Much of the AI used in health care today is tied up in helping the industry manage its revenue streams — think billing and coding assistance. But Dr. Toyin Ajayi knows there’s more AI can do to help manage the health of patients.
On this episode of The Dose podcast, Dr. Joel Bervell talks with Dr. Ajayi, a board-certified family physician and CEO of the primary care provider Cityblock Health, about the ways she is using AI to make meaningful differences in the health of the most underserved patients. Dr. Ajayi’s data-driven, outcomes-based care model has already demonstrated reductions in emergency room visits and higher patient engagement — as well as lower overall costs. Now, she's using AI to scale it.
“Technology, if done right, is one of the very few deflationary forces we can apply to health care costs right now.”
Transcript
JOEL BERVELL: My guest on this episode of The Dose is living in the future, and so are her patients. Dr. Toyin Ajayi is a board-certified family medicine physician and CEO of Cityblock Health, and she’s here for a conversation about how AI is improving the lives of the most underserved patients, those she focuses on exclusively in her clinics, and how everything from data sets to physician practices and patient outcomes can be improved with AI.
The AI hype is real, but this is not that, and it’s not charity work. Dr. Ajayi’s patients are what’s called dually eligible. That means enrolled in both Medicare and Medicaid, so they’re typically older, low income, and living with chronic disease. As a physician and visionary health care entrepreneur, Dr. Ajayi created an entire ecosystem of clinics for people in these circumstances, and it’s profitable and growing.
Dr. Ajayi is an Aspen Institute Fellow, and a member of the National Academy of Medicine. She understands the moment and has seized it for good.
Dr. Ajayi, thank you so much for joining me.
TOYIN AJAYI: Thanks for having me. It’s nice to be with you.
JOEL BERVELL: As I touched on briefly in the intro, your foundational idea launching Cityblock was to close a gap in care. Your mission is to serve people who are not accessing care at all or those most in need of better health care. And you had this insight, in your words, it’s within Medicaid that AI can make the most meaningful material differences in health and engagement. That’s from a paper you published this spring, and we’ll talk more about that. But I want to begin with this. What was the gap that you saw? And then we’ll get to how tech, and specifically AI, is accelerating your model’s reach and potential.
TOYIN AJAYI: I think the gap is very easy and obvious to spot for anybody who’s accessed the health care system, which is it’s a reflection of the way that our system is organized and it comes all the way from the incentives at the top. Our health care system is reactive by definition, and that is true in the acute care setting as true as it is in the primary care setting. We have historically deployed health care in a way that requires people to come to us to receive it and that reimburses physicians and health systems and practices for the care they provide, irrespective of the outcomes that it delivers.
So if I have a primary care panel, I have 30 patients to see on my schedule, if half of them don’t show up, I just focus on the ones who do show up. And I’m not incentivized actually in most traditional primary care models to actually think about the people who didn’t come, not only the people who aren’t on my schedule for the day, but the people I hadn’t seen in six months or 12 months or 18 months who might be in the most need of care.
And so we sort of ration and deliver care in a first-come, first-served basis. And the same is true for the acute care setting. If someone comes in with cellulitis or pneumonia or concern for a stroke, we do what we can in the acute care setting to patch them up and then we send them out. And if they happen to come back in in 30 days, as sometimes up to 20 percent of patients who are discharged from the acute care setting do, we just do the same thing again and we bill for it. Not surprisingly, that does not lead to better outcomes for a population of people. And it particularly doesn’t lead to better engagement, better quality, better experience for people who are likely to face barriers to accessing care. And those barriers may be transportation, childcare, financial resources. They may be bigger than that, existential, historical trauma and mistrust of the health care system. All of these things mean that the folks who often are most likely to benefit from really high-quality, coordinated, tech-enabled health care are not the ones who might necessarily show up in front of a clinician when they need it.
