, We were talking, a friend and I, about a patient with low sodium, hyponatremia. And this patient came in with a few things that are calculations and typical questions that we have when we are talking about board exams and so on. So the algorithm of hyponatremia is a heavy one, and some of the calculations and some of the interpretations can mean that you have to focus and refresh your memory about all of these things. When you use open evidence, in the end, it actually gives you what it most likely is, and it gives you the studies that it is connected to. So there's a bit of show-offism when someone does a consult, for example, when they ask me to consult on somebody for pulmonary problem, and in the end, I give the differential, and then I give the studies why this differential is. In the end, people read my consult and say, wow, this is an academic consult. So open evidence democratized that, made that easy to everyone. That does not need to go and search and spend time looking after the patient is done to go to the next patient. When you're busy and you have 15, 16, 20 patients to go through, it does it for you. So all you have to do is curate which study applies to the patient in the end because it already gave it to you. And it has the name recognition of the New England Journal and JAMA. So the space of ambient scribes isn't just all hip-hop. They're all HIPAA compliant. What it is is whether or not it gets into the pain points of physicians and how their burnout is made during that note-taking time.

So it's just making it easier, really, giving you more time. Now, you mentioned that you use a couple. So how do you specifically choose which one you're going to use for what? So you mentioned this— person with the, this patient with the low sodium?

I'll give you some of the bells and whistles. Doximity, for example, I like that it gives me exactly the input. What we talked about, no more, no less. In the assessment, it gives you what the assessment is, and in the plan, it gives you exactly the plan you spoke about to the patient. Doesn't add anything. Sometimes that's what we want. We do. I don't want you to add any more things. Open Evidence does that, but adds studies. But it's quite simplified. It has not yet matured in the complexity of the language it uses to give me that assessment and plan. If I go to Freed AI, for example, it has more text. You can go to the backend if you are savvy enough and change some of this. I can ask it to be more verbose. Or I can ask it to give me ICD-10 codes, or I can ask it to not give me studies or to give me studies depending. But you still have to have the human in the loop or here, a physician in the loop. So each one has their own bells and whistles. And the only way that you can know which one is to run them all on the same patient and see which one compared and contrast and see which one gives you the one that you like.

And running them all on the same patient wouldn't take away from its purpose. That it's like saving you time?

Not really, because in the end you still have 4 notes. Each one of them gave you a note. You just pick and choose which one you want. And you're not going to do that for the entire career. You just do it for an A/B testing period where you see which one fits your needs. I know, for example, my, one of my colleagues, I showed her Freed AI, but she liked one that's called Heidi, and they included it in their private practice. I know, for example, somebody that I showed Doximity to did not quite like the outputs, but they like the ICD-10 code. So they ran it next to another one that couldn't give the ICD-10 code. One gives a note and the other one gives the codes, and therefore that uses, that saves them time. So it really depends on the person. We are now in the era of customization. So the purpose of running 2 or 3 at the same time is to pick which one you want to customize for your own needs.

It's like what you're, which one you're more compatible with almost, no? Like what, which one is for you easier to train?

Correct.

Okay. As we're talking about saving time, let's talk about pajama time. Do these scribes actually give clinicians hours back or are we just shifting where the work happens?

So there still needs to be studies serious studies, not studies that are done by the manufacturers of that scrub. Serious studies. I am aware of one study at least where it actually looked without confounders and while correcting for a lot of factors that may actually affect it one way or the other. So all things considered, there was only a few minutes saved compared between Scribe and non-Scribe for the pajama time. Why? You still have to review the note and edit it in order for it to be trustworthy. Not only is it a note that is medically based on the patient's case and how you're going to make decisions about their health, but it's also a legal document. So on the one hand, you still need to edit that note. The one thing that I could say has been saved consistently is the cognitive load. Those of us that work 12-hour shifts in the ICU or in a hospital or have 20 to 30 patients to see a day as a primary care physician in the office, a lot of times you're tired towards the end of the day and you don't have as much sharpness as you had in the beginning of the day, so you may miss certain things, and AI, AmbientScribe, helps with that cognitive load. Pajama time, the jury's still out on that.

Okay, yeah, it makes sense that it's not the same time invested when you can do something without too much thought. Once you're tired, and that's what it's actually helping you with. In your experience, has it saved you pajama time?

