In Episode 9, host John Mangano sits down with Dr. Brad Bowman, MD, Chief Medical Officer and Head of Data Science and Quality at Healthgrades. Brad offers a unique perspective on the intersection of healthcare and artificial intelligence (AI). You may know that AI is already transforming medicine, from reading X-rays to predicting health outcomes. Brad shares why he believes it will do so much more and that human limitations are holding the field back.
The conversation dives into AI's potential to enhance patient care by reducing administrative burdens, improving diagnostics, and even rewriting doctors' roles. Brad also explores the ethical and emotional implications of letting machines take on more responsibilities in healthcare, but he highlights the surprising ways AI could restore humanity to a broken system. Tune in for an enlivening discussion on the opportunities and challenges ahead as AI reshapes healthcare—and what it means for patients, doctors, and health marketers.
Transcript:
John Mangano (JM): Welcome to Deep, the health marketing podcast. I'm John Mangano. Today I'm talking with Dr. Brad Bowman. Brad is both an MD and a data scientist, uniquely qualified to help us understand how AI is going to impact health and healthcare. We'll also explore how AI impacts marketing, and the parallels between healthcare and marketing. Let's dive in. Brad, great to have you on the show.
Brad Bowman (BB): Thanks, glad to be here.
JM: We talk to a lot of people about marketing and messaging, but right now AI is what everyone's talking about — usually in the context of media, but there's so much more to it, especially when it comes to health and how physicians and patients will ultimately receive care. You're a foremost expert in that space, so thank you for joining us.
BB: Thank you — I'm honestly a bit humbled to be here talking about this. I'm not sure I'm the foremost expert, but I'll do my best.
JM: For those who haven't worked with Brad — he's done some amazing things and touched a lot of people personally, including me. Brad has spent years using data science to understand physician success rates with patients, building the logic behind the rankings used at Healthgrades, and more recently through a service called My Health Match, which helps match patients with the doctor most likely to give them the best outcome. I'm personally grateful for that — a few years ago I pretty thoroughly destroyed my ankle doing something a 50-year-old really shouldn't be doing, skateboarding, and it was a call from the emergency room to you, Brad, that ultimately connected me with the best possible surgeon for my situation in all of DC. I know I'm not alone — I could point to at least ten other people I know personally whose lives, or a loved one's life, your service has genuinely changed.
BB: I appreciate that, and I'm glad it worked out.
JM: Brad, we talk about AI constantly. As much as marketing is my world, does it really compare to how AI could impact health outcomes directly, through doctors? There's a lot happening right now — what do you see as the biggest changes so far, and where are we headed?
BB: It's a fascinating topic. If you go back even a decade, we had machine learning, then deep learning, and now we call it AI — I'm honestly not a huge fan of the term, because it makes it sound like the machines are taking over. But depending on who you ask, people will tell you either that AI is going to have almost no real impact on healthcare, or that it's going to change absolutely everything. I lean toward the latter.
Going back even a decade, we already have computers that read chest X-rays and mammograms more accurately than humans. We have computers that read EKGs and predict heart attacks better than humans do. Especially with something like a mammogram, where you're really talking about computer vision at the pixel level — computers simply have better "eyes" than we do, and they learn associations better than we do.
When a human looks at an X-ray, they're generally checking for maybe ten specific things. When a computer looks at the same X-ray, it's evaluating millions of different patterns — this pixel next to that pixel, but not that other one, and so on. That combination means something to the computer in a way we can't always articulate. We don't even fully understand how these systems arrive at some of their conclusions — they've learned patterns we don't consciously grasp ourselves. We already trust machine learning and AI with big computational tasks, like deciding when to shock your heart if it stops — that's built into defibrillators today. For those kinds of tasks, it's just a much better tool than we are. And I think that trend is accelerating. Computers are now better at diagnosing certain conditions than we are, too. This has been slowly creeping in for a decade, but I think it's really speeding up right now.
JM: So this isn't new — we're just talking about it more. In a lot of ways, the shift already happened; we're just becoming more aware of it as it goes mainstream. Doctors have been relying on AI for a while now to help them be better doctors, right?
BB: Right, it's already been used in these applications — but I think physicians have largely held onto, or maybe "clung to" is more accurate, their cognitive role as the head decision-maker, treating AI more as a supporting tool. For example, today a radiologist still signs off on a chest X-ray even after a computer has already flagged what's on it — there's still human oversight built in.