And so the gap that we sought to close was exactly that. How do we build a responsive, proactive, data-driven health care system that focuses on finding people who need care and bringing it to them wherever they may be? And in so doing improves everything we’re looking for: cost, quality, experience, and outcomes.
JOEL BERVELL: Absolutely. And just to even rewind a little bit, about nine years ago, you and your cofounders launched the company after raising nearly $500 million. And on paper the business doesn’t seem like the type of idea that people would be calling sexy. So what made the pitch so compelling? Was it simply the value proposition? Was it that no one else was doing it? What was it that investors immediately saw?
TOYIN AJAYI: It’s a bit of both. I mean, health care is, in most parts of the country, the biggest employer, and in most states it is the single largest line item. It’s Medicaid. We spend over a trillion dollars on health care through Medicare and Medicaid. It is a massive, massive, massive industry and market, and we get quite poor outcomes. And when you think about the places where venture capital and where entrepreneurship are best suited to impact, it seemed like a place ripe for disruption. The business model that we deploy — value-based care, outcomes-based care — was not something new. We didn’t make that up, but we were the first to really apply it at scale to Medicaid. And I think that was what made it compelling. That there’s this whole segment of the population with high costs, high needs. You have a payer source, so you’ve got a customer built in. The customer has real pain points. We can all see them and feel them, and the real opportunity to disrupt with technology to get into that big, big market segment.
JOEL BERVELL: Absolutely. And you envisioned and built to transform the system where the incentive, as you mentioned, is outcomes, not units of care anymore. This spring, you wrote and posted a paper where you talked about your belief that Medicaid plus AI will set a new standard for innovation. I love that optimism. Can you share some of the applications that you’re implementing, iterating on, experimenting with that you’re convinced are going to make a difference? And maybe we can start with what happens for patients. I know that you see AI as a positive for relationship- and trust-building, which I think in some ways may seem counterintuitive to a lot of the listeners.
TOYIN AJAYI: Yeah. You talked about trust. What is so important is throughout all of this experience we’ve learned that if we can engender trust and real engagement with our members, we earn the right to have them share more information about what matters to them that we can then use to better provide care to them.
We’ve spent a lot of time serving our members and our patients, really trying to understand what creates trust. And the truth is that it’s as much about the humans as it is about the systems. What people tell us is that often when their trust has been broken in health care, it’s because the system broke, not because a person didn’t care. If a person says, “I’m going to make the referral to you to see an oncologist,” and then the referral never comes through. If they say, “I’m going to call you in four days and follow up to check on how you’re doing with your new medication,” and they don’t call. If they say, “You can call me anytime,” and then you ring and ring and ring and no one answers, these are things that break down trust. So how do we create trustworthy systems? Technology allows us to do that much, much, much more effectively and at much more scale so that we can make sure that every call gets answered, every task gets completed, every follow-up gets done, every loop gets closed. We use tech to do that.
Oftentimes, when I was in active primary care, I get a fax maybe a month later that told me my patient was in the emergency room. I had no idea in real time what was happening. And so solve the data problem, which we have solved at Cityblock because we’ve been doing this for so long, which is to make sure that we’ve got as much data as possible, as real time as possible on every single one of the hundreds of thousands of people that we serve at any given point in time. How do we integrate all of those real-time data sources and use it to help us actually identify who’s the most likely to need our help and what kind of help they might need?
So AI, incredible right? These are all really important, nuanced things we can optimize for with AI that help us understand truly, truly, truly at the patient level what is the right next best action for each person? And historically, again, in traditional primary care, we have very few tools. Bring them in, let them see a doctor. Maybe have the practice nurse call or the MA call if it’s something that requires a lab draw or some follow-up. That’s kinda it.
Well, in our toolkit, we’ve got a lot more tools. We can send you to a therapist. We can have a therapist come to you. We can have a pharmacist work with you or your doctors to manage your meds. We can send a nurse to your home to reconcile your medications. We can provide coaching. We can do video visits. We can do in-person visits. We can send paramedics and EMTs to get labs in your home. There’s a whole host of tools and deciding what the right next best action is and who to deploy to do that action. Also, massive optimization problem that AI can help tremendously with.
