Are AI scribes the first tech physicians actually like?

The electronic health record (EHR) promised to make medicine more efficient. Instead, it turned physicians into expensive data-entry clerks, spending evenings catching up on documentation while their patients' charts grew longer and their attention spans grew thinner. A decade of work-arounds — voice dictation, offshore transcriptionists — never quite solved what became known as "pajama time."

Ambient artificial intelligence (AI) scribes may be the closest thing yet to a fix.

At the University of California, San Francisco (UCSF), roughly 70% of physicians now use one, and 90% of those physicians say they love it. Robert Wachter, M.D., professor and chair of UCSF's Department of Medicine and author of “A Giant Leap: How AI is Transforming Healthcare and What That Means for Our Future,” sees that adoption rate as a signal of something bigger than documentation relief.

In his view, the scribe's real value isn't the minutes it saves, but the trust it builds between physicians and a technology they've long had reason to distrust.

The following interview, conducted in preparation for Medical Economics’ March-April 2026 Tech Issue, has been edited for length and clarity.

You’ve written about how EHRs helped create “pajama time” for physicians. Are ambient AI scribes able to realistically reverse that trend, or are we overstating their ability?

“Reverse” is a little strong, because the genesis of pajama time originally was documentation of the visit itself, and now it has also morphed into answering 100 emails from patients. Scribes do not address that, though hopefully other AI will.

It is the first technology we have seen that says to physicians: ‘We can give you back some of not only your time, but also some of the humanity of the doctor-patient visit.’

So it does not completely solve the problem of pajama time, but I think it is the first technology we have seen that says to physicians: We can give you back some of not only your time, but also some of the humanity of the doctor-patient visit. We can create a mechanism by which you can make eye contact with your patients and both appear to be, and actually be, paying real attention to them, as opposed to being a pretty expensive, grumpy data-entry clerk.

So in that way, it does save some time, but I think it is more important in resurrecting the humanity of the doctor-patient visit.

What is it about these tools that is really changing physician workflow? Is it mostly the time saved on documentation, or is it more about the connection with patients and the ability to maintain eye contact, for example?

It is a little bit of both. But I think the main thing it is doing is this: When we were on paper, I could scribble down whatever I wanted in my notes, sometimes in real time, sometimes afterwards, and still make eye contact with the patient.

What the electronic health record did was enable a whole bunch of new requirements for what needed to be in the note, because all of a sudden there was a mechanism by which my boss, Medicare, UnitedHealthcare, my malpractice lawyer, the quality measures and all sorts of other entities could be looking over my metaphorical shoulder and making me do things they could not make me do before. Given that opportunity, of course they took advantage of it.

So the note became bloated. It became full of checkboxes. It became disheartening, not only because it drove the physician away from what he or she thought they were supposed to be doing, which is listening to this other human talk about their issues, but also because it became both a time sink and a cognitive sink. You are spending so much of your energy inputting stuff that you are not fully present and listening to the patient.

What the scribes did was relieve you, to a large extent, of that documentation burden. It was very important that they did some real magic. This was not simply voice-to-text translation, which we have had for years. If all it did was produce a transcript of my conversation with a patient, it would be entirely worthless.

The real magic, and where generative AI came in, was the ability to take a conversation and figure out what should be included in the note, and what should be left out.

It could understand that the description of your chest pain belongs in the note, while a conversation about playing with your grandchildren last weekend or how well Johnny did in the soccer game probably does not. It could also pull together information in the formula that matters for documentation. If I start talking about chest pain, then shift to some unrelated topic and later mention shortness of breath, the AI needs to know those clinical details belong together in the note.

The tools became smart enough to do that, and doctors said: This solves a really important problem. I can now make eye contact with my patient, be engaged and not spend an enormous amount of time documenting.

Where I think we probably overhyped it was that we thought it would save five or 10 minutes a visit. It turns out it does not save that much time. It saves a little, but part of the reason is that some of that time is repurposed into actual, genuine human contact, which feels better for the doctor and feels better for the patient. And although it is hard to prove, I think that is ultimately better for health care because it makes it more likely I am really tuned in to what is going on with you and asking the right questions.

You’ve used the phrase “training wheels” for AI documentation. Why is documentation the right entry point for AI, and what makes ambient scribes lower risk than other clinical uses of AI?

AI in health care is not new. It was actually a hot issue in the late 1970s and early 1980s. But it was not ready for prime time for a lot of reasons. Our documentation was on paper, so if you wanted to use a computer to do AI, you had to reenter all the data. There was no cloud. And the early technology was built around if-then statements.

If-then works fine for simple problems. If the patient has a sore throat and swollen lymph nodes, maybe they have strep throat or mono. But it falls apart very quickly when patients have the problems real patients have, which is five things going on, three different diagnoses and nine medications.