What we're seeing now, and I think this is part of the industry's underlying tension, is that computers are starting to take over more of the actual diagnostic function, which more directly changes what it means to be a physician. But arguably, that's something that should have happened a while ago.
If you think about it, humans have a fairly limited capacity to learn and remember things — we can maybe track four or five things at once, while computers can track thousands, tens of thousands, even millions of things simultaneously, without ever forgetting. We're also not great at doing math in our heads, calculating odds ratios, that sort of thing. It's honestly not really a fair fight. And we're not getting any smarter as a species — we've kind of plateaued — while computers keep getting smarter every year. At some point, you could argue that humans are actually holding medicine back, and we should figure out a more assistive or oversight-based role for these machines. I think that will profoundly reshape what medicine looks like going forward.
JM: That's a great point — technology is really just a fancy word for "tool." Not that long ago, the calculator was new, and nobody was too worried about a calculator doing math for them, because the benefit was obvious. We're all fine using Excel, which handles things a bit more complex than a calculator. AI is doing something more sophisticated still, but the same principle applies — it becomes a tool. If I'm having heart surgery, I don't really care whether a human or a tool is doing it, as long as it gets me the best outcome. That's the direction we're heading. What percentage of medical tasks do you think AI will eventually handle better than a human's touch?
BB: The calculator analogy is a great one. I think in the future we'll use AI the way we already use a stethoscope or a CAT scan — we don't try to replicate those ourselves. More of the cognitive decision-making, memory, and medical judgment will rely on AI as a kind of calculator, with humans providing oversight.
Two people who've been talking about this for a long time, with very different views, are Vinod Khosla — co-founder of Sun Microsystems and founder of Khosla Ventures — and Dr. Eric Topol, formerly of the Cleveland Clinic, now at Scripps. Both are extremely sharp and usually right, so it'll be interesting to see how this plays out.
Khosla, for example, once told a class of medical residents at Stanford, more or less, "In five years, you'll all be obsolete — we won't need you anymore." Not exactly a popular message in a room full of doctors. I still think he may be directionally right, even if his timing was off — we obviously haven't replaced all doctors yet. But he told this fascinating story to that same class: something like, "What if every time the price of bananas in Brazil rises above a certain point, purse snatchings on a specific street in Atlanta double — and this pattern has held for a hundred years, never once wrong?" If that happened again, would you take it seriously?
In medicine, we get very hung up on causation — if we can't explain the cause-and-effect mechanism, we tend to discount the pattern entirely. That's foundational to how medical research works, but it also keeps the pace of medical progress fairly slow and incremental. If these AI systems can start identifying patterns in nature and in medicine directly, we could start making connections and improving treatment much faster. Khosla's message to those residents was essentially: stop trying to fully understand the mechanism — the machines will figure it out, and they'll consistently be right. His view is that there will be little or no remaining role for humans in this, because these cognitive tasks will eventually be redistributed entirely to machines.
JM: So that spans the whole process — diagnostics, treatment, surgery. What about mental health, and areas that are less physical? How does AI help there?
BB: That's an interesting one too. I used to think that the "humanness" piece was the one thing that could never be replaced — that a human would always need to be at the center of that kind of care. But a study published just a couple of months ago looked at responses to patient emails: a patient emails their doctor, the doctor emails back — and in this study, ChatGPT actually wrote some of the responses instead. The AI-written responses were rated better. The AI communicated more, wrote longer and more detailed answers, and came across as warmer and more empathetic than an email likely written by an exhausted physician at the end of a long day, short and a bit curt.
So I'm honestly not sure the humanity advantage is as clear-cut as we assume — computers seem to communicate quite well in some contexts. Eric Topol talks a lot about using AI to passively track things like mental health indicators — respiratory rate, how often someone sighs, mood, affect, whether someone's moving around less than usual — pulling in all these small, continuous data points and drawing conclusions from them. A system could simply alert you, or your doctor, that something subtle is changing, suggesting mood might be worsening or improving. It'd be wonderful to treat mental health with fewer medications if we could build more of these monitoring and communication systems. So I think there's a very large potential role for AI here too.
JM: It's funny you mention that — most of us are already wearing a device on our wrist that's probably using some form of AI to monitor our physical, or even mental, wellbeing. I have a sports watch that tracks something called "body battery," and I can watch it drop throughout the day as I get more tired. I know I feel tired, but now I can actually quantify it, and the watch picks up on more than just that. That gives you the ability to monitor yourself, and potentially share that data with a broader data set or your doctor.