And then finally, there’s an access problem for many populations, which is we just don’t have enough clinicians, particularly in primary care. And we certainly don’t have enough nonclinicians to help with the navigation and the coordination issues that people have. “My electric wheelchair broke down. What do I do? Who do I call? How do I make sure they’re going to follow up? This is essential to my livelihood.” We can use AI agents to help triage those types of concerns, and in many cases, to resolve those concerns using agentic voice modalities that sound and feel like a real person actually, but are always on, infinitely patient, and have full access to all the information necessary.
JOEL BERVELL: As you’re talking, there’s so many touchpoints that I literally see in my day-to-day every single time a patient comes in. I’m curious about what you’ve been able to measure so far. Have you seen reductions in hospitalizations, in the ER visits, costs, and improvements in patient engagement that make you confident that a model like this using AI to fill in these gaps and narrow them is working?
TOYIN AJAYI: Yeah, I think the reason why I’m so bullish and so excited about the implications of AI is because we’ve been at this for nine years. And so we built a model that has allowed us to gather and iterate on a ton of data and to prove out that these things work. So now the task for us is to say, okay, so do that but better and more and at more scale with higher velocity, with even better follow through using AI and technology. And so we’ve been able to start to see the proof points in terms of the amount of access we’re able to offer to people, the fidelity to our model that we’re able to maintain while using more AI and more technology to allow us to be more efficient, and the ability to scale into different populations to serve a broader group of people.
So that’s on top of a base of a model that has already demonstrated reductions in acute utilization, improvements in quality, much, much higher engagement, incredibly high patient satisfaction and retention, and overall ultimately reductions in total cost of care.
JOEL BERVELL: And AI, of course, works after staff go home for the day. How do you see that helping practices with patient compliance, outcomes, cost containment, et cetera?
TOYIN AJAYI: Yeah, I think hugely. I mean, part of what AI enables us to do is to say, how do we improve costs without restricting access? How do we improve costs while expanding access? It is actually definitionally one of the very few deflationary forces that we can apply to health care costs right now is technology, if done right. And so it is just such an incredible golden opportunity at a time when we are desperately in need of solutions to help make sure that we can continue to provide the care that people need.
As we see a population that is aging, a population that is more sick, more disabled, more dependent on public supports and resources, the answer is not going to be we just keep cutting and cutting and cutting access in order to make the budget work. We have to improve and expand access while also being really, really pragmatic fiduciaries of public resources. And that’s what technology allows us to do. It’s not enough to just have hope that this time, this massive revolution in technology will actually help people and not just create more wealth that gets circulated in a very small segment of the population.
JOEL BERVELL: And it’s a key in what you said there, if done right. I’m curious, where has AI not lived up to the hype? Or what have you tried that simply didn’t work or taught you an important lesson?
TOYIN AJAYI: Yeah, it’s so early. I think that’s the thing we have to remember. All of us who are doing this work are experimenting, and the stakes are different depending on what part of the value chain you’re experimenting in. We’ve been really fortunate to have a really great group of patients who spend time with us in our advisory council. So they actually give us feedback before we deploy anything that’s going to be patient-facing. We want to make sure that we get really solid feedback from them about what they need and want. And I’m always really pleasantly surprised. Our population tells us that they have a lot of smartphone access. They’re actually very open to using technology and to interacting with technology, but they’re sensitive to the ways in which historical biases subtly and unsubtly can be baked into everything about the experience.
And so even things as subtle as a turn of phrase, if a voice agent says, “Gosh, I can see how that would be really concerning to you.” Well, to some people that sounds patronizing. To others, it sounds empathetic, but to some people it sounds patronizing. And so being really careful about the ways in which we train these models.