In some ways, it was not ready for prime time, but the biggest problem was that the early founders of medical AI chose diagnosis as the place to start. I interviewed a lot of them for my last book, and I asked them why. They said they knew it was hard, but it was the most interesting problem. Very smart people made that choice, and it flamed out.

It flamed out partly because of the technological limits, but also because you do not start on the hardest problem, especially when, if you get it wrong, you can kill somebody. That is not a good strategy for winning over the trust of the incumbents, in this case doctors.

If I can have happier doctors, happier patients and maybe save a few minutes per visit, it is probably going to come close to paying for itself.

This time around, some of the old problems have been solved. The data are digital. We have cloud storage. The AI is no longer just a bunch of if-then statements and can now deal with the complexity of real medicine. But we also started in a smarter place.

If we had started with diagnosis again, that would have been a big mistake, because as good as AI is, it still sometimes gets things wrong. It still hallucinates. And when it does, if I am not paying attention, I can label a patient with the wrong diagnosis and maybe treat them for the wrong thing.

So I think this time we are being more thoughtful. We are starting with relatively easy things to do, where, if you get them wrong, you probably are not going to kill anybody. AI scribes fit that quadrant. There was already a model for this. At UCSF, we had human scribes 10 years ago, mostly premed students who came into the office and typed while I talked to the patient. So there was already a model for a nonexpert, nonphysician to take in the data and put it into the chart.

The AI tools turned out to be good enough to do that well. And even if they make an error, it is usually not catastrophic. The AI drafts the note, and I review it. At least in theory, I catch the error if it exists.

From a health system perspective, the economics made sense too. If I can have happier doctors, happier patients and maybe save a few minutes per visit, it is probably going to come close to paying for itself. And when you factor in the cost of recruiting a new physician, which for a primary care doctor is often estimated in the million-dollar range, it may absolutely be worth it.

All the ducks lined up for this being the first use case, and I think it has worked out pretty well. At UCSF, all of our 3,000 to 4,000 doctors now have access to an AI scribe. About 70% of them use it. About 90% of those say they love it. We have almost gotten to the point where, if we turned it off, we might lose a fair number of doctors, or it would be harder to recruit a doctor if we did not offer it. It has almost become an expectation of practice now.

What drawbacks or complications are physicians seeing so far?

It is not perfect. The tools are getting better, but ideally, you would want one that documents quite differently for a cardiologist than it does for a primary care doctor. Right now, they do a little of that, but not perfectly. I think that will get better over time as they get more data and more reps.

If it makes a mistake, it is not the end of the world, but it is not nothing. You would like the physician to be reviewing the note. But as I talk about in the book, the human-in-the-loop idea sounds better on paper than it is in reality.

If I have used the scribe for the last 49 notes and reviewed them carefully and they were perfect, am I really going to be fully attentive as I review note No. 50? If I am human, the answer is no. At some point, I will begin overtrusting the data.

We are not that worried here about deskilling. It is hard for me to imagine I am going to lose my ability to read over a note and see whether it is accurate. That feels more relevant in other areas, like diagnosis.

One downside is that the tool may have underdelivered a bit on the efficiency advantages. People thought it would save enough time to justify clear ROI [return on investment] through more visits, and the early studies suggest there has been some time savings, but not quite as much as people anticipated.

Probably the biggest potential downside is more of a cognitive tradeoff. On the one hand, I like that I am now listening really intently to the patient and making eye contact. On the other hand, when I used to type my own note, that was a cognitively active process. Sometimes while typing I would realize I forgot to ask an important question. There is something about the human act of thinking through something and writing it down that is probably more cognitively active than reviewing a draft someone or something else created.

I think that comes up across education and probably in journalism, too. If something drafts it for you, is something lost because you did not go through the hardship of creating the first draft? These are trade-offs. Doctors are voting with their feet and saying they are willing to make that trade-off. But to me, that is probably the biggest one, and it is hard to quantify. I have not seen a really good study that gets at that instinct.

You mentioned that 70% of UCSF physicians are using the AI scribe. Do you have any insight into the 30% who are not?

It is probably a little bit of everything. For some, what they are doing now works fine. It is not like if you do not use this tool, you cannot practice effectively. Some people have been doing things a certain way for 20 years.

The technology is not incredibly complex, but it is not nothing either. You have to learn a little bit about how to use it, how to get the note recorded on your phone and sync it with Epic and all of that. It is not massive amounts of training and change management, but it is not zero.

For an entirely elective tool that you do not have to use to have a 70% adoption rate in that amount of time, that is pretty good for health care.