BB: Yeah, I think we're very close to being able to combine all these different sensor inputs for genuine remote monitoring. The technology is there, the computational power is there — I think this is one of the areas where AI is going to play an increasingly large role.
JM: What are the top three things you think we'll see in healthcare over the next couple of years as a result of AI?
BB: The biggest surprise to me, honestly — something I didn't see coming — is AI's ability to take over a huge chunk of administrative work in healthcare. For every minute a doctor spends with a patient, they're probably spending another minute documenting it, filling out forms, recording everything that happened during the visit. I think we're at a point where a lot of that busywork can be offloaded to AI — something simply listening in the exam room, transcribing the conversation between doctor and patient, and documenting it automatically.
This is where Eric Topol has a genuinely optimistic take, maybe the most optimistic in healthcare — his thesis is that if computers take over this administrative burden, we can actually put more humanity back into healthcare: more direct conversation, more hands-on examination. Right now, I think most of us would agree healthcare is fairly broken. It's not really working well for anyone — patients aren't happy, doctors aren't happy, doctors are leaving the profession and even taking their own lives at record rates. The average visit is around seven minutes, and the average patient gets about 20 seconds to speak before being interrupted. That old-fashioned doctor-patient relationship really doesn't exist much anymore. But if AI can offload the administrative load and let doctors and patients spend more real, unhurried time together, maybe AI is our only real path to unbreaking a system we've badly broken.
JM: It's funny — as you were talking, it reminded me of something. I have handwriting-recognition note-taking software on my iPad; it somehow reads my handwriting better than I can and converts it to text. It recently got a new AI summarization feature. The first time I used it, I pulled up notes I'd taken during a conversation with someone who had just resigned.
As often happens in those conversations, she wasn't communicating clearly why she was leaving — it's an awkward conversation, even with someone you consider a friend. I'd shared my written notes afterward with our head of HR, since it matters to us why people leave, and we want to make sure we're doing right by our people. We were both a little puzzled at the time. Later, we spoke with her again, and she told us clearly, and it made total sense — the right decision for her. But afterward, I ran my original notes through the AI summarizer, and neither of us had been able to correctly interpret the real reason from what I'd written — yet the AI's summary matched almost exactly what we eventually learned directly from her in that follow-up conversation.
So beyond just freeing up more time to talk, AI may also help interpret what's actually being said — she was also from a different cultural background, which may have played a role. In this case it didn't matter much, since it was simply a personal opportunity for her, and we weren't discussing her health. But in a medical context, that same kind of interpretation — understanding what a patient is really saying, given how they describe their symptoms or feelings — could genuinely change an outcome. That could be really powerful.
BB: Right — that administrative-offloading piece was the biggest surprise to me, just how quickly systems are being adopted specifically to reduce that burden. The other thing — and this is where some in the industry think we probably won't go, though I'm not sure I fully agree — is AI moving deeper into actual treatment decisions.
These systems already diagnose more accurately in some areas. On a standard X-ray, the human false-negative rate — meaning something is present but missed — is roughly 30%. For computers, that's now down to single digits; they're just much, much better at it.
On the treatment side — U.S. doctors go through training, and we tend to lean on that training and get somewhat locked into what we learned, but it's genuinely impossible to keep up with the entirety of the world's medical literature every single day. AI systems can stay continuously updated on the most effective treatments for specific patient types — by sex, by ethnicity, by side-effect and complication profiles — and start factoring that directly into recommendations.
I remember, on our very first day of medical school, being told something like: over the next four years, we were going to be handed an overwhelming amount of information — literally drinking from a fire hose — and told to just do our best. They said none of us would retain more than 10 or 15% of what we were taught. I don't know if that number was accurate, but the point they were making was that each of us would retain a different 10%, and collectively, as a group, we'd have a shared body of knowledge. Now think about a system that retains everything, all the time, versus each of us individually operating on our own narrow slice. The disparity in capability is enormous. I think our role will increasingly be to sign off, double-check, and confirm — "Yes, this diagnosis makes sense, I'll go explain it to the patient" — but the underlying analysis will be extremely comprehensive. It's going to fundamentally rewrite the job description of what it means to be a doctor. Maybe it becomes more of a concierge role — or like the human sitting in a self-driving car, ready to grab the wheel only if something goes wrong. It'll definitely look very different.