More broadly, I think in our health care ecosystem, I think we’re seeing a lot of, it’s early innings, but the most initial use cases tend to be actually the opposite of what I described. They tend to be things that help with revenue cycle, billing and coding. It’s arbitrage around the existing system. It’s health systems trying to get paid more money for doing the same things that they did yesterday for the same people. It is payers trying to manage spend in the same ways that they have historically.
Those are obvious use cases because that’s the economic engine of our health care infrastructure. But my hope is that as we get past these early innings of AI adoption, that the use cases become more expansive and more geared towards that optimistic vision that we can actually expand access, improve outcomes, improve quality, and reduce cost. And that’s the place where I think the magic lies and where the real opportunity is. I just haven’t seen as many folks with the right set of incentives today and an appetite for innovation spend a ton of time and resources outside of us and a few folks like us who have those incentives in place really trying to do that work now.
JOEL BERVELL: I know Cityblock, particularly that the initial cluster of clinics, was mostly in the Northeast, but now you’ve expanded in North Carolina and Virginia and you’re planning to continue to grow in different rural areas. Why is reaching people in rural areas a priority strategically for the business? What challenges do you anticipate in that rollout?
TOYIN AJAYI: Yeah, I mean, the rural underserved populations are such a big part of a mission-aligned population for us to serve. We’ve been really focused on people who are underserved, and that exists in urban spaces for sure, but we’re seeing increasingly that access to care in rural parts of the country is just as challenged. And when we looked at our data and looked at our model, what we found was that today we serve a really meaningful portion of people who are outside of urban areas and increasingly rural geos. And that the same premise around our model applies there too, which is that we can reach people and engage them, build trust with them, and then provide them access to a much more expansive care model using tools and technology.
And so it’s part of our mission and our mandate to say, if we’ve got the capability to serve people who need it desperately, who are so left behind, why would we not deploy that? And so we’ve been increasingly focusing on scaling into rural populations and thinking about the rural underserved as an important stakeholder as we think about who we can help and how we can help them.
JOEL BERVELL: Absolutely. A lot of the work that I do outside of this podcast is around disparities. So I really want to dig in on that point that you made. As you alluded to, AI models are only as good as the data that they’re trained on. And if health care data already reflects decades of disparities, which your patients, as you mentioned, are seeing, how do you prevent AI from simply scaling those same inequities?
TOYIN AJAYI: Yeah, I think it’s such an important question. And actually for many people, it’s been a big barrier for folks wanting to innovate in this space. And I think understandably so, health care does not have a great track record of experimentation and innovation when applied to underserved populations. So I totally hear that. The truth is though, is that bias exists throughout our health care system, and in human form, it is very hard to audit it. It’s subtle, and you can see it in the data, but in real time, very hard. That nurse who maybe doesn’t check in quite as frequently on that one patient, or the language that seeps into their medical record — noncompliant, unadherent, disheveled. What does that mean? And how does that show up in the way that people who receive that person then care for them?
What is so interesting about technology, and I think so powerful, is it’s auditable. The record is there. And if we know what to look for and we have data sets that allow us to actually compare apples to apples and do evals and have humans then review what the outputs are, we’re much better positioned actually to train it out of a model than we are to train it out of people. So we actually have an opportunity with technology to train things with a really, really tight closed feedback loop. That’s incredible. And so if you’re diligent and you’re data-driven, I actually think it creates a real opportunity to train something different in a way that human and analog models are much more resistant to that evolution.
JOEL BERVELL: I love that perspective of being able to really see the data and be able to train it out in a way that we often can’t see in human interactions every single day.
I want to jump over to the tools for doctors now. I’m in my second year of residency, recently started it, and really see . . . It’s so fascinating because you see the system in a different way and you see the brokenness when you live within a system every single day and see where it breaks down for patients. I’m curious, what tools are doctors using that you’ve seen and loving? And then I’ll share what’s on my wishlist a little bit too.