Some people really do feel that the process of typing and writing the note is valuable to their thought process and do not want to give that up. It also depends to some extent on your specialty. In primary care, where the history is prolonged and you are taking in a lot of data, it may be more useful than in a highly specialized area where you are asking a relatively set series of questions and the documentation is already fairly checkbox-driven.

Some people tried it and found that it did make errors every now and then, and they felt they could not trust it. For some, the physical exam is a particularly important part of the process, and the technology is not great with that. It cannot really do the physical exam. You have to narrate it, which is a little weird if you have never done that before.

Some people have found that a few patients are a little put off by it. We ask for consent, and some patients are a little creeped out by the recording. So there are lots of reasons.

But the technology did not exist three years ago, and for an entirely elective tool that you do not have to use to have a 70% adoption rate in that amount of time, that is pretty good for health care.

Do you see overreliance as a real risk, and are there ways health systems can reduce that risk?

Yes. You are not going to forget how to proofread, but you may stop proofreading very thoughtfully as you come to trust the technology. That is just a general human tendency.

We are thinking about it. I was talking to a senior leader at one of the big tech companies today, and he said they are beginning to develop tools for areas like medical education that build in ways to make sure the user is actually reviewing things carefully.

I think one of the fundamental challenges in AI and health care, and maybe AI in general, is that the tools are now good enough to be useful, but not perfect. So it is going to be important to be strategic about what this AI-human partnership really looks like.

The mistake would be to say, “Here is the AI, here is the human in the loop, now we are done.” That is not a fail-safe system. We have to be intentional about how to make it work.

Maybe the system gives you a confidence signal — green when it is very sure, yellow when there was a word it did not understand or some conflict in the note, so you know to review the yellow note more carefully. Maybe there are other kinds of checks. I do not think anybody has really tested all of this out yet, but the key concept is that the AI-human dyad sounds safer than it is. It is not a fail-safe system, and we need to design it thoughtfully.

When health systems think about ROI for ambient AI scribes, is the real return financial, or is it more about physician retention and morale?

Today, you would have to say it is more about retention, physician experience, joy in practice and maybe to some extent patient experience as well, all of which I think do have quantifiable ROI.

When you factor in the cost of recruiting, retaining and replacing a physician, and probably the value of retaining patients who feel their doctor is actually paying attention to them, those things matter. We have certainly had patients say they do not want to come somewhere if it looks like the doctor is not paying attention.

There is probably another benefit that will be harder to quantify but may end up being even more important: winning over hearts and minds and opening the door for more and other AI tools.

If you just do the math on minutes saved and translate that into more patients per day, there is a management question there, too. If I save five minutes, do I turn that into five more minutes on the hamster wheel? If part of my motivation is to create a happier doctor and patient, probably I should not do all of that.

I would say today the AI scribe probably kind of pays for itself in pure throughput and time savings. But the real benefit is joy in practice, recruitment, retention and maybe patient experience.

There is probably another benefit that will be harder to quantify but may end up being even more important: winning over hearts and minds and opening the door for more and other AI tools. The home run here is not the scribe. The home run is robust clinical decision support that helps make diagnoses, suggests the right tests and suggests the right treatments. That is where the clinical benefit really is.

You did not want to start there because the stakes were too high. But I think when the story of AI scribes is finally written, their biggest contribution may be that they made physicians more receptive to the bigger, broader and ultimately more impactful tools that come next.

In your new book, “A Giant Leap,” you take a broader view of AI in health care. What should physicians keep in mind about AI scribes and the wider role of AI in medicine?

The overall theme of the book is that I came out of it pretty optimistic about AI’s role in health care, in part because this is an example of a tool that really does solve a real problem and the users are happy.

As one person told me, this is the first tool they have seen in information technology where people run to the chief information officer’s office and say, “I really want this.” Usually it is the other way around. Usually people run to the CIO’s office and say, “Why did you bring this in, and couldn’t we have found something better?”

It is not going to be easy, and it is not going to be perfect. There will be plenty of speed bumps. But because the tools really do amazing things, solve important problems and the system is in desperate need of transformation, I think they have an important role to play.

I am optimistic because the tools have remarkable capabilities. Three years ago, the idea that you could replace a human scribe with an AI scribe would have been inconceivable. And in a way, I feel a little sorry for the companies, because they built these unbelievable tools and within two years they became commoditized and prices started falling. But the tools really are remarkable in solving important problems.

The other reason I am optimistic is that the health care system is so broken that we need tools like this just to get through the day and do our jobs. Nobody was sitting around saying the health care system was perfect and did not need help.

So in some ways, the theme of the book is that I am quite optimistic. It is not going to be easy, and it is not going to be perfect. There will be plenty of speed bumps. But because the tools really do amazing things, solve important problems and the system is in desperate need of transformation, I think they have an important role to play.