JM: That feels like a normal progression for almost any career. My father, who's passed now, started his career in the 1950s, and his frame of reference was completely different — he'd have entire teams handling things that you and I are stuck doing ourselves in PowerPoint. Roles evolve, and maybe what's left for humans is more the bedside manner a computer can't quite replicate — or maybe it's someone to help you navigate insurance paperwork and hospital bills, which honestly seems like the hardest part of the entire healthcare system.
BB: Actually, one of the coolest uses of AI I've heard of is writing prior-authorization letters to insurance companies. AI can write a genuinely persuasive letter, laying out the potential liability and risk of denying a particular test, and it's just as useful for writing appeals for peer-to-peer reviews. That's something almost every doctor struggles with and isn't particularly good at. I have no doubt ChatGPT can write a far more persuasive letter to an insurance company than most providers can — honestly, that seems like a genuinely great use of the technology.
JM: Absolutely. What's the biggest thing to be worried about with AI in medicine right now?
BB: I think we're all still concerned about current reliability, though I expect that to improve. On my team, every week during standup, someone shares a new example of getting an AI model to hallucinate — coming back with a genuinely ridiculous answer. So for now, I'd expect models to operate with guardrails and human oversight until trust is fully established. Some people argue AI will never be fully trusted in medicine precisely because it isn't reliable or predictable enough, and we can't hand over full decision-making authority.
My personal worry was that AI would make medicine feel colder and more impersonal, stripping out some of the humanity. So far, that hasn't really played out that way. I actually love Topol's optimistic take — that it might add humanity back into healthcare, which would honestly be wonderful.
JM: Brad, not everyone has your background — starting as an MD and now doing incredible work in AI. What was your path to get here?
BB: Honestly, my path was indecision. I liked computer science, and I liked medicine, and I genuinely couldn't choose between them. As an undergrad, I was taking software engineering courses alongside the chemistry and biology courses needed to pass the MCAT. At the last minute, my career counselor told me I should go into medicine, because "this computer stuff isn't going anywhere."
JM: This was in the Pacific Northwest, where everything was headed in exactly the opposite direction.
BB: Right — this was literally across the lake from Microsoft at the time. Anyway, I became fascinated with where the two fields intersected. Even back then, I was really interested in decision analysis and the application of Monte Carlo simulations to medical diagnosis, and the broader role of computers in making diagnostic decisions. There were a few other programmers at medical school too, and we'd build software together on the side, but my intention was always to practice medicine and treat software as a hobby.
After residency, I went into private practice, and eventually started building software for myself — I wanted an electronic medical record system, and nothing really existed at the time, so I built my own. I also built a health risk assessment tool, because I wanted to systematically screen my patients, and a system to track them, since patients don't always call to tell you when something's wrong. We built something that would flag it on a calendar to call and check in if we hadn't heard from someone in a while.
Somewhere along the way, this hobby got out of hand, and other organizations became interested in the software. I ended up starting a company called WellMed, which I eventually sold to two friends who were at WebMD at the time. They acquired the company, and I agreed to stay on for a few years to see where things would go — and I just never got around to going back to practicing medicine. I still say that's the plan, but at this point that may be too far gone.
I talk about this a lot: whenever two industries bump up against each other, there's usually a lot of opportunity right at that seam. In healthcare and technology specifically, that intersection is really interesting, especially when you're building healthcare technology directly. My general goal has been to know more medicine than the other engineers, and more engineering than the other doctors — that creates a role for myself as a kind of translator or conduit between the two teams. Doctors and engineers genuinely think differently, solve problems differently, and approach things differently, so having someone who can translate between the two worlds is a uniquely useful spot to occupy. I really enjoy having a foot in both.
JM: In many ways, you're the healthcare provider of the future, because you've been building the platforms that use AI and technology to deliver better care through other healthcare providers. Looking at what you've built with My Health Match — you're helping patients get to the best available treatment for their specific situation. For those who don't know it, I'd describe it as using data and AI to determine which doctor is genuinely the best fit for a specific procedure or condition a patient needs. Is that a fair way to describe it, or how would you put it?