TOYIN AJAYI: I was like, I know this is your podcast, you should ask all the questions, but I am dying to know, if you wouldn’t mind sharing, what your experience has been with AI and with technology as a resident, and what the stance is in general in academic settings for you right now. And then I’m happy to answer your question, but I’m so curious, it’s too good an opportunity to pass up.
JOEL BERVELL: Yeah. What I’ll say is there’s no standardization around AI right now. And so I think some programs have even taken the stance of don’t use it at all. Some are like, use it within discretion. Mine is in the middle ground. And one of the things I always say is not all AI is created equal. And so depending on what you source information from. I’m a big proponent of things like open evidence where it sources from actual the New England Journal of Medicine and JAMA. We’re using it to find the most recent up-to-date literature that’s out there about specific conditions. Two, obviously the time-saving. So when it comes to scribing, I think that’s something that’s nice. We are not allowed to use audio scribing so that it listens to a patient and automatically writes it, but we can use it for synthesizing a note really quickly or making it a little bit better or taking our own notes and then making it into a more coherent way of understanding for someone else who picks it up afterwards.
Another big one is, let’s say you go see a patient, we have our differential, we’re talking, and then we’re like, okay, let’s actually see what are we missing now? What have we not thought about that could be really interesting from these additional labs that we got? Never using it to replace clinical judgment, but always using it to add onto it.
TOYIN AJAYI: I love that. I love that. Especially the last use case, all of those things are a hypothesis generation. All those things are so ripe for technology. And then there’s the next best action support. So that’s the place where today almost 80 percent of U.S. physicians are using open evidence. A lot of docs are going to tools to say, “Help me figure out what to do next or what’s going on for this patient.” And open evidence is a really amazing tool, but how much better would it be if you had something like that was embedded directly in the EHR [electronic health record] that pulled from that patient’s story that you could chat with like you wouldn’t with Claude or something? You’re actually just querying it, having a conversation, hypothesizing, looking for data to support this next step, asking about what the right next investigation might be. It automatically ingests and so recognizes when the lab comes back, when the stress test results come back, what’s the next best action?
We can build those types of tools directly into a record that are assistive of clinicians as they generate hypotheses and as they figure out what the right next action is, and then help them take those on and then update continually as the data gets in there. So those are some of the things that we’re thinking about.
JOEL BERVELL: Absolutely. How would you describe the adoption rate and conversation in this sector? The moment seems to suggest there’s some urgency. There was about, I mean, one trillion in cutbacks to health care that are coming as a result of the OBDV Act, or what the Trump administration calls the Big Beautiful Bill. And already millions of Americans lack health insurance or now face losing insurance. Is this threat accelerating uptake?
TOYIN AJAYI: That’s a good question. I think there’s a broad recognition that we have a massive opportunity to solve. I certainly have been very heartened in my conversations with folks in CMS [the Centers for Medicare and Medicaid Services] and HHS [the U.S. Department of Health and Human Services] who I think very much have seized on the notion that technology can be a big part of the solution here. We’re seeing CMMI, so the Center for Medicare and Medicaid Innovation, pushing out models that are really focused on technology-driven solutions. We’re seeing Health and Human Services and CMS setting up specific offices focused on technology and AI. So I think there’s real recognition that this is going to be part of a solution.
I think the industry has yet to adopt as quickly as it ought to, partly because, again, I think it’s not inherently good or bad. It’s a tool. And therefore, what happens with AI is as much about the wielder of the tool. And I fear that because we are still layering technology on what is essentially a broken system as it pertains to the incentives, we’re not going to get the best use cases. And until we see broad adoption of value-based or outcomes-based models that mean that the incentives are to use this tool to improve quality, improve outcomes, improve access, and reduce cost, we’re going to see more of the same. AI is just going to enable us to do the things we were doing before better. And the things we’re doing before weren’t working in terms of actually building a healthy and sustainable health care system.
