U.S. medicine is among the most advanced in the world, fueling breakthroughs in precision gene therapies, robotic surgeries, cutting-edge devices, organ transplants and more. But healthcare infrastructure lags woefully behind. Patients face long waits and steep price tags that can make even the most basic care feel out of reach.
If you ask Jonathan Kolstad, a health economist and the co-founder and CEO of Knit Health, this is a matching problem.
The issue isn’t the system’s capacity or capabilities. It’s the system’s inability to get the right patient to the right care at the right time, Kolstad told Healthcare Dive in an interview at the HLTH conference in Las Vegas.
Knit is building a healthcare world model — an AI that understands the complex dynamics of the clinical ecosystem — on the bet that there’s a big market for a system that can predict patients’ needs and direct their medical journey.
The startup, which was launched out of the University of California, Berkeley, last year, announced a $35 million Series A round on Monday — financing that will accelerate the commercial deployment of its AI model, Med-BEAM, which has been trained on a whopping 140 million patient records.
Healthcare Dive sat down with Kolstad at HLTH to talk about why healthcare needs world models, hospital demand for the technology and how Knit is differentiated in the competitive AI space.
Editor’s note: This interview has been edited for clarity and brevity.
HEALTHCARE DIVE: Tell me about Knit Health. What problem are you trying to solve?

JONATHAN KOLSTAD: Most of the knowledge in medicine is not written down anywhere. It’s in real clinicians’ heads from treating thousands of patients, and it’s in the realities of the healthcare system. It’s not any one Dr. House doing everything. How and when you navigate through the system is ultimately what determines your healthcare. And if you take that perspective, you come to AI from a very different angle. This looks a lot more like a problem of intelligently navigating through a space. How and when do I want referrals to specialists? How and when do I want to stay with the primary care provider? When should I go to step-down physical therapy? What treatments are going to work for me, and how will that unfold over time? All of those problems are much better conceived as a model of navigation through a space — a world model.
What is a world model?
World models are typically thought of in a physical space — navigating a self-driving car through the streets, for example. What’s doing that is not an LLM looking at the driver’s handbook. It’s a world model of, what would a driver do with the steering wheel given all these things in front of it? Should I brake? Should I turn? So this is a fundamentally different approach to AI in healthcare — it’s a world model of the entire healthcare system. And this, in conjunction with LLMs — which can do written language, which have the clinical context — they’re much more than the sum of their parts. We’ve seen it deployed directly by itself, but also in conjunction with LLMs as the infrastructure layer, making LLMs much more efficient and safer.
How do you train this model if the information you need is locked inside clinicians’ heads?
It turns out you can, in a sense, clone a doctor, or lots of collective doctors. Back to the self-driving car example — when you train those models, what they do is they basically look at, here’s all the stimulus in front of you from the radar, the lidar, everything else. And what did the human do? That’s nesting an incredible amount of driving experience, right? A rich understanding of what you should do next. We’ve done that for clinicians.
You need billions and billions of real decisions in response to real patients. So we initially built this using data from Berkeley, showing it could learn these paths. We asked, how do we get the sequences of clinical decisions given all the information available to that doctor at that point in time. And the EMR is an amazing source for that — the ubiquity of electronic medical records means that we’re measuring almost every single action that takes place.
We have the largest electronic medical record dataset in the world that anyone’s building AI on. So that was a focus: how to scale that data to learn what’s in clinicians’ heads. And when the model is learning, it’s similar to the way it learns the semantic structure of language, right? Like, ‘if someone wrote all these words in this paragraph, they’re going to write this next paragraph’. We’re learning the semantic structure of medicine.
A few companies are building similar models — Mayo Clinic, for one, in tandem with Microsoft. What differentiates Med-BEAM from these other efforts?
There’s a lot of announcements around it. What’s really unique about us is we are the first to actually do it by some measure. We’ve been working on it a lot longer than any of these other teams. But most importantly, we have that scale of data to build this massive foundation model at a scale beyond anything you’d ever get at Mayo.
Mayo, candidly, probably doesn’t have the right scale. You need a lot of data to make this happen. And then when we partner with a health system, we’re able to actually learn that health system’s specific clinical practice pattern. So it’s both the scale and the focus and the granularity, the customization that Knit can do, which is totally unique relative to anything else out there.
There are some more niche domain competitors that are doing, say, surgical world models, for a physical space. Epic has their Cosmos suite, a Curiosity suite that is building sequence models on EMR data. That’s the closest to us. They’re building sequences of actions. I’m not sure whether they’re calling it a world model these days.
Med-BEAM can specialize based on a health system’s specific patterns. But what if Hospital A has better triaging or outcomes than Hospital B? Wouldn’t Hospital B want to have insights surfaced about how to operate better, instead of just predictions based on its own data?
Totally. You’re getting to the next thing that’s unique to Knit. Today, I can understand your system as it is. But where we’re going is the ability to behaviorally clone the best performers for that patient. So what Knit is building now is a new form of causal personalized evidence that we have never had before. If you go to your clinician with a set of symptoms, you want to know what the best doctor for your particular situation at that moment in time would do for you next. And that’s hugely informative.
How does this work in practice?
There’s a lot of applications. If I can get the patient to the right person for them, that’s already a massive unlock. I don’t actually need to go into the EMR and tell the doctor, ‘Actually, my AI knows what to do better than you’. Ultimately, we may get there, we may not get there, right? There’s different ways in which that can or cannot be valuable.
Today, Knit’s focus is on the observation that the vast majority of patients are not even getting to the right place at the right time in the first place. The way we’re deploying the model is to solve that matching problem. The core of healthcare is the most complicated matching problem in the whole economy.
This seems like it would be most useful if leveraged in primary care settings, when a patient first enters the healthcare system. Why are you selling to hospitals?
You’re spot on. The way we see Knit being transformational down the line is, once you have this intelligence at the level of unit economics, and it becomes cheaper and ubiquitously available, you can start to rearrange how medicine routes patients. But even today, fusing Knit into the upstream scheduling infrastructure is an unlock. Let’s suppose I have knee pain. I could go to an orthopedist. I could go to a pain clinic. I could go to rheumatology. Today it's kind of at the whim of your primary care doctor. And you could have to wait a long time and end up in the wrong queue. For rheumatology, some systems are booking out 12 months. And then anywhere from a third to a half of patients that arrive after that wait time don’t even have rheumatologic disease. So by infusing Knit at that operational layer, you’re rewiring the system to leverage the skill you have today.
Why not start with payers? They’d be very interested in a system able to steer patients to the most appropriate site of care.
We started with providers because our view is that if you’re going to be a clinically intelligent, trusted AI, you need to have doctors trust you. With payers, though, there is a lot of alignment. This can ensure patients get the treatments they need and not end up in the hospital for other things. And they want people to stay in network. And a major leakage out of network is that people get referred but have a wait time for nine months, so they go somewhere else. So we’re starting to do some initial work with payers.
What’s your deployment schedule?
Our first three go-lives are in process to be deployed early 2027. We’re aiming for February or March. And our design partners are really excited. One of them started with one use case, and their internal team came up with 97 other applications. Once you get a system thinking about how supply meets demand, it becomes this ‘Aha!’ moment for them. And then the funnel, if you will, is very full of larger health systems that are interested. Because nearly every system we encounter have specialist access at the core of their roadmap. So we’re in pretty deep contracting and sales discussions with them. And we’ve also seen a ton of interest from safety-net hospitals, which are under a ton of pressure and have patients who are hard to track. So demand kind of runs the gamut here.