BB: That's a good way to describe it. There are a number of interesting problems in healthcare, and one of the ones we started out trying to solve was the huge variation in care across different hospitals. Most patients assume either that all hospitals are basically the same, or that there are simply "good" and "bad" hospitals — neither is really true. What actually tends to happen is that a given hospital is excellent at some things and mediocre at others, and vice versa. So part of what we do is understand which hospitals get the best outcomes for which specific procedures — maybe your orthopedic surgery should happen at one hospital, and your cardiac care at a completely different one. The difference can be enormous — sometimes five times the complication rate between Hospital A and Hospital B. Complication and mortality rates are generally low overall, but you still want to put the odds in your favor wherever possible.
It's similar with doctors — depending on their level of specialization, you generally want someone who performs a high volume of the specific procedure you need, and does it well. There's a well-established concept that you need a certain number of cases to become proficient at something, and considerably more to become truly exceptional, and to maintain that level over time. And of course, doctors operate within broader systems of care — someone else handles your pre-op care, someone else handles your recovery. Ideally, you want a doctor who does a lot of exactly what you need, working at a hospital that's also excellent in that specific area.
That's really the core of My Health Match — finding the most experienced provider, finding the best facility, and identifying where those two overlap. Then we layer in other factors, like patient satisfaction — which matters more for some specialties than others. Your brain surgeon doesn't necessarily need great bedside manner, as long as they're excellent at brain surgery. But for your primary care physician, communication matters a lot more. Putting those pieces together is the concept. What's genuinely fun about it is that you can find the best doctor in Atlanta, or across the Southeast, or, for something extremely rare like CIDP, identify the true national experts — sometimes there's a very short list, essentially a top ten in the entire country. You may have to travel to see them, but at least you now have access to that information.
JM: You touched on how doctors communicate — a great example being prior-authorization forms, where AI can help free up time for patient care instead of paperwork. On the flip side, insurance companies can also use AI to review those same letters. Does that create new friction, or does it actually make the whole process more seamless and efficient — faster, or does it just create a stalemate where both sides cancel each other out?
BB: That's a genuine unknown. Like I said, there'll probably be some back-and-forth — one side's technology improves, then the other side catches up, and we land back at a kind of stalemate, over and over. We're already seeing a bit of that. I actually heard Marc Andreessen talking last week about how AI right now is in something of a "race to the bottom," because these large language models are becoming commoditized — anyone can build one, and they're all roughly comparable, so companies are competing without really differentiating. I think we might see something similar here: parts of this process become commoditized and ubiquitous, and some underlying dynamics simply won't improve much, no matter how much the technology advances.
JM: Or maybe it centralizes — the data used to generate the pre-op letter gets fed into some shared decisioning system that both doctors and insurers reference, and they arrive at the same conclusion simultaneously. That could make things faster.
BB: Hopefully we do get better decision-making out of this. Ideally, it's in an insurance company's own interest to get patients treated accurately and quickly the first time — that it's not purely a demand-management or utilization tool, but that they genuinely want the right care to reach the right people. We just need to make sure the right people are actually getting approved for the surgery or treatment they need. But at the end of the day, it does come down to dollars, and insurers remain fairly strict gatekeepers.
JM: It's interesting — if diagnostic accuracy and treatment matching improve enough, the waste insurance companies are trying to eliminate, and treatment happening exactly when and as needed, could actually align into something close to an ideal system.
BB: It's been shown repeatedly that better care actually saves money. You can see the opposite effect happening right now in emergency rooms, where private equity firms are buying up ERs, reducing physician headcount, and increasing reliance on physician assistants or nurse practitioners. On the surface, that looks like cost savings driven by better financial performance. But these mid-level providers often order more tests, and patients end up getting discharged and readmitted at a much higher rate. That actually drives costs up for insurers and patients, even as it likely increases revenue for the hospitals themselves. It's genuinely complicated. Ideally, you're on a fact-finding mission, searching for the right answer and the right decision — but it's always complicated. I think AI could help cut through some of that, at least by pointing toward what's probably the best decision in a given case.
JM: When it comes to AI and the time and communication between HCPs and patients, does this really open the door to more human interaction — more emotional support around injuries, diagnoses, and news none of us are ever really ready to hear?
BB: Potentially, yes — but I think it'll take real effort. A lot has changed in healthcare over the last 10 to 15 years. If you look back, most providers used to be small business owners, self-employed. Now the majority of doctors are employees, working for a larger entity, with managers. The industry created RVUs — relative value units — to measure physician productivity, which is part of why visit times shrank: to hit your productivity numbers, you need to see a certain number of patients per day, order a certain number of tests, and so on. That fundamentally reshaped the entire landscape of healthcare.