Incredibly so. Incredibly so. The difference is, this is why I said you have to correct for factors. The difference is, I mentioned in one of my previous podcasts, I work on these things when it's inconsequential. For example, I use AI and I learned how to prompt. I talked about prompt engineering before, and now I'm talking about context engineering. AI works best when you give it a prompt of what you want and context for why you want it and in what way you want it back. In that, I am trained in this outside of being busy. So when I am busy, it is easier for me to craft a prompt that gives me what I want immediately. And that's how I'm able to take those AIs and give them sort of commands as to what I want so that it saves time from that. If I didn't, I would be constantly frustrated that it's not giving me what I want and it's taken away from my pajama time anyways. So it doesn't matter if I do it or didn't do it. The studies comparing people that are professional in this, or proficient, shall I say, in this, and people that are not, have not been done. So pajama time is not yet compared head to head. But for me personally, it has.

Okay, so it is a combination of how advanced the technology is, or the tools are, and how, you use a word, how—

Proficient the person is.

Or how literate. You are in AI, right?

I like that better. AI literacy is exactly what I do. So yes, how literate the person is in AI, much as they would be literate in ultrasound or literate in intensive care unit lingo or literate in research. Absolutely.

So those who are more literate in an ultrasound will do it faster. You are correct. Awesome. So who's really building for physicians here and who's just playing them with shiny tech?

One would say that the electronic medical records are building for physicians. That is true. On the other hand, I feel like they're building for the economic impact. They're building for the bottom line. They're building to create more than just an ambient scribe. They're building to create a practice management tool. It's no secret now that they're going into billing, coding. They're looking at making sure that they have optimized capturing all the money value that is in that encounter. So I don't know if this is all building for physicians. One thing I will say is I was quite impressed with, this is why I joined them. I was quite impressed with Doximity building directly for physicians. When I was in COVID, I discovered their dialer, and that is where you get the ability from within the Doximity app to get issued a phone number and to call from that. That would be your phone number. You can video with patients I've used it a lot in the ICU during COVID You can also get a fax number that is HIPAA compliant. They also created many, many other tools such as a Doc Defender. And when you sign up to it for free, it actually goes online and takes your information from a variety of websites across the internet so that you don't have your private information there. So I feel like Doximity has an ethos of building for physicians. And I was completely shocked in a good way, the amount of resources they put to build that Scribe, the meetings that we had, the amount of fine-tuning that they did to the tool, the customization to each specialty built by physicians for physicians. That, I would say, I was skeptical at first, but slowly I got to say this is built for physicians. When I looked, for example, at open evidence, that is built for physicians. You don't just start to add evidence there unless you have been a physician within a physician interaction. In a physician's shoes, because those are things that only physicians know that are their own pain point in the workflow. That's the key. Build for the workflow. Don't take the, the physician out of the workflow to adopt your shiny product.

Okay, so if I understand, building for a clinician is that is trained by clinicians, speaks the language, knows exactly what the clinician needs. But EMRs also do that, but they just don't have the interest of the clinician in terms of making their life easier. There's other interests involved, involved in that.

That is correct.

And as a teaching clinician, do scribes change how you train residents? Do you spend less time correcting their SOAP notes and more time teaching medicine itself?

This is a burgeoning field. It is just starting and it's moving fast. When we talk about ambient scribes, we cannot talk about them without talking about the informed consent and letting the patients know. On the other hand, the patients have been tainted already by what AI pejoratively could mean. So usually what I say is I have a note-taker. That helps the patient understand that this is a transcription service, not a recording service. Doximity does that. Open Evidence does that. But others, they actually keep the clip of the voice. So for me, when I look at specifically an AI, how it does, I don't want it to take away from learning.

Think.