We spend more money in the United States on health care per capita for worse outcomes than any other developed country, more disparities, poor outcomes writ large. And I think AI could help us do that better, but that’s not what we’re trying to do here. We’re trying to do something else. And to do that, we’ve got to change the incentives alongside deploying technology. Technology alone will not save us.
JOEL BERVELL: I want to end with this question here, and it’s an extension of what you’re just saying right now. But if we were to sit down again in 2035, so 10 years or nine years from now, I guess, what would you hope has fundamentally changed about that experience of being a patient and a physician? And what part of medicine do you hope remains deeply human no matter how advanced AI becomes?
TOYIN AJAYI: For folks on Medicaid, what I would love to see is every single person on Medicaid by 2035, sooner even, have a phone number that they can call or an app they can pull up where they can get personalized, truly, truly precision coaching and support for their health and well-being. So that means they can ask all the questions. “I’m going to Wendy’s to celebrate my birthday. My doctor says I should be on a diabetic diet. What should I get off the menu?” “I’ve got an oncology appointment on Tuesday, I think. Can you remind me what date it is? How do I get there?” “My wheelchair broke. Who do I talk to?” Or, “I’m having chest pains. I’m feeling dizzy. I’m lonesome and I think I might hurt myself.” All of these are things that can be triaged and provide real-time access to using AI. And I would hope that every single person has at their fingertips access to health care that looks and feels like it was tailored for them because it is embedded in their record, it has context and knowledge and recall, and it has the ability to, in real time, get them to a human when they need to be in front of a human. That would be transformational.
When they reach a human, I would hope that that human has at their fingertips access to the best possible research and evidence to support the best possible decision-making about what that right next best action is for that patient in front of them. Whether it is a social need, we should get this person housing because their substance use disorder and their serious mental illness and their homelessness are all intertwined and actually housing is going to be the most important solution for them. All the way through to we should trial this drug and not that drug because it’s more likely to be more efficacious for them and better tolerated. I would like for every clinician, no matter where they are, to have access to that best possible tools and technology so that they can use that to benefit their patients.
Then where I think we must remain deeply human is in two places. And actually one place may be surprising to you. I think we’re going to have to think very differently about how do we build technology tools because so much of the trust and the relationship is actually going to be embedded and encoded in technology. And so the human aspect of who builds technology and how they build it is going to be as important as the human at the bedside. And we have to really think about infusing that humanity in product build in a way that we’ve not done before. So that’s one place.
The other place is in our particularly vulnerable populations. So our frail seniors, folks who are going to have challenges actually coming in to see clinicians or using technology. I think children, I think pregnant women, postpartum families, that’s places where we just hands-on need to continue to have that human touch. But I think so much of that human touch can be enabled and will be enabled by also very human technology. And the ability to scale trust in that high-touch and high-tech way is going to be really, really important to how we deploy the health care system in the future.
JOEL BERVELL: Absolutely. Well, Dr. Ajayi, I want to say thank you so much. This has been a fascinating conversation. I think so many people hear AI in health care, and immediately they think about robots replacing doctors. But I think what you’ve described is something very different, which is using technology to strengthen relationships, expand access, and then building a health care system that works better for the people who have been left out of that. And sometimes the data can help us see where that’s actually happened and those touch points. So thank you for sharing your vision, your honesty, and the work that you’re leading at Cityblock. It’s been a real pleasure having you here.
TOYIN AJAYI: Thank you. It’s been wonderful talking with you. Thanks so much for having me.
JOEL BERVELL: This episode of The Dose was produced by Jody Becker, Richard Shapiro, and Naomi Leibowitz. Special thanks to Barry Scholl for editorial support, Jen Wilson and Rose Wong for art and design, and Paul Frame for web support. Our theme music is “Arizona Moon” by Blue Dot Sessions. If you want to check us out online, visit thedose.show. There you’ll be able to learn more about today’s episode and explore other resources. That’s it for The Dose. I’m Dr. Joel Bervell, and thank you for listening.