So in order to get back to a place where the doctor-patient relationship, and real humanity, are restored, you'd need to adjust those RVU-based productivity targets, and actively promote real communication and connection — giving doctors and patients enough time to actually interact, and enough time to properly examine a patient. You see strange things today, like doctors examining patients right through their shirt — a kind of ceremonial gesture rather than an actual exam. There's very little real hands-on diagnosis happening anymore. My hope is that there's enough dissatisfaction with the current model — because providers don't want to practice this way either, they genuinely want to feel like they're delivering real care — that if AI can be the tool that gets us there, that would be the most optimistic and best possible outcome.
JM: Does AI have the potential to expand care into healthcare deserts and benefit the broader population, or will the benefits mostly accrue to specific groups who already have better access?
BB: I think AI genuinely has the potential to reduce variation in care, because if there's a shared, universal knowledge base that everyone is drawing from, everyone starts working from roughly the same playbook — that's a real opportunity. The equity question is more complicated, though, because of what's sometimes called "healthcare concordance." My team came across this looking at hospital data — certain ethnic groups consistently receive different treatment, or get misdiagnosed more often, but interestingly, that tends to show up between different institutions rather than within the same one. We've actually seen patients drive right past a five-star facility to go to a one-star facility, simply because that's where people like them tend to go, where they feel most comfortable — there's a real social dynamic driving a lot of that. Part of the goal is just making people aware of what the objectively best facility actually is, so they can make that choice for themselves.
JM: Brad, thank you so much for your time today. This has been incredibly insightful.
BB: Thank you — this was fun.
JM: I'm here with Andi Shehu, who leads data science here, and also does research and teaches as an adjunct professor at NYU. Welcome, Andi.
Andi Shehu (AS): Hello, John. Thank you, nice to be here.
JM: What struck me most about what Brad shared is that everything is going to change — we probably don't fully know the extent or the timeline, and there's a lot of debate around that, but change is clearly coming. If even a portion of the impact Brad described is already showing up in the practice of medicine — arguably one of the most complex human endeavors — it seems like AI is going to do remarkable things in how we manage marketing campaigns too. What are the biggest parallels you see between what Brad described and what we're seeing in marketing?
AS: Two things stood out to me. First was the communication between doctor and patient — think of that as an interface: here are the results, what's the best way to actually communicate them to the patient? In that study, the AI-written email was actually preferred over the human, doctor-written one.
The parallel in our world is that we work with a lot of data, and sometimes our main job is really just collecting it, analyzing it, building the charts — but how we communicate that to the end user, whether that's a client or someone else, is where AI can be genuinely powerful. Imagine being a very strong analyst — you have all the data, and you can just feed in some keywords and key demographics, and prompt a chatbot to generate the best way to communicate those results to a specific audience.
That could be a patient you're trying to reach, where maybe the messaging itself isn't quite landing — or it could be a client you're trying to communicate results to.
JM: So it could adjust based on whether the audience is a data scientist like you, an analyst like me — even though we're looking at the same thing, we approach it differently — or a brand-side person who understands marketing and the patient at more of a strategic level, without the analytics background.
AS: Exactly. And people can sound confident because they know what they mean to say, but they're not using precise terminology — and getting the terminology wrong, whether it's how you describe comparing samples or confidence intervals, can genuinely get you into trouble. With AI, you don't have to worry about that as much — you give it the right instructions, and it does a much better job than manually sifting through several reference books to get the messaging right and check all the edge cases.
JM: So the underlying numbers might stay the same, but how they're represented — or more importantly, how they're actually absorbed by whoever's receiving the message — can vary in style and focus, and AI can help tailor that.
AS: Right. You mentioned the MRI example earlier — I actually went through something similar. I had an injury, got an MRI report back, and had no idea what most of it meant. I had to Google individual terms one at a time. Now you can imagine having a couple of report formats — the standard clinical report, and then a second version specifically designed to communicate that same information clearly to the patient. Same underlying science, but written in a way that actually gets the message across.
JM: And that matters whether it's someone's health, a marketing campaign, or even communicating with finance.