It's interesting because when you tell someone, I have an AI that's going to, I have an AI that's gonna help you write. Let's say you went to a writer. It can write the whole book for you. Well, writing is thinking. So it's through writing that medical note that you think through the diagnostic path that the patient is going through, integrating it in order to help your patient. Is AI going to replace that? Hardly. It may come up with predictions. It may come up with summarizations. For it to come to the decision support tool, that's a different field, and I don't think it's mature for it just yet. The residents that I round with, I share with them what an AmbientScribe does, but I tell them, write your own note, compare it to the AmbientScribe, improve your note before you start to rely on AI from the get-go. And I do a lot of in education, medical education, I do a lot of human versus AI competitions and they are fun. And you can see in the end how people's perception of AI completely changes because they realize that AI harped on one portion of that conversation and took a whole different conversation direction. And because it is AI, it multiplied it and magnified it when in fact that was a small portion of the conversation that probably wasn't as important. And only somebody who was astute enough, human enough, paying attention enough, in the room enough would have picked up on that and put it in lower level of probability, whereas AI took it and ran with it. So afterwards, in the end, when they compare notes, they realize that that's not it. This is the weight loss was due to an increased catabolic demand from the slow sepsis that we found to be from a small abscess in her thigh rather than weight loss that AI took into a cancer diagnosis and suddenly asking for CAT scans and PET scans and oncology consults and the anxiety to the patient and so on and so forth. So just because it is able to give us this shiny note doesn't mean we should let go of our thinking process.

That's a very important point right there. And to those who are very skeptical about, you know, training resonance with AI, I think this is actually better than before because you actually put the resonance to compare and that way understand the importance of being present with with what you're doing.

Absolutely. We don't want just faster. We also want better and more polished, and we do not want to get elegant hallucinations. It's not about speed. It's about safety. First, do no harm. We didn't just start sending our patients to a CAT scan or start putting biopsies into people until we understood What is the risk? What is the benefit? What's the comparison between them? And involve the patient in the decision sharing mode. Some patients told me they don't want AI in the room, and that's fine. It's their prerogative.

Okay, that's also really nice to understand that you can always, you can and you probably, I don't know if must is the word, but you, am I right? You must. Actually let in the patient in that decision-making because at the end of the day, it is his life, their life.

We will always find that the physicians side with the patients by and large. So any shiny object that comes in and challenges the physician to adopt it without including the patient and without putting there that there's trust trust, first the trust in the patient, the trust of the patient in us, and our trust that the tool will deliver something that will help us help our patient. If this is not in the equation, the adoption in healthcare will not happen.

But I think that's only going back to the values and the oath, right? That's really at the core of what you have to do as a physician when you're adopting something new.

Indeed.

Remember where you come from and what your purpose as a physician is, right?

That is what governs every day of our practice.

So what happens when the scribe gets it wrong, misses a nuance, adds a wrong med, or hallucinates diagnosis? Do you see clinicians becoming better editors or worse note— Note checkers.

Nothing will be perfect. There's no plan that goes perfect. In fact, the status quo isn't perfect. There are two sides to this. The first side is there is an initial phase of skepticism, and then there's a wow phase, and then there's a phase where I've been had. Look at what it made me do, so I'm going to be more mindful. This adoption sort of curve will make the physicians be much more mindful of how to edit notes. Is it going to make mistakes? Of course. Are we going to make mistakes? Of course. Regardless, I remember the, to err is human, and part of our humanity is that there will be situations in which, in which we would have, we would wish that we, it was different. But at the end of the day, that note is mine and AI made me do it is not going to fly. So the liability of it is the second portion. And that's what makes some people did not even adopt EMR. They're still struggling with it years later. So there are some physicians and some patients that will just choose the old-fashioned way, and there is nothing wrong with that. We should have space for the old-fashioned way. There are things I do the old-fashioned way. I still calculate my A gradient. I still calculate my anion gap, my osmolar gap. There are things that are still the old-fashioned way. The question that is still on everybody's mind is, will there be liability for not using AI as a tool available to you? And that is still— the jury is still out because is there a liability that I would not check the electronic medical record for a history of something and then because of it, I don't take an action and I become liable for it? Absolutely. So will there be a time where if you don't use AI, you'll be found liable? It's probable. We're not there yet, but it's a question on everybody's mind.

Okay, Dr. B, so let's be blunt now. Okay. Every EHR promised to reduce burnout, and most made it worse. Are scribes the first AI burnout cure that might actually deliver, or will we be laughing at these promises in 5 years?

We will be laughing at these promises in 5 years. No question. I don't think the burnout is one thing and one thing only, but the scribes are helping with one pain point in that, which is the note-taking, the tediousness of it, and the fact that we have moved our eyesight from our patient to our computer. We want to move them back to the patient so they can feel seen, heard, They can feel the connection. That is something that they are helping with once we become proficient and literate in AI and in scribes. The issue with it is scribes now are becoming more commoditized. They are free. Doximity or Open Evidence are free. So the question becomes, if the scribe is free, we probably are the product. And so there's now advertisement in the Scribe platform. So exactly who owns that note and who trains on it? And do we want our patient conversations to fuel somebody's model? I think that train left the station when Epic showed already that EMR or their Scribe has already trained on millions of encounters. That train has left the station when we see that the AI has been trained in the UK in the NHS, 5.8 million encounters. Nobody was asked. And so who owns that note? At the end of the day, we are the product if the product is coming to us that easy and for free.

So in 5 years ahead, in 5 years from now, when we're laughing at this, What do you foresee will be, we, we will be using instead? Is it in the same direction of Scribe or of Scribes, or is it a different product or is just better?

5 years from now, Ambient Scribes are going to be like the first iPhone. It's a remote memory, and if we were to use it now, we will be talking about what is this fossil? On the other hand, the current iPhones or any smartphones or even what, Alexa, we know they're listening in on our conversation and we've accepted it. We have accepted that they're listening to our children's conversations, to our most intimate conversation. They are in our TVs. We used to think this is conspiracy, and now we know that actually it isn't. There is— they had to put a lot of, shall I say, walls, firewalls in order to make that privacy stay private. One patient told me, so now we've got Alexa in the room, huh? And that's the reality. But trust depends on transparency. When an AI comes and tells you where is it trained at, what is their open model, What is the dataset? When you have access to all of that, that transparency fosters trust. We know nothing is going to be 100% private. The second part to that is transparency in what uses you have for it. We know now that Epic wants to use its scribes and its chatbot to not just facilitate— that is the marketing message. What it does is it's going to allow for a more complete capture of the money value of that encounter. It will have a prediction in terms of risk of hospitalization or who might not make it out of this hospitalization or who might. All of these risks now are going to be AI-generated. What they're going to be doing with them That's the question. The ones that are transparent with it are the ones that might have a better deployment because we will trust them much more.

Interesting. So that's not necessarily at the interest of, of the physician nor of the patient, but I think this conversation will, will go on for a little while. Now, regarding to this episode, Dr. Bee, before we close, let's do a quick Pulse Check. I'll run through this week's big AI headlines, and you've got 30 seconds each for your take. OK. Ready?

Ready.

OK, let's do this. Epic's Cosmos AI trained on billions of encounters. Does it scale equally safety at the bedside?

Scale versus safety. We have billions of encounters. Do we trust the averages, or do we trust the outliers? Why can't we have both?

Oracle's cloud EHR, agentic, voice-driven. Is this innovation or just Cerner in disguise?

Unfortunately, having a chatbot that I can speak to does not improve health, does not improve outcomes. So I still need to see the proof in the pudding.

Open evidence visits, real-time decision support inside the note. Brilliant or model clinical clutter? Clutter.

As long as it is close-ended questions, it is helpful. As long as somebody knows what they're doing, it is helpful. If you are just starting out, it's clinical clutter.

GPT-5 buzz. More powerful models coming, but are we just getting more confident mistakes?

GPT-5 has more powerful reasoning, but you have to ask it to reason. It all comes down to AI literacy. If I'm literal or if I'm AI literate, then I will pick up on hallucinations and mistakes. Otherwise, it's just a better confident mistakes kind of tool.

Cultural intelligence in AI. If AI can't understand a patient's culture, is it really safe?

It's not, period. If it doesn't understand it, it's not.

Okay, Dr. B. So that's today's Pulse Check, quick hits and hot takes from Dr. B. Subscribe so you don't miss the next one. AI scribes may be the native product of AI revolution in healthcare, but as we've heard, not all scribes are created equal. Some give the time back, others just add noise. And as Dr. B reminded us, the real winners will be the ones who put clinicians and patients first. If this episode gave you something to think about, share it with a colleague and come back next week. Until then, stay sharp, stay skeptical, and stay AI ready.

Stay AI ready. Being a skeptic, not a cynic. It's healthy. For the first time in decades, Scribes has given us back the eye contact. Isn't it worth an hour of our time to understand this? In order to give this back to our patients? Bring back some questions and stay curious.