AS: Exactly — or even on the creative side. Say a brand is launching a new drug and trying to craft the right message. What's the best way to communicate it? Maybe they're missing something specific to a demographic or a region. With AI, it becomes much easier to generate many different creative variations at once — something tailored for the Northeast, something else for the Southwest. It really lowers the barrier to producing that range of tailored messages.
JM: And with one-to-one messaging, it can adapt based on the individual — geography, or whatever else turns out to be predictive, demographically.
AS: Exactly, exactly.
JM: Beyond communications, what else really resonated with you from what Brad shared?
AS: The diagnostic piece really stood out, and it applies directly to our own work too. We run a campaign, collect a huge amount of data, and then need to diagnose what's actually happening — what's working, what isn't, why, and what's actionable from that. We might run a six-month campaign, collect all that data, go back to the client, and try to optimize. What's the best way to actually process all that information?
What Brad mentioned about diagnosing X-rays is a great parallel — it's already been established that AI can diagnose X-rays more accurately than human doctors. Just a few days ago, a study was published in JAMA comparing diagnostic accuracy across three groups: doctors alone, doctors using a chatbot AI as an assistant, and the chatbot AI alone. Doctors using the chatbot performed slightly better than doctors alone — only by about 3 to 4%. But the chatbot alone, with no human involvement at all, outperformed both groups by roughly 14%, landing around 90% accuracy. That's genuinely fascinating.
JM: So at this point, computers are already outperforming us.
AS: Yes — and interestingly, when the computer produces an output, human intervention often actually makes it worse. The computer says, "this is the correct result," and then the human overrides it, saying, "No, I want B, because that's my opinion, based on my own experience." If the human is the last one to override the decision, accuracy actually drops.
JM: So the same thing is likely happening with our own communications around a campaign, budgeting, or specific micro-level decisions. What struck me most was the X-ray example — computers analyzing individual pixels, while as an analytics person, I'm looking at campaigns with tens of thousands, sometimes hundreds of thousands, of data variations. I can only really process a fraction of that, which means the vast majority of the fine-grained detail simply never gets accounted for — arguably it gets ignored. This could solve for that too.
AS: We already use models where, similarly, we don't fully understand exactly how they arrive at their output — but they still perform better than models we do fully understand. At some point, you just have to let go of needing to understand every mechanism, and focus instead on the goal: which model is performing better. Too much human interference can actually lower accuracy — if you're constantly tweaking a live campaign instead of letting it run for, say, 30 days to collect enough data, you might actually be reducing overall performance. In effect, you become the "doctor" who keeps stepping in and inadvertently making the results worse.
JM: And presumably AI can also predict, in advance, what the outcome would be if you insist on making a manual change — so even if you go ahead with a human override, it can tell you upfront, "This will likely reduce your performance by X."
AS: That's a great point — that's actually where tool design comes in. As data scientists and marketers, we need to think about how we build these interfaces. At the end of the day, a human is still going to make the final call on things like budget or campaign duration, so it's worth thinking carefully about how we design these tools and what features we build into that decision-making interface.
JM: What other parallels resonated for you, from HCP-to-patient communication over to marketing-service-provider-to-brand or marketer?
AS: I'd say those are the two biggest ones. The underlying theme, though, is that new AI tools are entering the market constantly, and there's a real gap in how quickly we learn to use them properly. In the diagnostic example, instead of treating the AI like a search engine, the better approach is to feed it the full data set directly and let it generate the diagnostic report itself. There's an educational component here — how do we, as practitioners, learn these tools as they emerge, and then communicate their proper use to our clients? Because yes, they're just tools, and yes, they should make campaign optimization easier — but we still need to genuinely understand each tool well enough to use it properly.
JM: As someone who does this professionally and is also active at NYU working with students — and speaking as someone starting to feel like I'm falling behind, watching people years younger than me seem to know more about everything — what would you recommend to someone just starting their career, versus someone more experienced who simply hasn't incorporated AI yet? How do we all stay ahead of this?
AS: That's a great question. I actually come from academia myself — my background is physics, and my PhD was in quantum mechanics. Shifting from theoretical physics into AI meant going through some of those same learning phases myself.
For students, there's a lot of debate right now, at NYU and at many universities, about students using ChatGPT to write essays, generate code, or solve homework problems. At what point is that still learning, and at what point are they just using a tool — or do they need to learn an entirely new skill set instead? I use AI to generate code myself. Coding keeps evolving — we used to write everything in C++; very few people in data science write raw C++ anymore. Now we use wrappers like Python and R, and the next "wrapper," in a sense, is going to be these AI chatbots.
For students, I think the right approach is to embrace it. For working professionals, the best way to learn is by doing — whenever there's a new task, set aside some time to just test the tool out, because you'll often be surprised. First figure out what it's genuinely good at, and let it handle those things. Then figure out what it's not good at — that's where human focus should go. The same logic applies to doctors and to marketers: AI will be very good at certain things, and once you hit its limits, that's where you, as a company or a brand, should concentrate your own effort.
JM: And I think those limits keep shifting — what it was good at six months ago isn't where the limits are today. Until things plateau, it feels like continuing medical education for doctors — as understanding of conditions grows, doctors have to keep learning. I think it's the same with AI: if you went deep on it six months ago, you'd basically need to relearn it now, because six months in AI terms feels like a lifetime.
AS: Exactly — and that accelerating pace is actually the exciting part. We don't want things to stall; we want them to accelerate, to keep improving. Keep evaluating new models, keep evaluating how you're running campaigns, how you're diagnosing issues, how you're writing code. Even for me personally, whenever I'm analyzing a data set, I have to keep checking what new models have come out that might improve accuracy.
JM: I honestly suspect that in ten years we won't even be talking about "AI" as a distinct thing — it'll just be another everyday tool, woven into daily life, not something remarkable in itself.
AS: I agree completely. One analogy I like: twenty years ago, taking your car to a mechanic meant a lot of manual diagnostic labor. Now, more and more, the "diagnosis" is just plugging in a piece of software that tells you exactly what's wrong. I think we're heading toward that same place with a lot of this — we'll just accept that it works, it does the job, and we move on to more interesting problems.
JM: One thing Brad raised was concern that AI might strip the humanity out of healthcare. Do you think AI takes the humanity out of the things we do — whether that's healthcare, marketing, trading stocks, or anything else?
AS: Maybe not — it could actually go the other way. Some of the technology we've already built isn't especially human-friendly. Even email isn't particularly human-friendly — you probably wouldn't want to email your grandparents the way you email a colleague. In teaching, for example, instead of spending hours grading homework, if AI handles that, I get to spend more actual face time with students. Doctors might get more real face time with patients instead of writing endless reports and fighting with insurance companies. So it could genuinely go the other way — the technology we've built so far just hasn't been especially human-friendly, and this might be our chance to build technology that actually is.
JM: I love that. And maybe humans have an incredible capacity for great things and kindness, but we also have some real, persistent flaws — and maybe those flaws are exactly what can be reduced, letting us be the best version of ourselves while getting past whatever those limitations are, whether that's oversight failures, bias, or something else.
AS: Right, and you can actually build those limits directly into the model. One last thing worth mentioning — if AI takes on all the manual labor nobody really wants to do — emailing, fighting with insurance, endless hours sifting through data from every direction — that frees up an enormous amount of untapped human potential. Doctors could spend more time on actual research, developing new treatments. That part genuinely excites me — whether it's new research, new drugs, or new treatments, that potential is really exciting.
JM: Or simply spending more time with patients, which on its own is a wonderful thing.
AS: Exactly, exactly.
JM: Andi, thank you so much for joining us today — your perspective on AI is genuinely refreshing for those of us who aren't data scientists.
AS: Thanks for having me, John — this was exciting, and I'm genuinely excited about the future.
JM: I want to thank Brad Bowman for joining us today, and Andi for his great insights. And thank you for listening. That was Deep.
Announcer (Voiceover): You've been listening to Deep, the health marketing podcast. Deep is a presentation of DeepIntent. Opinions shared by guests represent their own perspectives, not the views of their company or organization. If you'd like to learn more about the guests, the show, or the topics discussed, check out this episode's show notes in your podcast app, or visit deeppodcast.com. If you have a question or a suggestion for the podcast, drop us a note at podcast@deepintent.com. If you like the show, please leave a comment and give us a five-star rating on your favorite listening platform, and be sure to subscribe so you never miss an episode.
The Deep Podcast features original music by Diaphonic. The show is produced by Robert Haskett, with Ben Abramowitz, and hosted by John Mangano. Thanks for joining us.
Links
Relevant Blog Posts:
Advertising Data and AI Is Fueling Better Outcomes for Patients and Brands
